Datasets:

Abuelnour
/

json_1000_Scientific_Paper

Dataset card Viewer Files Files and versions Community

Split (1)

train · 947 rows

Text stringlengths 45 130k	Id stringlengths 8 8	Summary stringlengths 55 2.67k
three generative lexicalized models for statistical parsing in this paper we first propose a new statistical parsing model which is a generative model of lexicalised contextfree grammar we then extend the model to include a probabilistic treatment of both subcategorisation and whmovement results on wall street journal text show that the parser performs at 881875 constituent precisionrecall an average improvement of 23 over generative models of syntax have been central in linguistics since they were introduced in each sentencetree pair in a language has an associated topdown derivation consisting of a sequence of rule applications of a grammarthese models can be extended to be statistical by defining probability distributions at points of nondeterminism in the derivations thereby assigning a probability p to each pairprobabilistic context free grammar was an early example of a statistical grammara pcfg can be lexicalised by associating a headword with each nonterminal in a parse tree thus far and which both make heavy use of lexical information have reported the best statistical parsing performance on wall street journal textneither of these models is generative instead they both estimate p directlythis paper proposes three new parsing modelsmodel 1 is essentially a generative version of the model described in in model 2 we extend the parser to make the complementadjunct distinction by adding probabilities over subcategorisation frames for headwordsin model 3 we give a probabilistic treatment of whmovement which is derived from the analysis given in generalized phrase structure grammar the work makes two advances over previous models first model 1 performs significantly better than and models 2 and 3 give further improvements our final results are 881875 constituent precisionrecall an average improvement of 23 over second the parsers in and produce trees without information about whmovement or subcategorisationmost nlp applications will need this information to extract predicateargument structure from parse treesin the remainder of this paper we describe the 3 models in section 2 discuss practical issues in section 3 give results in section 4 and give conclusions in section 5in general a statistical parsing model defines the conditional probability p for each candidate parse tree t for a sentence s the parser itself is an algorithm which searches for the tree tbt that maximises pa generative model uses the observation that maximising p is equivalent to maximising p 1 to a topdown derivation of the treein a pcfg for a tree derived by n applications of contextfree rewrite rules lh si rhs 1 i n the rewrite rules are either internal to the tree where lhs is a nonterminal and rhs is a string of one or more nonterminals or lexical where lhs is a part of speech tag and rhs is a worda pcfg can be lexicalised2 by associating a word w and a partofspeech tag t with each nonterminal x in the treethus we write a nonterminal as x where x and x is a constituent labeleach rule now has the form3 h is the headchild of the phrase which inherits the headword h from its parent p l1l7 and are left and right modifiers of h either n or m may be zero and n m 0 for unary rulesfigure 1 shows a tree which will be used as an example throughout this paperthe addition of lexical heads leads to an enormous number of potential rules making direct estimation of p infeasible because of sparse data problemswe decompose the generation of the rhs of a rule such as given the lhs into three steps first generating the head then making the independence assumptions that the left and right modifiers are generated by separate 0thorder markov processes 4 for example the probability of the rule s np np vp would be estimated as but in general the probabilities could be conditioned on any of the preceding modifiersin fact if the derivation order is fixed to be depthfirst that is each modifier recursively generates the subtree below it before the next modifier is generated then the model can also condition on any structure below the preceding modifiersfor the moment we exploit this by making the approximations where distancei and distance are functions of the surface string from the head word to the edge of the constituent the distance measure is the same as in a vector with the following 3 elements is the string of zero length does the string contain a verb does the string contain 0 1 2 or 2 commas i p h h distance the distance is a function of the surface string from the word after h to the last word of r2 inclusivein principle the model could condition on any structure dominated by h r1 or r2 distinction and subcategorisation the tree in figure 1 is an example of the importance of the complementadjunct distinctionit would be useful to identify quotmarksquot as a subject and quotlast weekquot as an adjunct but this distinction is not made in the tree as both nps are in the same position from here on we will identify complements by attaching a quotcquot suffix to nonterminals figure 3 gives an example treea postprocessing stage could add this detail to the parser output but we give two reasons for making the distinction while parsing first identifying complements is complex enough to warrant a probabilistic treatmentlexical information is needed for example knowledge that quotweekquot is likely to be a temporal modifierknowledge about subcategorisation preferences for example that a verb takes exactly one subject is also requiredthese problems are not restricted to nps compare quotthe spokeswoman said quot vs quotbonds beat shortterm investments quot where an sbar headed by quotthatquot is a complement but an sbar headed by quotbecausequot is an adjunctthe second reason for making the complementadjunct distinction while parsing is that it may help parsing accuracythe assumption that complements are generated independently of each other often leads to incorrect parses see figure 4 for further explanationadjuncts in the penn treebank we add the quotcquot suffix to all nonterminals in training data which satisfy the following conditions in addition the first child following the head of a prepositional phrase is marked as a complementthe model could be retrained on training data with the enhanced set of nonterminals and it might learn the lexical properties which distinguish complements and adjuncts however it would still suffer from the bad independence assumptions illustrated in figure 4to solve these kinds of problems the generative process is extended to include a probabilistic choice of left and right subcategorisation frames other leads to errorsin the probability of generating both quotdreyfusquot and quotfundquot as subjects p i s vp was p i s vp was is unreasonably high is similar p vpc i vp vb was p i vp vb was p i vp vb was is a bad independence assumptionprceach subcat frame is a multiset6 specifying the complements which the head requires in its left or right modifiers spectivelythus the subcat requirements are added to the conditioning contextas complements are generated they are removed from the appropriate subcat multisetmost importantly the probability of generating the stop symbol will be 0 when the subcat frame is nonempty and the probability of generating a complement will be 0 when it is not in the subcat frame thus all and only the required complements will be generatedthe probability of the phrase s np npc vp is now here the head initially decides to take a single npc to its left and no complements to its rightnpc is immediately generated as the required subject and npc is removed from lc leaving it empty when the next modifier np is generatedthe incorrect structures in figure 4 should now have low probability because pic and p are smallanother obstacle to extracting predicateargument structure from parse trees is whmovementthis section describes a probabilistic treatment of extraction from relative clausesnoun phrases are most often extracted from subject position object position or from within pps it might be possible to write rulebased patterns which identify traces in a parse treehowever we argue again that this task is best integrated into the parser the task is complex enough to warrant a probabilistic treatment and integration may help parsing accuracya couple of complexities are that modification by an sbar does not always involve extraction quot and it is not uncommon for extraction to occur through several constituents quotthe second reason for an integrated treatment of traces is to improve the parameterisation of the modelin particular the subcategorisation probabilities are smeared by extractionin examples 1 2 and 3 above bought is a transitive verb but without knowledge of traces example 2 in training data will contribute to the probability of bought being an intransitive verbformalisms similar to gpsg handle np extraction by adding a gap feature to each nonterminal in the tree and propagating gaps through the tree until they are finally discharged as a trace complement in extraction cases the penn treebank annotation coindexes a trace with the whnp head of the sbar so it is straightforward to add this information to trees in training datagiven that the lhs of the rule has a gap there are 3 ways that the gap can be passed down to the rhs head the gap is passed to the head of the phrase as in rule in figure 5left right the gap is passed on recursively to one of the left or right modifiers of the head or is discharged as a trace argument to the leftright of the headin rule it is passed on to a right modifier the s complementin rule a trace is generated to the right of the head vbwe specify a parameter pg where g is either head left or rightthe generative process is extended to choose between these cases after generating the head of the phrasethe rest of the phrase is then generated in different ways depending on how the gap is propagated in the head case the left and right modifiers are generated as normalin the left right cases a gap requirement is added to either the left or right subcat variablethis requirement is fulfilled when a trace or a modifier nonterminal which has the gap feature is generatedfor example rule sbar whnp sc has probability in rule right is chosen so the gap requirement is added to rcgeneration of sc ulfills both the sc and gap requirements in rcin rule right is chosen againnote that generation of trace satisfies both the npc and gap subcat requirementstable 1 shows the various levels of backoff for each type of parameter in the modelnote that we decompose p l i p h w t a lc into the product pli i phwtalc x pl2 and then smooth these two probabilities separately in each case the final estimate is where e1 e2 and e3 are maximum likelihood estimates with the context at levels 1 2 and 3 in the table and ai a2 and a3 are smoothing parameters where 0 ai 1all words occurring less than 5 times in training data and words in test data which have never been seen in training are replaced with the quotunknownquot tokenthis allows the model to robustly handle the statistics for rare or new wordspart of speech tags are generated along with the words in this modelwhen parsing the pos tags allowed for each word are limited to those which have been seen in training data for that wordfor unknown words the output from the tagger described in is used as the single possible tag for that worda cky style dynamic programming chart parser is used to find the maximum probability tree for each sentence the parser was trained on sections 02 21 of the wall street journal portion of the penn treebank and tested on section 23 we use the parseval measures to compare performance number of correct constituents in proposed parse number of constituents in proposed parse number of correct constituents in proposed parse number of constituents in treebank parse crossing brackets number of constituents which violate constituent boundaries with a constituent in the treebank parsefor a constituent to be correct it must span the same set of words and have the same label as a constituent in the treebank parsetable 2 shows the results for models 1 2 and 3the precisionrecall of the traces found by model 3 was 933901 where three criteria must be met for a trace to be quotcorrectquot it must be an argument to the correct headword it must be in the correct position in relation to that head word it must be dominated by the correct nonterminal labelfor example in figure 5 the trace is an argument to bought which it follows and it is dominated by a vpof the 436 cases 342 were stringvacuous extraction from subject position recovered with 971982 precisionrecall and 94 were longer distance cases recovered with 76606 precisionrecall 9model 1 is similar in structure to the major differences being that the quotscorequot for each bigram dependency is ps 8 collapses advp and prt to the same label for comparison we also removed this distinction when calculating scores9we exclude infinitival relative clauses from these figures for example quoti called a plumber trace to fix the sinkquot where plumber is coindexed with the trace subject of the infinitivalthe algorithm scored 4118 precisionrecall on the 60 cases in section 23 but infinitival relatives are extremely difficult even for human annotators to distinguish from purpose clauses rather than ps and that there are the additional probabilities of generating the head and the stop symbols for each constituenthowever model 1 has some advantages which may account for the improved performancethe model in is deficient that is for most sentences s et p 1 because probability mass is lost to dependency structures which violate the hard constraint that no links may crossfor reasons we do not have space to describe here model 1 has advantages in its treatment of unary rules and the distance measurethe generative model can condition on any structure that has been previously generated we exploit this in models 2 and 3 whereas is restricted to conditioning on features of the surface string alone also uses a lexicalised generative modelin our notation he decomposes p as p x penn treebank annotation style leads to a very large number of contextfree rules so that directly estimating p may lead to sparse data problems or problems with coverage the complementadjunct distinction and traces increase the number of rules compounding this problem proposes 3 dependency models and gives results that show that a generative model similar to model 1 performs best of the threehowever a pure dependency model omits nonterminal information which is importantfor example quothopequot is likely to generate a vp modifier whereas quotrequirequot is likely to generate an s modifier but omitting nonterminals conflates these two cases giving high probability to incorrect structures such as quoti hope jim to sleepquot or quoti require to sleepquot extends a generative dependency model to include an additional state variable which is equivalent to having nonterminals his suggestions may be close to our models 1 and 2 but he does not fully specify the details of his model and does not give results for parsing accuracy describe a model where the rhs of a rule is generated by a markov process although the process is not headcenteredthey increase the set of nonterminals by adding semantic labels rather than by adding lexical headwords describe a historybased approach which uses decision trees to estimate pour models use much less sophisticated ngram estimation methods and might well benefit from methods such as decisiontree estimation which could condition on richer history than just surface distancethere has recently been interest in using dependencybased parsing models in speech recognition for example it is interesting to note that models 1 2 or 3 could be used as language modelsthe probability for any sentence can be estimated as p et p or as p pwe intend to perform experiments to compare the perplexity of the various models and a structurally similar pure pcfg1this paper has proposed a generative lexicalised probabilistic parsing modelwe have shown that linguistically fundamental ideas namely subcategorisation and whmovement can be given a statistical interpretationthis improves parsing performance and more importantly adds useful information to the parser outputi would like to thank mitch marcus jason eisner dan melamed and adwait ratnaparkhi for many useful discussions and comments on earlier versions of this paperthis work has also benefited greatly from suggestions and advice from scott miller	P97-1003	three generative lexicalized models for statistical parsingin this paper we first propose a new statistical parsing model which is a generative model of lexicalised contextfree grammarwe then extend the model to include a probabilistic treatment of both subcategorisation and whmovementresults on wall street journal text show that the parser performs at 881875 constituent precisionrecall an average improvement of 23 over we provide a 29million word parsed corpus from the wall street journal
automatic detection of text genre as the text databases available to users become larger and more heterogeneous genre becomes increasingly important for computational linguistics as a complement to topical and structural principles of classification we propose a theory of genres as of correlate with various surface cues and argue that genre detection based on surface cues is as successful as detection based on deeper structural properties computational linguists have been concerned for the most part with two aspects of texts their structure and their contentthat is we consider texts on the one hand as formal objects and on the other as symbols with semantic or referential valuesin this paper we want to consider texts from the point of view of genre that is according to the various functional roles they playgenre is necessarily a heterogeneous classificatory principle which is based among other things on the way a text was created the way it is distributed the register of language it uses and the kind of audience it is addressed tofor all its complexity this attribute can be extremely important for many of the core problems that computational linguists are concerned withparsing accuracy could be increased by taking genre into account similarly for postagging in wordsense disambiguation many senses are largely restricted to texts of a particular style such as colloquial or formal in information retrieval genre classification could enable users to sort search results according to their immediate interestspeople who go into a bookstore or library are not usually looking simply for information about a particular topic but rather have requirements of genre as well they are looking for scholarly articles about hypnotism novels about the french revolution editorials about the supercollider and so forthif genre classification is so useful why has not it figured much in computational linguistics before nowone important reason is that up to now the digitized corpora and collections which are the subject of much cl research have been for the most part generically homogeneous so that the problem of genre identification could be set asideto a large extent the problems of genre classification do not become salient until we are confronted with large and heterogeneous search domains like the worldwide webanother reason for the neglect of genre though is that it can be a difficult notion to get a conceptual handle on particularly in contrast with properties of structure or topicality which for all their complications involve wellexplored territoryin order to do systematic work on automatic genre classification by contrast we require the answers to some basic theoretical and methodological questionsis genre a single property or attribute that can be neatly laid out in some hierarchical structureor are we really talking about a multidimensional space of properties that have little more in common than that they are more or less orthogonal to topicalityand once we have the theoretical prerequisites in place we have to ask whether genre can be reliably identified by means of computationally tractable cuesin a broad sense the word quotgenrequot is merely a literary substitute for quotkind of textquot and discussions of literary classification stretch back to aristotlewe will use the term quotgenre here to refer to any widely recognized class of texts defined by some common communicative purpose or other functional traits provided the function is connected to some formal cues or commonalities and that the class is extensiblefor example an editorial is a shortish prose argument expressing an opinion on some matter of immediate public concern typically written in an impersonal and relatively formal style in which the author is denoted by the pronoun webut we would probably not use the term quotgenrequot to describe merely the class of texts that have the objective of persuading someone to do something since that class which would include editorials sermons prayers advertisements and so forth has no distinguishing formal propertiesat the other end of the scale we would probably not use quotgenrequot to describe the class of sermons by john donne since that class while it has distinctive formal characteristics is not extensiblenothing hangs in the balance on this definition but it seems to accord reasonably well with ordinary usagethe traditional literature on genre is rich with classificatory schemes and systems some of which might in retrospect be analyzed as simple attribute systems dubrow fowler frye hernadi hobbes staiger and todorov we will refer here to the attributes used in classifying genres as generic facetsa facet is simply a property which distinguishes a class of texts that answers to certain practical interests and which is moreover associated with a characteristic set of computable structural or linguistic properties whether categorical or statistical which we will describe as quotgeneric cuesquot in principle a given text can be described in terms of an indefinitely large number of facetsfor example a newspaper story about a balkan peace initiative is an example of a broadcast as opposed to directed communication a property that correlates formally with certain uses of the pronoun youit is also an example of a narrative as opposed to a directive suasive or descriptive communication and this facet correlates among other things with a high incidence of preterite verb formsapart from giving us a theoretical framework for understanding genres facets offer two practical advantagesfirst some applications benefit from categorization according to facet not genrefor example in an information retrieval context we will want to consider the opinion feature most highly when we are searching for public reactions to the supercollider where newspaper columns editorials and letters to the editor will be of roughly equal interestfor other purposes we will want to stress narrativity for example in looking for accounts of the storming of the bastille in either novels or historiessecondly we can extend our classification to genres not previously encounteredsuppose that we are presented with the unfamiliar category financial analystsreportby analyzing genres as bundles of facets we can categorize this genre as institutional and as nonsuasive or nonargumentative whereas a system trained on genres as atomic entities would not be able to make sense of an unfamiliar categorythe first linguistic research on genre that uses quantitative methods is that of biber which draws on work on stylistic analysis readability indexing and differences between spoken and written languagebiber ranks genres along several textual quotdimensionsquot which are constructed by applying factor analysis to a set of linguistic syntactic and lexical featuresthose dimensions are then characterized in terms such as quotinformative vs involvedquot or quotnarrative vs nonnarrativequot factors are not used for genre classification rather factors are used to validate hypotheses about the functions of various linguistic featuresan important and more relevant set of experiments which deserves careful attention is presented in karlgren and cutting they too begin with a corpus of handclassified texts the brown corpusone difficulty here however is that it is not clear to what extent the brown corpus classification used in this work is relevant for practical or theoretical purposesfor example the category popular lorequot contains an article by the decidedly highbrow harold rosenberg from commentary and articles from model railroader and gourmet surely not a natural class by any reasonable standardin addition many of the text features in karlgren and cutting are structural cues that require taggingwe will replace these cues with two new classes of cues that are easily computable characterlevel cues and deviation cuesthis section discusses generic cues the observableproperties of a text that are associated with facetsexamples of structural cues are passives nominalizations topicalized sentences and counts of the frequency of syntactic categories these cues are not much discussed in the traditional literature on genre but have come to the fore in recent work for purposes of automatic classification they have the limitation that they require tagged or parsed textsmost facets are correlated with lexical cuesexamples of ones that we use are terms of address which predominate in papers like the new york times latinate affixes which signal certain highbrow registers like scientific articles or scholarly works and words used in expressing dates which are common in certain types of narrative such as news storiescharacterlevel cues are mainly punctuation cues and other separators and delimiters used to mark text categories like phrases clauses and sentences such features have not been used in previous work on genre recognition but we believe they have an important role to play being at once significant and very frequentexamples include counts of question marks exclamations marks capitalized and hyphenated words and acronymsderivative cues are ratios and variation measures derived from measures of lexical and characterlevel featuresratios correlate in certain ways with genre and have been widely used in previous workwe represent ratios implicitly as sums of other cues by transforming all counts into natural logarithmsfor example instead of estimating separate weights a 3 and y for the ratios words per sentence characters per word and words per type respectively we express this desired weighting the 55 cues in our experiments can be combined to almost 3000 different ratiosthe log representation ensures that all these ratios are available implicitly while avoiding overfitting and the high computational cost of training on a large set of cuesvariation measures capture the amount of variation of a certain count cue in a text this type of useful metric has not been used in previous work on genrethe experiments in this paper are based on 55 cues from the last three groups lexical characterlevel and derivative cuesthese cues are easily computable in contrast to the structural cues that have figured prominently in previous work on genrethe corpus of texts used for this study was the brown corpusfor the reasons mentioned above we used our own classification system and eliminated texts that did not fall unequivocally into one of our categorieswe ended up using 499 of the 802 texts in the brown corpusfor our experiments we analyzed the texts in terms of three categorical facets brownarrative and genrebrow characterizes a text in terms of the presumptions made with respect to the required intellectual background of the target audienceits levels are popular middle uppermiddle and highfor example the mainstream american press is classified as middle and tabloid newspapers as popularthe narrative facet is binary telling whether a text is written in a narrative mode primarily relating a sequence of eventsthe genre facet has the values reportage editorial scitech legalnonfiction fictionthe first two characterize two types of articles from the daily or weekly press reportage and editorialsthe level scitech denominates scientific or technical writings and legal characterizes various types of writings about law and government administrationfinally nonfiction is a fairly diverse category encompassing most other types of expository writing and fiction is used for works of fictionour corpus of 499 texts was divided into a training subcorpus and an evaluation subcorpus the evaluation subcorpus was designed to have approximately equal numbers of all represented combinations of facet levelsmost such combinations have six texts in the evaluation corpus but due to small numbers of some types of texts some extant combinations are underrepresentedwithin this stratified framework texts were chosen by a pseudo randomnumber generatorthis setup results in different quantitative compositions of training and evaluation setfor example the most frequent genre level in the training subcorpus is reportage but in the evaluation subcorpus nonfiction predominateswe chose logistic regression as our basic numerical methodtwo informal pilot studies indicated that it gave better results than linear discrimination and linear regressionlr is a statistical technique for modeling a binary response variable by a linear combination of one or more predictor variables using a logit link function and modeling variance with a binomial random variable ie the dependent variable log is modeled as a linear combination of the independent variablesthe model has the form g xi3 where r is the estimated response probability xi is the feature vector for text i and 3 is the weight vector which is estimated from the matrix of feature vectorsthe optimal value of i3 is derived via maximum likelihood estimation using splus for binary decisions the application of lr was straightforwardfor the polytomous facets genre and brow we computed a predictor function independently for each level of each facet and chose the category with the highest predictionthe most discriminating of the 55 variables were selected using stepwise backward selection based on the aic criterion a separate set of variables was selected for each binary discrimination taskin order to see whether our easilycomputable surface cues are comparable in power to the structural cues used in karlgren and cutting we also ran lr with the cues used in their experimentbecause we use individual texts in our experiments instead of the fixedlength conglomerate samples of karlgren and cutting we averaged all count features over text lengthbecause of the high number of variables in our experiments there is a danger that overfitting occurslr also forces us to simulate polytomous decisions by a series of binary decisions instead of directly modeling a multinomial responsefinally classical lr does not model variable interactionsfor these reasons we ran a second set of experiments with neural networks which generally do well with a high number of variables because they protect against overfittingneural nets also naturally model variable interactionswe used two architectures a simple perceptron and a multilayer perceptron with all input units connected to all units of the hidden layer and all units of the hidden layer connected to all output unitsfor binary decisions such as determining whether or not a text is narrative the output layer consists of one sigmoidal output unit for polytomous decisions it consists of four or six softmax units the size of the hidden layer was chosen to be three times as large as the size of the output layer for binary decisions the simple perceptron fits a logistic model just as lr doeshowever it is less prone to overfitting because we train it using threefold crossvalidationvariables are selected by summing the crossentropy error over the three validation sets and eliminating the variable that if eliminated results in the lowest crossentropy errorthe elimination cycle is repeated until this summed crossentropy error starts increasingbecause this selection technique is timeconsuming we only apply it to a subset of the discriminationstable 1 gives the results of the experimentsfor each genre facet it compares our results using surface cues against results using karlgren and cutting structural cues on the one hand and against a baseline on the other each text in the evaluation suite was tested for each facetthus the number 78 for narrative under method quotlr allquot means that when all texts were subjected to the narrative test 78 of them were classified correctlythere are at least two major ways of conceiving what the baseline should be in this experimentif the machine were to guess randomly among k categories the probability of a correct guess would be 1k ie 12 for narrative 16 for genre and 14 for browbut one could get dramatic improvement just by building a machine that always guesses the most populated category nonfict for genremiddle for brow and no for narrativethe first approach would be fair because our machines in fact have no prior knowledge of the distribution of genre facets in the evaluation suite but we decided to be conservative and evaluate our methods against the latter baselineno matter which approach one takes however each of the numbers in the table is significant at p 05 by a binomial distributionthat is there is less than a 5 chance that a machine guessing randomly could have come up with results so much better than the baselineit will be recalled that in the lr models the facets with more than two levels were computed by means of binary decision machines for each level then choosing the level with the most positive scoretherefore some feeling for the internal functioning of our algorithms can be obtained by seeing what the performance is for each of these binary machines and for the sake of comparison this information is also given for some of the neural net modelstable 2 shows how often each of the binary machines correctly determined whether a text did or did not fall in a particular facet levelhere again the appropriate baseline could be determined two waysin a machine that chooses randomly performance would be 50 and all of the numbers in the table would be significantly better than chance but a simple machine that always guesses no would perform much better and it is against this stricter standard that we computed the baseline in table 2here the binomial distribution shows that some numbers are not significantly better than the baselinethe numbers that are significantly better than chance at p 05 by the binomial distribution are starredtables 1 and 2 present aggregate results when all texts are classified for each facet or leveltable 3 by contrast shows which classifications are assigned for texts that actually belong to a specific known levelfor example the first row shows that of the 18 texts that really are of the reportage genre level 83 were correctly classified as reportage 6 were misclassified as editorial and 11 as nonfictionbecause of space constraints we present this amount of detail only for the six genre levels with logistic regression on selected surface variablesthe experiments indicate that categorization decisions can be made with reasonable accuracy on the basis of surface cuesall of the facet level assignments are significantly better than a baseline of always choosing the most frequent level and the performance appears even better when one considers that the machines do not actually know what the most frequent level iswhen one takes a closer look at the performance of the component machines it is clear that some facet levels are detected better than otherstable 2 shows that within the facet genre our systems do a particularly good job on reportage and fiction trend correctly but not necessarily significantly for scitech and nonfiction but perform less well for editorial and legal textswe suspect that the indifferent performance in scitech and legal texts may simply reflect the fact that these genre levels are fairly infrequent in the brown corpus and hence in our training settable 3 sheds some light on the other casesthe lower performance on the editorial and nonfiction tests stems mostly from misclassifying many nonfiction texts as editorialsuch confusion suggests that these genre types are closely related to each other as in fact they areeditorials might best be treated in future experiments as a subtype of nonfiction perhaps distinguished by separate facets such as opinion and institutional authorshipalthough table 1 shows that our methods predict brow at abovebaseline levels further analysis indicates that most of this performance comes from accuracy in deciding whether or not a text is high browthe other levels are identified at near baseline performancethis suggests problems with the labeling of the brow feature in the training datain particular we had labeled journalistic texts on the basis of the overall brow of the host publication a simplification that ignores variation among authors and the practice of printing features from other publicationswe plan to improve those labelings in future experiments by classifying brow on an articlebyarticle basisthe experiments suggest that there is only a small difference between surface and structural cuescomparing lr with surface cues and lr with structural cues as input we find that they yield about the same performance averages of 770 vs 775 for all variables and 784 vs 789 for selected variableslooking at the independent binary decisions on a taskbytask basis surface cues are worse in 10 cases notenumbers are the percentage of the evaluation subcorpus which were correctly assigned to the appropriate facet level the baseline column tells what percentage would be correct if the machine always guessed the most frequent levellr is logistic regression over our surface cues or karlgren and cutting structural cues 2lp and 3lp are 2 or 3layer perceptrons using our surface cuesunder each experimentall tells the results when all cues are used and sel tells the results when for each level one selects the most discriminating cuesa dash indicates that an experiment was not runnotenumbers are the percentage of the evaluation subcorpus which was correctly classified on a binary discrimination taskthe baseline column tells what percentage would be got correct by guessing no for each levelheaders have the same meaning as in table 1 means significantly better than baseline at p 05 using a binomial distribution notenumbers are the percentage of the texts actually belonging to the genre level indicated in the first column that were classified as belonging to each of the genre levels indicated in the column headersthus the diagonals are correct guesses and each row would sum to 100 but for rounding error and better in 8 casessuch a result is expected if we assume that either cue representation is equally likely to do better than the other 041we conclude that there is at best a marginal advantage to using structural cues an advantage that will not justify the additional computational cost in most casesour goal in this paper has been to prepare the ground for using genre in a wide variety of areas in natural language processingthe main remaining technical challenge is to find an effective strategy for variable selection in order to avoid overfitting during trainingthe fact that the neural networks have a higher performance on average and a much higher performance for some discriminations indicates that overfitting and variable interactions are important problems to tackleon the theoretical side we have developed a taxonomy of genres and facetsgenres are considered to be generally reducible to bundles of facets though sometimes with some irreducible atomic residuethis way of looking at the problem allows us to define the relationships between different genres instead of regarding them as atomic entitieswe also have a framework for accommodating new genres as yet unseen bundles of facetsfinally by decomposing genres into facets we can concentrate on whatever generic aspect is important in a particular application further practical tests of our theory will come in applications of genre classification to tagging summarization and other tasks in computational linguisticswe are particularly interested in applications to information retrieval where users are often looking for texts with particular quite narrow generic properties authoritatively written documents opinion pieces scientific articles and so onsorting search results according to genre will gain importance as the typical data base becomes increasingly heterogeneouswe hope to show that the usefulness of retrieval tools can be dramatically improved if genre is one of the selection criteria that users can exploit	P97-1005	automatic detection of text genreas the text databases available to users become larger and more heterogeneous genre becomes increasingly important for computational linguistics as a complement to topical and structural principles of classificationwe propose a theory of genres as bundles of facets which correlate with various surface cues and argue that genre detection based on surface cues is as successful as detection based on deeper structural propertieswe believe that parsing and wordsense disambiguation can also benefit from genre classificationwe avoid structured markers since they require tagged or parsed text and replace them with characterlevel markers and derivative markers ie ratios and variation measures derived from measure of lexical and characterlevel markers
using syntactic dependency as local context to resolve word sense ambiguity most previous corpusbased algorithms disambiguate a word with a classifier trained from previous usages of the same word separate classifiers have to be trained for different words we present an algorithm that uses the same knowledge sources to disambiguate different words the algorithm does not require a sensetagged corpus and exploits the fact that two different words are likely to have similar meanings if they occur in identical local contexts given a word its context and its possible meanings the problem of word sense disambiguation is to determine the meaning of the word in that contextwsd is useful in many natural language tasks such as choosing the correct word in machine translation and coreference resolutionin several recent proposals statistical and machine learning techniques were used to extract classifiers from handtagged corpusyarowsky proposed an unsupervised method that used heuristics to obtain seed classifications and expanded the results to the other parts of the corpus thus avoided the need to handannotate any examplesmost previous corpusbased wsd algorithms determine the meanings of polysemous words by exploiting their local contextsa basic intuition that underlies those algorithms is the following two occurrences of the same word have identical meanings if they have similar local contextsin other words most previous corpusbased wsd algorithms learn to disambiguate a polysemous word from previous usages of the same wordthis has several undesirable consequencesfirstly a word must occur thousands of times before a good classifier can be learnedin yarowsky experiment an average of 3936 examples were used to disambiguate between two sensesin ng and lee experiment 192800 occurrences of 191 words were used as training examplesthere are thousands of polysemous words eg there are 11562 polysemous nouns in wordnetfor every polysemous word to occur thousands of times each the corpus must contain billions of wordssecondly learning to disambiguate a word from the previous usages of the same word means that whatever was learned for one word is not used on other words which obviously missed generality in natural languagesthirdly these algorithms cannot deal with words for which classifiers have not been learnedin this paper we present a wsd algorithm that relies on a different intuition two different words are likely to have similar meanings if they occur in identical local contextsconsider the sentence the new facility will employ 500 of the existing 600 employees the word quotfacilityquot has 5 possible meanings in wordnet 15 installation proficiencytechnique adeptness readiness toiletbathroomto disambiguate the word we consider other words that appeared in an identical local context as quotfacilityquot in table 1 is a list of words that have also been used as the subject of quotemployquot in a 25millionword wall street journal corpusthe quotfreqquot column are the number of times these words were used as the subject of quotemployquotorg includes all proper names recognized as organizations the loga column are their likelihood ratios the meaning of quotfacilityquot in can be determined by choosing one of its 5 senses that is most similar to the meanings of words in table 1this way a polysemous word is disambiguated with past usages of other wordswhether or not it appears in the corpus is irrelevantour approach offers several advantages the required resources of the algorithm include the following an untagged text corpus a broadcoverage parser a concept hierarchy such as the wordnet or roget thesaurus and a similarity measure between conceptsin the next section we introduce our definition of local contexts and the database of local contextsa description of the disambiguation algorithm is presented in section 3section 4 discusses the evaluation resultspsychological experiments show that humans are able to resolve word sense ambiguities given a narrow window of surrounding words most wsd algorithms take as input ito be defined in section 31 a polysemous word and its local contextdifferent systems have different definitions of local contextsin the local context of a word is an unordered set of words in the sentence containing the word and the preceding sentencein a local context of a word consists of an ordered sequence of 6 surrounding partofspeech tags its morphological features and a set of collocationsin our approach a local context of a word is defined in terms of the syntactic dependencies between the word and other words in the same sentencea dependency relationship is an asymmetric binary relationship between a word called head and another word called modifier dependency grammars represent sentence structures as a set of dependency relationshipsnormally the dependency relationships form a tree that connects all the words in a sentencean example dependency structure is shown in the local context of a word w is a triple that corresponds to a dependency relationship in which w is the head or the modifier where type is the type of the dependency relationship such as subj adjn comp i etc word is the word related to w via the dependency relationship and posit ion can either be head or modthe position indicates whether word is the head or the modifier in dependency relationsince a word may be involved in several dependency relationships each occurrence of a word may have multiple local contextsthe local contexts of the two nouns quotboyquot and quotdogquot in are as follows boy dog using a broad coverage parser to parse a corpus we construct a local context databasean entry in the database is a pair where lc is a local context and c is a set of tripleseach triple specifies how often word occurred in lc and the likelihood ratio of lc and wordthe likelihood ratio is obtained by treating word and c as a bigram and computed with the formula in the database entry corresponding to table 1 is as followsthe polysemous words in the input text are disambiguated in the following steps step a parse the input text and extract local contexts of each wordlet lc denote the set of local contexts of all occurrences of w in the input textstep bsearch the local context database and find words that appeared in an identical local context as w they are called selectors of w step c select a sense s of w that maximizes the similarity between w and selectorsstep d the sense s is assigned to all occurrences of w in the input textthis implements the quotone sense per discoursequot heuristic advocated in step c needs further explanationin the next subsection we define the similarity between two word senses we then explain how the similarity between a word and its selectors is maximizedthere have been several proposed measures for similarity between two concepts all of those similarity measures are defined directly by a formulawe use instead an informationtheoretic definition of similarity that can be derived from the following assumptions where cornmon is a proposition that states the commonalities between a and b is the amount of information contained in the proposition s assumption 2 the differences between a and b is measured by where describe is a proposition that describes what a and b areassumption 3 the similarity between a and b sim is a function of their commonality and differencesthat is sim f i the domain of f is ix 0 y 0 y xassumption 4 similarity is independent of the unit used in the information measureaccording to information theory log bp where p is the probability of s and b is the unitwhen b 2 is the number of bits needed to encode s since log bx 12412 assumption 4 means that the function f must satisfy the following condition vc 0 f f consists of two independent parts then the sim is the sum of the similarities computed when each part of the commonality is consideredin other words f f fa corollary of assumption 5 is that vy f f f 0 which means that when there is no commonality between a and b their similarity is 0 no matter how different they arefor example the similarity between quotdepthfirst searchquot and quotleather sofaquot is neither higher nor lower than the similarity between quotrectanglequot and quotinterest ratequotassumption 6 the similarity between a pair of identical objects is 1when a and b are identical knowning their commonalities means knowing what they are ie i i therefore the function f must have the following property vx f 1assumption 7 the function f is continuoussimilarity theorem the similarity between a and b is measured by the ratio between the amount of information needed to state the commonality of a and b and the information needed to fully describe what a and b are proof to prove the theorem we need to show f s since f f we only need to show that when li is a rational number f the result can be generalized to all real numbers because f is continuous and for any real number there are rational numbers that are infinitely close to itsuppose m and n are positive integersthus f f substituting fi for x in this equation qedfor examplefigure 1 is a fragment of the wordnetthe nodes are concepts the links represent isa relationshipsthe number attached to a node c is the probability p that a randomly selected noun refers to an instance of c the probabilities are estimated by the frequency of concepts in semcor a sensetagged subset of the brown corpusif x is a hill and y is a coast the commonality between x and p is that quotx is a geoform and y is a geoformquotthe information contained in this statement is 2 x logpthe similarity between the concepts hill and coast is where p of both c and cwe now provide the details of step c in our algorithmthe input to this step consists of a polysemous word v110 and its selectors wi w2 wothe word wi has ni senses sii sin step c1 construct a similarity matrix the rows and columns represent word sensesthe matrix is divided into x blocksthe blocks on the diagonal are all osthe elements in block sii are the similarity measures between the senses of wi and the senses of itsimilarity measures lower than a threshold 0 are considered to be noise and are ignoredin our experiments 9 02 was usedstep c5 modify the similarity matrix to remove the similarity values between other senses of wi and senses of other wordsfor all 1 j m such that 1 e 1ni and 1 0 imax and j imax and m e 1 nil let us consider again the word quotfacilityquot in it has two local contexts subject of quotemployquot and modifiee of quotnewquot table 1 lists words that appeared in the first local contexttable 2 lists words that appeared in the second local contextonly words with top20 likelihood ratio were used in our experimentsthe two groups of words are merged and used as the selectors of quotfacilityquotthe words quotfacilityquot has 5 senses in the wordnetsenses 1 and 5 are subclasses of artifactsenses 2 and 3 are kinds of statesense 4 is a kind of abstractionmany of the selectors in tables 1 and table 2 have artifact senses such as quotpostquot quotproductquot quotsystemquot quotunitquot quotmemory devicequot quotmachinequot quotplantquot quotmodelquot quotprogramquot etctherefore senses 1 and 5 of quotfacilityquot received much more support 537 and 242 respectively than other sensessense 1 is selectedconsider another example that involves an unknown proper name we treat unknown proper nouns as a polysemous word which could refer to a person an organization or a locationsince quotdreamlandquot is the subject of quotemployedquot its meaning is determined by maximizing the similarity between one of person organization locaton and the words in table 1since table 1 contains many quotorganizationquot words the support for the quotorganizationquot sense is much higher than the otherswe used a subset of the semcor to evaluate our algorithmgeneralpurpose lexical resources such as wordnet longman dictionary of contemporary english and roget thesaurus strive to achieve completenessthey often make subtle distinctions between word sensesas a result when the wsd task is defined as choosing a sense out of a list of senses in a generalpurpose lexical resource even humans may frequently disagree with one another on what the correct sense should bethe subtle distinctions between different word senses are often unnecessarytherefore we relaxed the correctness criteriona selected sense sanswer is correct if it is quotsimilar enoughquot to the sense tag s key in semcorwe experimented with three interpretations of quotsimilar enoughquotthe strictest interpretation is sim1 which is true only when sanswerskeythe most relaxed interpretation is sim 0 which is true if sanswer and skey are the descendents of the same toplevel concepts in wordnet a compromise between these two is sim 027 where 027 is the average similarity of 50000 randomly generated pairs in which w and w belong to the same roget categorywe use three words quotdutyquot quotinterestquot and quotlinequot as examples to provide a rough idea about what sim 027 meansthe word quotdutyquot has three senses in wordnet 15the similarity between the three senses are all below 027 although the similarity between senses 1 and 2 is very close to the thresholdthe word quotinterestquot has 8 sensessenses 1 and 7 are merged2 senses 3 4 and 5 are mergedthe word quotinterestquot is reduced to a 5way ambiguous wordthe other three senses are 2 6 and 8 the word quotlinequot has 27 sensesthe similarity threshold 027 reduces the number of senses to 14the reduced senses are where each group is a reduced sense and the numbers are original wordnet sense numberswe used a 25millionword wall street journal corpus to construct the local context databasethe text was parsed in 126 hours on a sparcultra 1140 with 96mb of memorywe then extracted from the parse trees 8665362 dependency relationships in which the head or the modifier is a nounwe then filtered out pairs with a likelihood ratio lower than 5 the resulting database contains 354670 local contexts with a total of 1067451 words in them since the local context database is constructed from wsj corpus which are mostly business news we only used the quotpress reportagequot part of semcor which consists of 7 files with about 2000 words eachfurthermore we only applied our algorithm to nounstable 3 shows the results on 2832 polysemous nouns in semcorthis number also includes proper nouns that do not contain simple markers to indicate its categorysuch a proper noun is treated as a 3way ambiguous word person organization or locationwe also showed as a baseline the performance of the simple strategy of always choosing the first sense of a word in the wordnetsince the wordnet senses are ordered according to their frequency in semcor choosing the first sense is roughly the same as choosing the sense with highest prior probability except that we are not using all the files in semcorit can be seen from table 3 that our algorithm performed slightly worse than the baseline when the strictest correctness criterion is usedhowever when the condition is relaxed its performance gain is much lager than the baselinethis means that when the algorithm makes mistakes the mistakes tend to be close to the correct answerthe step c in section 32 is similar to resnik noun group disambiguation although he did not address the question of the creation of noun groupsthe earlier work on wsd that is most similar to ours is they proposed a set of heuristic rules that are based on the idea that objects of the same or similar verbs are similarour algorithm treats all local contexts equally in its decisionmakinghowever some local contexts hardly provide any constraint on the meaning of a wordfor example the object of quotgetquot can practically be anythingthis type of contexts should be filtered out or discounted in decisionmakingour assumption that similar words appear in identical context does not always holdfor example where per refers to proper names recognized as personsnone of these is similar to the quotbody partquot meaning of quotheartquotin fact quotheartquot is the only body part that beatswe have presented a new algorithm for word sense disambiguationunlike most previous corpusbased wsd algorithm where separate classifiers are trained for different words we use the same local context database and a concept hierarchy as the knowledge sources for disambiguating all wordsthis allows our algorithm to deal with infrequent words or unknown proper nounsunnecessarily subtle distinction between word senses is a wellknown problem for evaluating wsd algorithms with generalpurpose lexical resourcesour use of similarity measure to relax the correctness criterion provides a possible solution to this problem	P97-1009	using syntactic dependency as local context to resolve word sense ambiguitymost previous corpusbased algorithms disambiguate a word with a classifier trained from previous usages of the same wordseparate classifiers have to be trained for different wordswe present an algorithm that uses the same knowledge sources to disambiguate different wordsthe algorithm does not require a sensetagged corpus and exploits the fact that two different words are likely to have similar meanings if they occur in identical local contextswe define the similarity between two objects to be the amount of information contained in the commonolity between the objects divided by the amount of information in the descriptions of the objects
the rhetorical parsing of unrestricted natural language texts we derive the rhetorical structures of texts by means of two new surfaceformbased algorithms one that identifies discourse usages of cue phrases and breaks sentences into clauses and one that produces valid rhetorical structure trees for unrestricted natural language texts the algorithms use information that was derived from a corpus analysis of cue phrases researchers of natural language have repeatedly acknowledged that texts are not just a sequence of words nor even a sequence of clauses and sentenceshowever despite the impressive number of discourserelated theories that have been proposed so far there have emerged no algorithms capable of deriving the discourse structure of an unrestricted texton one hand efforts such as those described by asher lascarides asher and oberlander kamp and reyle grover et al and prust scha and van den berg take the position that discourse structures can be built only in conjunction with fully specified clause and sentence structuresand hobbs theory assumes that sophisticated knowledge bases and inference mechanisms are needed for determining the relations between discourse unitsdespite the formal elegance of these approaches they are very domain dependent and therefore unable to handle more than a few restricted exampleson the other hand although the theories described by grosz and sidner polanyi and mann and thompson are successfully applied manually they are too informal to support an automatic approach to discourse analysisin contrast with this previous work the rhetorical parser that we present builds discourse trees for unrestricted textswe first discuss the key concepts on which our approach relies and the corpus analysis that provides the empirical data for our rhetorical parsing algorithmwe discuss then an algorithm that recognizes discourse usages of cue phrases and that determines clause boundaries within sentenceslastly we present the rhetorical parser and an example of its operation the mathematical foundations of the rhetorical parsing algorithm rely on a firstorder formalization of valid text structures the assumptions of the formalization are the following1the elementary units of complex text structures are nonoverlapping spans of text2rhetorical coherence and cohesive relations hold between textual units of various sizes3relations can be partitioned into two classes paratactic and hypotacticparatactic relations are those that hold between spans of equal importancehypotactic relations are those that hold between a span that is essential for the writer purpose ie a nucleus and a span that increases the understanding of the nucleus but is not essential for the writer purpose ie a satellite4the abstract structure of most texts is a binary treelike structure5if a relation holds between two textual spans of the tree structure of a text that relation also holds between the most important units of the constituent subspansthe most important units of a textual span are determined recursively they correspond to the most important units of the immediate subspans when the relation that holds between these subspans is paratactic and to the most important units of the nucleus subspan when the relation that holds between the immediate subspans is hypotacticin our previous work we presented a complete axiomatization of these principles in the context of rhetorical structure theory and we described an algorithm that starting from the set of textual units that make up a text and the set of elementary rhetorical relations that hold between these units can derive all the valid discourse trees of that textconsequently if one is to build discourse trees for unrestricted texts the problems that remain to be solved are the automatic determination of the textual units and the rhetorical relations that hold between themin this paper we show how one can find and exploit approximate solutions for both of these problems by capitalizing on the occurrences of certain lexicogrammatical constructssuch constructs can include tense and aspect certain patterns of pronominalization and anaphoric usages itclefts and discourse markers or cue phrases in the work described here we investigate how far we can get by focusing our attention only on discourse markers and lexicogrammatical constructs that can be detected by a shallow analysis of natural language textsthe intuition behind our choice relies on the following facts the discourse segments that they relatein other words we assume that the texts that we process are wellformed from a discourse perspective much as researchers in sentence parsing assume that they are wellformed from a syntactic perspectiveas a consequence we assume that one can bootstrap the full syntactic semantic and pragmatic analysis of the clauses that make up a text and still end up with a reliable discourse structure for that textgiven the above discussion the immediate objection that one can raise is that discourse markers are doubly ambiguous in some cases their use is only sentential ie they make a semantic contribution to the interpretation of a clause and even in the cases where markers have a discourse usage they are ambiguous with respect to the rhetorical relations that they mark and the sizes of the textual spans that they connectwe address now each of these objections in turnsentential and discourse usages of cue phrasesempirical studies on the disambiguation of cue phrases have shown that just by considering the orthographic environment in which a discourse marker occurs one can distinguish between sentential and discourse usages in about 80 of caseswe have taken hirschberg and litman research one step further and designed a comprehensive corpus analysis that enabled us to improve their results and coveragethe method procedure and results of our corpus analysis are discussed in section 3discourse markers are ambiguous with respect to the rhetorical relations that they mark and the sizes of the units that they connectwhen we began this research no empirical data supported the extent to which this ambiguity characterizes natural language textsto better understand this problem the corpus analysis described in section 3 was designed so as to also provide information about the types of rhetorical relations rhetorical statuses and sizes of textual spans that each marker can indicatewe knew from the beginning that it would be impossible to predict exactly the types of relations and the sizes of the spans that a given cue markshowever given that the structure that we are trying to build is highly constrained such a prediction proved to be unnecessary the overall constraints on the structure of discourse that we enumerated in the beginning of this section cancel out most of the configurations of elementary constraints that do not yield correct discourse treesconsider for example the following text although discourse markers are ambiguous one can use them to build discourse trees for unrestricted texts2 this will lead to many new applications in natural language processing3 for the sake of the argument assume that we are able to break text into textual units as labelled above and that we are interested now in finding rhetorical relations between these unitsassume now that we can infer that although marks a concessive relation between satellite 1 and nucleus either 2 or 3 and the colon an elaboration between satellite 3 and nucleus either 1 or 2if we use the convention that hypotactic relations are represented as firstorder predicates having the form rhet_rel and that paratactic relations are represented as predicates having the form rhet_rel a correct representation for text is then the set of two disjunctions given in rhet_re rhet _rel v rhet_re despite the ambiguity of the relations the overall rhetorical structure constraints will associate only one discourse tree with text namely the tree given in figure 1 any discourse tree configuration that uses relations rhet_re and rhet_re will be ruled outfor example relation rhet_re will be ruled out because unit 1 is not an important unit for span 12 and as mentioned at the beginning of this section a rhetorical relation that holds between two spans of a valid text structure must also hold between their most important units the important unit of span 12 is unit 2 ie the nucleus of the relation r het felwe used previous work on cue phrases to create an initial set of more than 450 potential discourse markersfor each potential discourse marker we then used an automatic procedure that extracted from the brown corpus a set of text fragmentseach text fragment contained a quotwindowquot of approximately 200 words and an emphasized occurrence of a markeron average we randomly selected approximately 19 text fragments per marker having few texts for the markers that do not occur very often in the corpus and up to 60 text fragments for markers such as and which we considered to be highly ambiguousoverall we randomly selected more than 7900 textsall the text fragments associated with a potential cue phrase were paired with a set of slots in which an analyst described the following1the orthographic environment that characterizes the usage of the potential discourse markerthis included occurrences of periods commas colons semicolons etc2the type of usage sentential discourse or both3the position of the marker in the textual unit to which it belonged beginningmedial or end4the right boundary of the textual unit associated with the marker5the relative position of the textual unit that the unit containing the marker was connected to before or after6the rhetorical relations that the cue phrase signaled7the textual types of the units connected by the discourse marker from clause to multiple_paragraph8the rhetorical status of each textual unit involved in the relation nucleus or satellitethe algorithms described in this paper rely on the results derived from the analysis of 1600 of the 7900 text fragmentsafter the slots for each text fragment were filled the results were automatically exported into a relational databasethe database was then examined semiautomatically with the purpose of deriving procedures that a shallow analyzer could use to identify discourse usages of cue phrases break sentences into clauses and hypothesize rhetorical relations between textual unitsfor each discourse usage of a cue phrase we derived the followingat the time of writing we have identified 1253 occurrences of cue phrases that exhibit discourse usages and associated with each of them procedures that instruct a shallow analyzer how the surrounding text should be broken into textual unitsthis information is used by an algorithm that concurrently identifies discourse usages of cue phrases and determines the clauses that a text is made ofthe algorithm examines a text sentence by sentence and determines a set of potential discourse markers that occur in each sentenceit then applies left to right the procedures that are associated with each potential markerthese procedures have the following possible effects they can cause an immediate breaking of the current sentence into clausesfor example when an 11 althoughquot marker is found a new clause whose right boundary is just before the occurrence of the marker is createdthe algorithm is then recursively applied on the text that is found having a discourse usagefor example when the cue phrase quotalthoughquot is identified it is also assigned a discourse usagethe decision of whether a cue phrase is considered to have a discourse usage is sometimes based on the context in which that phrase occurs ie it depends on the occurrence of other cue phrasesfor example an quotandquot will not be assigned a discourse usage in most of the cases however when it occurs in conjunction with quotalthoughquot ie quotand althoughquot it will be assigned such a rolethe most important criterion for using a cue phrase in the marker identification procedure is that the cue phrase is used as a discourse marker in at least 90 of the examples that were extracted from the corpusthe enforcement of this criterion reduces on one hand the recall of the discourse markers that can be detected but on the other hand increases significantly the precisionwe chose this deliberately because during the corpus analysis we noticed that most of the markers that connect large textual units can be identified by a shallow analyzerin fact the discourse marker that is responsible for most of our algorithm recall failures is andsince a shallow analyzer cannot identify with sufficient precision whether an occurrence of and has a discourse or a sentential usage most of its occurrences are therefore ignoredit is true that in this way the discourse structures that we build lose some potential finer granularity but fortunately from a rhetorical analysis perspective the loss has insignificant global repercussions the vast majority of the relations that we miss due to recall failures of and are joint and sequence relations that hold between adjacent clausesevaluationto evaluate our algorithm we randomly selected three texts each belonging to a different genre three independent judges graduate students in computational linguistics broke the texts into clausesthe judges were given no instructions about the criteria that they had to apply in order to determine the clause boundaries rather they were supposed to rely on their intuition and preferred definition of clausethe locations in texts that were labelled as clause boundaries by at least two of the three judges were considered to be quotvalid clause boundariesquotwe used the valid clause boundaries assigned by judges as indicators of discourse usages of cue phrases and we determined manually the cue phrases that signalled a discourse relationfor example if an quotandquot was used in a sentence and if the judges agreed that a clause boundary existed just before the quotandquot we assigned that quotandquot a discourse usageotherwise we assigned it a sentential usagehence we manually determined all discourse usages of cue phrases and all discourse boundaries between elementary unitswe then applied our marker and clause identification algorithm on the same textsour algorithm found 808 of the discourse markers with a precision of 895 a result that outperforms hirschberg and litman the same algorithm identified correctly 813 of the clause boundaries with a precision of 903 we are not aware of any surfaceformbased algorithms that achieve similar resultsthe rhetorical parsing algorithm is outlined in figure 2in the first step the marker and clause identification algorithm is appliedonce the textual units are determined the rhetorical parser uses the procedures derived from the corpus analysis to hypothesize rhetorical relations between the textual unitsa constraintsatisfaction procedure similar to that described in then determines all the valid discourse trees for detailsthe rhetorical parsing algorithm has been fully implemented in cdiscourse is ambiguous the same way sentences are more than one discourse structure is usually produced for a textin our experiments we noticed at least for english that the quotbestquot discourse trees are usually those that are skewed to the rightwe believe that the explanation of this observation is that text processing is essentially a lefttoright processusually people write texts so that the most important ideas go first both at the paragraph and at the text levelthe more text writers add the more they elaborate on the text that went before as a consequence incremental discourse building consists mostly of expansion of the right branchesin order to deal with the ambiguity of discourse the rhetorical parser computes a weight for each valid discourse tree and retains only those that are maximalthe weight function reflects how skewed to the right a tree isconsider the following text from the november 1996 issue of scientific american the words in italics denote the discourse markers the square brackets denote in fact journalists are trained to employ this quotpyramidquot approach to writing consciously the boundaries of elementary textual units and the curly brackets denote the boundaries of parenthetical textual units that were determined by the rhetorical parser for details the numbers associated with the square brackets are identification labelswith its distant orbit 50 percent farther from the sun than earth and slim atmospheric blanket mars experiences frigid weather conditions2 surface temperatures typically average about 60 degrees celsius at the equator and can dip to 123 degrees c near the poles3 only the midday sun at tropical latitudes is warm enough to thaw ice on occasion but any liquid water formed in this way would evaporate almost instantly5 because of the low atmospheric pressure6 although the atmosphere holds a small amount of water and waterice clouds sometimes develop7 most martian weather involves blowing dust or carbon dioxide8 each win terfor example a blizzard of frozen carbon dioxide rages over one pole and a few meters of this dryice snow accumulate as previously frozen carbon dioxide evaporates from the opposite polar cap9 yet even on the summer pole in this case the algorithm constructs 8 different treesthe trees are ordered according to their weightsthe quotbestquot tree for text has weight 3 and is fully represented in figure 3the postscript file corresponding to figure 3 was automatically generated by a backend algorithm that uses quotdotquot a preprocessor for drawing directed graphsthe convention that we use is that nuclei are surrounded by solid boxes and satellites by dotted boxes the links between a node and the subordinate nucleus or nuclei are represented by solid arrows and the links between a node and the subordinate satellites by dotted linesthe occurrences of parenthetical information are marked in the text by a p and a unique subordinate satellite that contains the parenthetical informationwe believe that there are two ways to evaluate the correctness of the discourse trees that an automatic process buildsone way is to compare the automatically derived trees with trees that have been built manuallyanother way is to evaluate the impact that the discourse trees that we derive automatically have on the accuracy of other natural language processing tasks such as anaphora resolution intention recognition or text summarizationin this paper we describe evaluations that follow both these avenuesunfortunately the linguistic community has not yet built a corpus of discourse trees against which our rhetorical parser can be evaluated with the effectiveness that traditional parsers areto circumvent this problem two analysts manually built the discourse trees for five texts that ranged from 161 to 725 wordsalthough there were some differences with respect to the names of the relations that the analysts used the agreement with respect to the status assigned to various units and the overall shapes of the trees was significantin order to measure this agreement we associated an importance score to each textual unit in a tree and computed the spearman correlation coefficients between the importance scores derived from the discourse trees built by each analyst2 the spearman correlation coefficient between the ranks assigned for each textual unit on the bases of the discourse trees built by the two analysts was very high 0798 at p 00001 level of significancethe differences between the two analysts came mainly from their interpretations of two of the texts the discourse trees of one analyst mirrored the paragraph structure of the texts while the discourse trees of the other mirrored a logical organization of the text which that analyst believed to be importantthe spearman correlation coefficients with respect to the importance of textual units between the discourse trees built by our program and those built by each analyst were 0480 p 00001 and 0449 p 00001these lower correlation values were due to the differences in the overall shape of the trees and to the fact that the granularity of the discourse trees built by the program was not as fine as that of the trees built by the analystsbesides directly comparing the trees built by the program with those built by analysts we also evaluated the impact that our trees could have on the task of summarizing texta summarization program that uses the rhetorical parser described here recalled 66 of the sentences considered important by 13 judges in the same five texts with a precision of 68in contrast a random procedure recalled on average only 384 of the sentences considered important by the judges with a precision of 384and the microsoft office 97 summarizer recalled 41 of the important sentences with a precision of 39we discuss at length the experiments from which the data presented above was derived in the rhetorical parser presented in this paper uses only the structural constraints that were enumerated in section 2corelational constraints focus theme anaphoric links and other syntactic semantic and pragmatic factors do not yet play a role in our system but we nevertheless expect them to reduce the number of valid discourse trees that can be associated with a textwe also expect that other robust methods for determining coherence relations between textual units such as those described by harabagiu and moldovan will improve the accuracy of the routines that hypothesize the rhetorical relations that hold between adjacent unitswe are not aware of the existence of any other rhetorical parser for englishhowever sumita et al report on a discourse analyzer for japaneseeven if one ignores some computational quotbonusesquot that can be easily exploited by a japanese discourse analyzer there are still some key differences between sumita work and oursparticularly important is the fact that the theoretical foundations of sumita et al analyzer do not seem to be able to accommodate the ambiguity of discourse markers in their are independent of each other against the alternative hypothesis that the rank of a variable is correlated with the rank of another variablethe value of the statistic ranges from 1 indicating that high ranks of one variable occur with low ranks of the other variable through 0 indicating no correlation between the variables to 1 indicating that high ranks of one variable occur with high ranks of the other variable system discourse markers are considered unambiguous with respect to the relations that they signalin contrast our system uses a mathematical model in which this ambiguity is acknowledged and appropriately treatedalso the discourse trees that we build are very constrained structures as a consequence we do not overgenerate invalid trees as sumita et al dofurthermore we use only surfacebased methods for determining the markers and textual units and use clauses as the minimal units of the discourse treesin contrast sumita et al use deep syntactic and semantic processing techniques for determining the markers and the textual units and use sentences as minimal units in the discourse structures that they builda detailed comparison of our work with sumita et al and others work is given in we introduced the notion of rhetorical parsing ie the process through which natural language texts are automatically mapped into discourse treesin order to make rhetorical parsing work we improved previous algorithms for cue phrase disambiguation and proposed new algorithms for determining the elementary textual units and for computing the valid discourse trees of a textthe solution that we described is both general and robustacknowledgementsthis research would have not been possible without the help of graeme hirst there are no right words to thank him for iti am grateful to melanie baljko phil edmonds and steve green for their help with the corpus analysisthis research was supported by the natural sciences and engineering research council of canada	P97-1013	the rhetorical parsing of unrestricted natural language textswe derive the rhetorical structures of texts by means of two new surfaceformbased algorithms one that identifies discourse usages of cue phrases and breaks sentences into clauses and one that produces valid rhetorical structure trees for unrestricted natural language textsthe algorithms use information that was derived from a corpus analysis of cue phraseswe describe a method for text summarization based on a nuclearity and selective retention of hierarchical fragments
machine transliteration it is challenging to translate names and technical terms across languages with different alphabets and sound inventories these items are commonly transliterated ie replaced with approximate phonetic equivalents example english comes out in japanese translating such items from japanese back to english is even more challenging and of practical interest as transliterated items make up the bulk of text phrases not found in bilingual dictionaries we describe and evaluate a method for performing backwards transliterations by machine this method uses a generative model incorporating several distinct stages in the transliteration process translators must deal with many problems and one of the most frequent is translating proper names and technical termsfor language pairs like spanishenglish this presents no great challenge a phrase like antonio gil usually gets translated as antonio gilhowever the situation is more complicated for language pairs that employ very different alphabets and sound systems such as japaneseenglish and arabicenglishphonetic translation across these pairs is called transliterationwe will look at japaneseenglish transliteration in this paperjapanese frequently imports vocabulary from other languages primarily from englishit has a special phonetic alphabet called katakana which is used primarily to write down foreign names and loanwordsto write a word like golfbag in katakana some compromises must be madefor example japanese has no distinct l and ft sounds the two english sounds collapse onto the same japanese sounda similar compromise must be struck for english ii and f also japanese generally uses an alternating consonantvowel structure making it impossible to pronounce lfb without intervening vowelskatakana writing is a syllabary rather than an alphabetthere is one symbol for ga another for gi another for gu etcso the way to write golfbag in katakana is 1 7 7 roughly pronounced goruhubagguhere are a few more examples notice how the transliteration is more phonetic than orthographic the letter h in johnson does not produce any katakanaalso a dotseparator is used to separate words but not consistentlyand transliteration is clearly an informationlosing operation aisukuri imu loses the distinction between ice cream and i screamtransliteration is not trivial to automate but we will be concerned with an even more challenging problemgoing from katakana back to english ie backtransliterationautomating backtransliteration has great practical importance in japaneseenglish machine translationkatakana phrases are the largest source of text phrases that do not appear in bilingual dictionaries or training corpora however very little computational work has been done in this area briefly mentions a patternmatching approach while discuss a hybrid neuralnetexpertsystem approach to transliterationthe informationlosing aspect of transliteration makes it hard to inverthere are some problem instances taken from actual newspaper articles1 english translations appear later in this paperhere are a few observations about backtransliteration like most problems in computational linguistics this one requires full world knowledge for a 100 solutionchoosing between katarina and catalina might even require detailed knowledge of geography and figure skatingat that level human translators find the problem quite difficult as well so we only aim to match or possibly exceed their performancebilingual glossaries contain many entries mapping katakana phrases onto english phrases eg it is possible to automatically analyze such pairs to gain enough knowledge to accurately map new katakana phrases that come along and learning approach travels well to other languages pairshowever a naive approach to finding direct correspondences between english letters and katakana symbols suffers from a number of problemsone can easily wind up with a system that proposes iskrym as a backtransliteration of aisukuriimutaking letter frequencies into account improves this to a more plausiblelooking isclimmoving to real words may give is crime the i corresponds to ai the corresponds to su etcunfortunately the correct answer here is ice creamafter initial experiments along these lines we decided to step back and build a generative model of the transliteration process which goes like this this divides our problem into five subproblemsfortunately there are techniques for coordinating solutions to such subproblems and for using generative models in the reverse directionthese techniques rely on probabilities and bayes rulesuppose we build an english phrase generator that produces word sequences according to some probability distribution pand suppose we build an english pronouncer that takes a word sequence and assigns it a set of pronunciations again probabilistically according to some pgiven a pronunciation p we may want to search for the word sequence w that maximizes pbayesrule let us us equivalently maximize p p exactly the two distributions we have modeledextending this notion we settled down to build five probability distributions given a katakana string o observed by ocr we want to find the english word sequence w that maximizes the sum over all e j and k of following we implement p in a weighted finitestate acceptor and we implement the other distributions in weighted finitestate transducers a wfsa is an statetransition diagram with weights and symbols on the transitions making some output sequences more likely than othersa wfst is a wfsa with a pair of symbols on each transition one input and one outputinputs and outputs may include the empty symbol c also following we have implemented a general composition algorithm for constructing an integrated model p from models p and p treating wfsas as wfsts with identical inputs and outputswe use this to combine an observed katakana string with each of the models in turnthe result is a large wfsa containing all possible english translationswe use dijkstra shortestpath algorithm to extract the most probable onethe approach is modularwe can test each engine independently and be confident that their results are combined correctlywe do no pruning so the final wfsa contains every solution however unlikelythe only approximation is the viterbi one which searches for the best path through a wfsa instead of the best sequence this section describes how we designed and built each of our five modelsfor consistency we continue to print written english word sequences in italics english sound sequences in all capitals japanese sound sequences in lower case and katakana sequences naturally the first model generates scored word sequences the idea being that ice cream should score higher than ice creme which should score higher than aice kreemwe adopted a simple unigram scoring method that multiplies the scores of the known words and phrases in a sequenceour 262000entry frequency list draws its words and phrases from the wall street journal corpus an online english name list and an online gazeteer of place names2 a portion of the wfsa looks like this an ideal word sequence model would look a bit differentit would prefer exactly those strings which are actually grist for japanese transliteratorsfor example people rarely transliterate auxiliary verbs but surnames are often transliteratedwe have approximated such a model by removing highfrequency words like has an are am were them and does plus unlikely words corresponding to japanese sound bites like coup and ohwe also built a separate word sequence model containing only english first and last namesif we know that the transliterated phrase is a personal name this model is more precisethe next wfst converts english word sequences into english sound sequenceswe use the english phoneme inventory from the online cmu pronunciation dictionary3 minus the stress marksthis gives a total of 40 sounds including 14 vowel sounds 25 consonant sounds plus our special symbol the dictionary has pronunciations for 110000 words and we organized a phonemetree based wfst from it ee note that we insert an optional pause between word pronunciationsdue to memory limitations we only used the 50000 most frequent wordswe originally thought to build a general lettertosound wfst on the theory that while wrong pronunciations might occasionally be generated japanese transliterators also mispronounce wordshowever our lettertosound wfst did not match the performance of japanese translit2 available from the acl data collection initiative httpwwwspeechcscmueducgibincmudict erators and it turns out that mispronunciations are l ow l ow modeled adequately in the next stage of the cascadei a r 0 0 r o o next we map english sound sequences onto japanese sound sequencesthis is an inherently informationlosing process as english r and l sounds collapse onto japanese r the 14 english vowel sounds collapse onto the 5 japanese vowel sounds etcwe face two immediate problems an obvious target inventory is the japanese syllabary itself written down in katakana or a roman equivalent with this approach the english sound k corresponds to one of t are or depending on its contextunfortunately because katakana is a syllabary we would be unable to express an obvious and useful generalization namely that english k usually corresponds to japanese k independent of contextmoreover the correspondence of japanese katakana writing to japanese sound sequences is not perfectly onetoone so an independent sound inventory is wellmotivated in any caseour japanese sound inventory includes 39 symbols 5 vowel sounds 33 consonant sounds and one special symbol an english sound sequence like might map onto a japanese sound sequence like note that long japanese vowel sounds are written with two symbols instead of just one this scheme is attractive because japanese sequences are almost always longer than english sequencesour wfst is learned automatically from 8000 pairs of englishjapanese sound sequences eg we were able to produce these pairs by manipulating a small englishkatakana glossaryfor each glossary entry we converted english words into english sounds using the previous section model and we converted katakana words into japanese sounds using the next section modelwe then applied the estimationmaximization algorithm to generate symbolmapping probabilities shown in figure 1our them training goes like this 1for each englishjapanese sequence pair compute all possible alignments between their elementsin our case an alignment is a drawing that connects each english sound with one or more japanese sounds such that all japanese sounds are covered and no lines crossfor example there are two ways to align the pair we then build a wfst directly from the symbolmapping probabilities our wfst has 99 states and 283 arcswe have also built models that allow individual english sounds to be quotswallowedquot however these models are expensive to compute and lead to a vast number of hypotheses during wfst compositionfurthermore in disallowing quotswallowingquot we were able to automatically remove hundreds of potentially harmful pairs from our training set eg because no alignments are possible such pairs are skipped by the learning algorithm cases like these must be solved by dictionary lookup anywayonly two pairs failed to align when we wished they hadboth involved turning english y uw into japanese you as in cu kurerenote also that our model translates each english sound without regard to contextwe have built also contextbased models using decision trees recoded as wfstsfor example at the end of a word english t is likely to come out as rather than however contextbased models proved unnecessary case as learned by estimationmaximizationonly mappings with conditional probabilities greater than 1 are shown so the figures may not sum to 1 for backtransliterationthey are more useful for englishtojapanese forward transliterationto map japanese sound sequences like onto katakana sequences like we manually constructed two wfstscomposed together they yield an integrated wfst with 53 states and 303 arcsthe first wfst simply merges long japanese vowel sounds into new symbols aa uu ee and oothe second wfst maps japanese sounds onto katakana symbolsthe basic idea is to consume a whole syllable worth of sounds before producing any katakana eg o 3 this fragment shows one kind of spelling variation in japanese long vowel sounds are usually written with a long vowel mark but are sometimes written with repeated katakana we combined corpus analysis with guidelines from a japanese textbook to turn up many spelling variations and unusual katakana symbols and so onspelling variation is clearest in cases where an english word like switch shows up transliterated variously in different dictionariestreating these variations as an equivalence class enables us to learn general sound mappings even if our bilingual glossary adheres to a single narrow spelling conventionwe do not however and harmfully restrictive in their unsinoothed incarnations generate all katakana sequences with this model for example we do not output strings that begin with a subscripted vowel katakanaso this model also serves to filter out some illformed katakana sequences possibly proposed by optical character recognitionperhaps uncharitably we can view optical character recognition as a device that garbles perfectly good katakana sequencestypical confusions made by our commercial ocr system include for 1 for 1quot 7 for 7 and 7 for 1to generate preocr text we collected 19500 characters worth of katakana words stored them in a file and printed them outto generate postocr text we ocrd the printoutswe then ran the them algorithm to determine symbolmapping probabilitieshere is part of that table this model outputs a superset of the 81 katakana symbols including spurious quote marks alphabetic symbols and the numeral 7we can now use the models to do a sample backtransliterationwe start with a katakana phrase as observed by ocrwe then serially compose it with the models in reverse ordereach intermediate stage is a wfsa that encodes many possibilitiesthe final stage contains all backtransliterations suggested by the models and we finally extract the best onewe start with the masutaazutoonamento problem from section 1our ocr observes q jthis string has two recognition errors for and 1 for t we turn the string into a chained 12state11arc wfsa and compose it with the p modelthis yields a fatter 12state15arc wfsa which accepts the correct spelling at a lower probabilitynext comes the p model which produces a 28state31arc wfsa whose highestscoring sequence is masutaazutoochiment o next comes p yielding a 62state241arc wfsa whose best sequence is next to last comes p which results in a 2982state4601arc wfsa whose best sequence is masters tone am ent awe this english string is closest phonetically to the japanese but we are willing to trade phonetic proximity for more sensical english we rescore this wfsa by composing it with p and extract the best translation we have performed two largescale experiments one using a fulllanguage p model and one using a personal name language modelin the first experiment we extracted 1449 unique katakana phrases from a corpus of 100 short news articlesof these 222 were missing from an online 100000entry bilingual dictionarywe backtransliterated these 222 phrasesmany of the translations are perfect technical program sex scandal omaha beach new york times ramon diazothers are close tanya harding nickel simpson danger washington world capsome miss the mark nancy care again plus occur patriot miss realwhile it is difficult to judge overall accuracysome of the phases are onomatopoetic and others are simply too hard even for good human translatorsit is easier to identify system weaknesses and most of these lie in the p modelfor example nancy kerrigan should be preferred over nancy care againin a second experiment we took katakana versions of the names of 100 yous politicians eg 712 y q 7l and q y7 we backtransliterated these by machine and asked four human subjects to do the samethese subjects were native english speakers and newsaware we gave them brief instructions examples and hintsthe results were as follows human machine 27 64 7 12 66 24 there is room for improvement on both sidesbeing english speakers the human subjects were good at english name spelling and yous politics but not at japanese phoneticsa native japanese speaker might be expert at the latter but not the formerpeople who are expert in all of these areas however are rareon the automatic side many errors can be correcteda firstnamelastname model would rank richard bryan more highly than richard briana bigram model would prefer orren hatch over olin hatchother errors are due to unigram training problems or more rarely incorrect or brittle phonetic modelsfor example quotlongquot occurs much more often than quotronquot in newspaper text and our word selection does not exclude phrases like quotlong islandquot so we get long wyden instead of ron wydenrare errors are due to incorrect or brittle phonetic modelsstill the machine performance is impressivewhen word separators are removed from the katakana phrases rendering the task exceedingly difficult for people the machine performance is unchangedwhen we use ocr7 of katakana tokens are misrecognized affecting 50 of test strings but accuracy only drops from 64 to 52we have presented a method for automatic backtransliteration which while far from perfect is highly competitiveit also achieves the objectives outlined in section 1it ports easily to new language pairs the p and p models are entirely reusable while other models are learned automaticallyit is robust against ocr noise in a rare example of highlevel language processing being useful in improving lowlevel ocrwe plan to replace our shortestpath extraction algorithm with one of the recently developed kshortest path algorithms we will then return a ranked list of the k best translations for subsequent contextual disambiguation either by machine or as part of an interactive manmachine systemwe also plan to explore probabilistic models for arabicenglish transliterationsimply identifying which arabic words to transliterate is a difficult task in itself and while japanese tends to insert extra vowel sounds arabic is usually written without any vowelsfinally it should also be possible to embed our phonetic shift model p inside a speech recognizer to help adjust for a heavy japanese accent although we have not experimented in this areawe would like to thank alton earl ingram yolanda gil bonnie gloverstalls richard whitney and kenji yamada for their helpful commentswe would correct phonetically equivalent but misspelled incorrect also like to thank our sponsors at the department of defense	P97-1017	machine transliterationit is challenging to translate names and technical terms across languages with different alphabets and sound inventoriesthese items are commonly transliterated ie replaced with approximate phonetic equivalentsfor example computer in english comes out as in japanesetranslating such items from japanese back to english is even more challenging and of practical interest as transliterated items make up the bulk of text phrases not found in bilingual dictionarieswe describe and evaluate a method for performing backwards transliterations by machinethis method uses a generative model incorporating several distinct stages in the transliteration processwe propose to compose a set of weighted finite state transducers to solve the problem of backtransliteration from japanese katakana to english
predicting the semantic orientation of adjectives we identify and validate from a large corpus constraints from conjunctions on the positive or negative semantic orientation of the conjoined adjectives a loglinear regression model uses these constraints to predict whether conjoined adjectives are of same or different orientations achieving 82 accuracy in this task when each conjunction is considered independently combining the constraints across many adjectives a clustering algorithm separates the adjectives into groups of different orientations and finally adjectives are labeled positive or negative evaluations on real data and simulation experiments indicate high levels of performance classification precision is more than 90 for adjectives that occur in a modest number of conjunctions in the corpus the semantic orientation or polarity of a word indicates the direction the word deviates from the norm for its semantic group or lexical field it also constrains the word usage in the language due to its evaluative characteristics for example some nearly synonymous words differ in orientation because one implies desirability and the other does not in linguistic constructs such as conjunctions which impose constraints on the semantic orientation of their arguments the choices of arguments and connective are mutually constrained as illustrated by the tax proposal was simple and wellreceived simplistic but wellreceived simplistic and wellreceived by the publicin addition almost all antonyms have different semantic orientationsif we know that two words relate to the same property but have different orientations we can usually infer that they are antonymsgiven that semantically similar words can be identified automatically on the basis of distributional properties and linguistic cues identifying the semantic orientation of words would allow a system to further refine the retrieved semantic similarity relationships extracting antonymsunfortunately dictionaries and similar sources do not include semantic orientation information 2 explicit links between antonyms and synonyms may also be lacking particularly when they depend on the domain of discourse for example the opposition bear bull appears only in stock market reports where the two words take specialized meaningsin this paper we present and evaluate a method that automatically retrieves semantic orientation information using indirect information collected from a large corpusbecause the method relies on the corpus it extracts domaindependent information and automatically adapts to a new domain when the corpus is changedour method achieves high precision and while our focus to date has been on adjectives it can be directly applied to other word classesultimately our goal is to use this method in a larger system to automatically identify antonyms and distinguish near synonymsour approach relies on an analysis of textual corpora that correlates linguistic features or indicators with i exceptions include a small number of terms that are both negative from a pragmatic viewpoint and yet stand in an antonyrnic relationship such terms frequently lexicalize two unwanted extremes eg verboseterse semantic orientationwhile no direct indicators of positive or negative semantic orientation have been proposed3 we demonstrate that conjunctions between adjectives provide indirect information about orientationfor most connectives the conjoined adjectives usually are of the same orientation compare fair and legitimate and corrupt and brutal which actually occur in our corpus with fair and brutal and corrupt and legitimate which are semantically anomalousthe situation is reversed for but which usually connects two adjectives of different orientationsthe system identifies and uses this indirect information in the following stages in the following sections we first present the set of adjectives used for training and evaluationwe next validate our hypothesis that conjunctions constrain the orientation of conjoined adjectives and then describe the remaining three steps of the algorithmafter presenting our results and evaluation we discuss simulation experiments that show how our method performs under different conditions of sparseness of datafor our experiments we use the 21 million word 1987 wall street journal corpus automatically annotated with partofspeech tags using the parts tagger in order to verify our hypothesis about the orientations of conjoined adjectives and also to train and evaluate our subsequent algorithms we need a 3certain words inflected with negative affixes tend to be mostly negative but this rule applies only to a fraction of the negative wordsfurthermore there are words so inflected which have positive orientation eg independent and unbiasedpositive adequate central clever famous intelligent remarkable reputed sensitive slender thriving negative contagious drunken ignorant lanky listless primitive strident troublesome unresolved unsuspecting set of adjectives with predetermined orientation labelswe constructed this set by taking all adjectives appearing in our corpus 20 times or more then removing adjectives that have no orientationthese are typically members of groups of complementary qualitative terms eg domestic or medicalwe then assigned an orientation label to each adjective using an evaluative approachthe criterion was whether the use of this adjective ascribes in general a positive or negative quality to the modified item making it better or worse than a similar unmodified itemwe were unable to reach a unique label out of context for several adjectives which we removed from consideration for example cheap is positive if it is used as a synonym of inexpensive but negative if it implies inferior qualitythe operations of selecting adjectives and assigning labels were performed before testing our conjunction hypothesis or implementing any other algorithms to avoid any influence on our labelsthe final set contained 1336 adjectives figure 1 shows randomly selected terms from this setto further validate our set of labeled adjectives we subsequently asked four people to independently label a randomly drawn sample of 500 of these adjectivesthey agreed with us that the positivenegative concept applies to 8915 of these adjectives on averagefor the adjectives where a positive or negative label was assigned by both us and the independent evaluators the average agreement on the label was 9738the average interreviewer agreement on labeled adjectives was 9697these results are extremely significant statistically and compare favorably with validation studies performed for other tasks in the pastthey show that positive and negative orientation are objective properties that can be reliably determined by humansto extract conjunctions between adjectives we used a twolevel finitestate grammar which covers complex modification patterns and nounadjective appositionrunning this parser on the 21 million word corpus we collected 13426 conjunctions of adjectives expanding to a total of 15431 conjoined adjective pairsafter morphological transformations the remaining 15048 conjunction tokens involve 9296 distinct pairs of conjoined adjectives each conjunction token is classified by the parser according to three variables the conjunction used the type of modification and the number of the modified noun 4 validation of the conjunction hypothesis using the three attributes extracted by the parser we constructed a crossclassification of the conjunctions in a threeway tablewe counted types and tokens of each conjoined pair that had both members in the set of preselected labeled adjectives discussed above 2748 of all conjoined pairs and 4024 of all conjunction occurrences met this criterionwe augmented this table with marginal totals arriving at 90 categories each of which represents a triplet of attribute values possibly with one or more quotdo not carequot elementswe then measured the percentage of conjunctions in each category with adjectives of same or different orientationsunder the null hypothesis of same proportions of adjective pairs of same and different orientation in a given category the number of same or differentorientation pairs follows a binomial distribution with p 05 we show in table 1 the results for several representative categories and summarize all results below different orientations there are rather surprisingly small differences in the behavior of conjunctions between linguistic environments there are a few exceptions eg appositive and conjunctions modifying plural nouns are evenly split between same and different orientationbut in these exceptional cases the sample is very small and the observed behavior may be due to chancethe analysis in the previous section suggests a baseline method for classifying links between adjectives since 7784 of all links from conjunctions indicate same orientation we can achieve this level of performance by always guessing that a link is of the sameorientation typehowever we can improve performance by noting that conjunctions using but exhibit the opposite pattern usually involving adjectives of different orientationsthus a revised but still simple rule predicts a differentorientation link if the two adjectives have been seen in a but conjunction and a sameorientation link otherwise assuming the two adjectives were seen connected by at least one conjunctionmorphological relationships between adjectives also play a roleadjectives related in form almost always have different semantic orientationswe implemented a morphological analyzer which matches adjectives related in this mannerthis process is highly accurate but unfortunately does not apply to many of the possible pairs in our set of 1336 labeled adjectives 102 pairs are morphologically related among them 99 are of different orientation yielding 9706 accuracy for the morphology methodthis information is orthogonal to that extracted from conjunctions only 12 of the 102 morphologically related pairs have been observed in conjunctions in our corpusthus we add to the predictions made from conjunctions the differentorientation links suggested by morphological relationshipswe improve the accuracy of classifying links derived from conjunctions as same or different orientation with a loglinear regression model exploiting the differences between the various conjunction categoriesthis is a generalized linear model with a linear predictor wtx where x is the vector of the observed counts in the various conjunction categories for the particular adjective pair we try to classify and w is a vector of weights to be learned during trainingthe response y is nonlinearly related to 77 through the inverse logit function en note that y e with each of these endpoints associated with one of the possible outcomeswe have 90 possible predictor variables 42 of which are linearly independentsince using all the 42 independent predictors invites overfitting we have investigated subsets of the full loglinear model for our data using the method of iterative stepwise refinement starting with an initial model variables are added or dropped if their contribution to the reduction or increase of the residual deviance compares favorably to the resulting loss or gain of residual degrees of freedomthis process led to the selection of nine predictor variableswe evaluated the three prediction models discussed above with and without the secondary source of morphology relationsfor the loglinear model we repeatedly partitioned our data into equally sized training and testing sets estimated the weights on the training set and scored the model performance on the testing set averaging the resulting scorestable 2 shows the results of these analysesalthough the loglinear model offers only a small improvement on pair classification than the simpler but prediction rule it confers the important advantage 5when morphology is to be used as a supplementary predictor we remove the morphologically related pairs from the training and testing sets of rating each prediction between 0 and 1we make extensive use of this in the next phase of our algorithmthe third phase of our method assigns the adjectives into groups placing adjectives of the same orientation in the same groupeach pair of adjectives has an associated dissimilarity value between 0 and 1 adjectives connected by sameorientation links have low dissimilarities and conversely differentorientation links result in high dissimilaritiesadjective pairs with no connecting links are assigned the neutral dissimilarity 05the baseline and but methods make qualitative distinctions only for them we define dissimilarity for sameorientation links as one minus the probability that such a classification link is correct and dissimilarity for differentorientation links as the probability that such a classification is correctthese probabilities are estimated from separate training datanote that for these prediction models dissimilarities are identical for similarly classified linksthe loglinear model on the other hand offers an estimate of how good each prediction is since it produces a value y between 0 and 1we construct the model so that 1 corresponds to sameorientation and define dissimilarity as one minus the produced valuesame and differentorientation links between adjectives form a graphto partition the graph nodes into subsets of the same orientation we employ an iterative optimization procedure on each connected component based on the exchange method a nonhierarchical clustering algorithm we define an objective function 4 scoring each possible partition p of the adjectives into two subgroups ci and c2 as where ici stands for the cardinality of cluster i and d is the dissimilarity between adjectives x and ywe want to select the partition pmin that minimizes subject to the additional constraint that for each adjective x in a cluster c where c is the complement of cluster c ie the other member of the partitionthis constraint based on rousseeuw silhouettes helps correct wrong cluster assignmentsto find pi we first construct a random partition of the adjectives then locate the adjective that will most reduce the objective function if it is moved from its current clusterwe move this adjective and proceed with the next iteration until no movements can improve the objective functionat the final iteration the cluster assignment of any adjective that violates constraint is changedthis is a steepestdescent hillclimbing method and thus is guaranteed to convergehowever it will in general find a local minimum rather than the global one the problem is npcomplete we can arbitrarily increase the probability of finding the globally optimal solution by repeatedly running the algorithm with different starting partitionsthe clustering algorithm separates each component of the graph into two groups of adjectives but does not actually label the adjectives as positive or negativeto accomplish that we use a simple criterion that applies only to pairs or groups of words of opposite orientationwe have previously shown that in oppositions of gradable adjectives where one member is semantically unmarked the unmarked member is the most frequent one about 81 of the timethis is relevant to our task because semantic markedness exhibits a strong correlation with orientation the unmarked member almost always having positive orientation we compute the average frequency of the words in each group expecting the group with higher average frequency to contain the positive termsthis aggregation operation increases the precision of the labeling dramatically since indicators for many pairs of words are combined even when some of the words are incorrectly assigned to their groupsince graph connectivity affects performance we devised a method of selecting test sets that makes this dependence explicitnote that the graph density is largely a function of corpus size and thus can be increased by adding more datanevertheless we report results on sparser test sets to show how our algorithm scales upwe separated our sets of adjectives a and conjunction and morphologybased links l into training and testing groups by selecting for several values of the parameter a the maximal subset of a aa which includes an adjective x if and only if there exist at least a links from l between x and other elements of aathis operation in turn defines a subset of l la which includes all links between members of aawe train our loglinear model on l la compute predictions and dissimilarities for the links in la and use these to classify and label the adjectives in aa a must be at least 2 since we need to leave some links for trainingtable 3 shows the results of these experiments for 2 to 5our method produced the correct classification between 78 of the time on the sparsest test set up to more than 92 of the time when a higher number of links was presentmoreover in all cases the ratio of the two group frequencies correctly identified the positive subgroupthese results are extremely significant statistically when compared with the baseline method of randomly assigning orientations to adjectives or the baseline method of always predicting the most frequent category figure 2 shows some of the adjectives in set a4 and their classificationsclassified as positive bold decisive disturbing generous good honest important large mature patient peaceful positive proud sound stimulating straightforward strange talented vigorous witty classified as negativea strong point of our method is that decisions on individual words are aggregated to provide decisions on how to group words into a class and whether to label the class as positive or negativethus the overall result can be much more accurate than the individual indicatorsto verify this we ran a series of simulation experimentseach experiment measures how our algorithm performs for a given level of precision p for identifying links and a given average number of links k for each wordthe goal is to show that even when p is low given enough data we can achieve high performance for the groupingas we noted earlier the corpus data is eventually represented in our system as a graph with the nodes corresponding to adjectives and the links to predictions about whether the two connected adjectives have the same or different orientationthus the parameter p in the simulation experiments measures how well we are able to predict each link independently of the others and the parameter k measures the number of distinct adjectives each adjective appears with in conjunctionsp therefore directly represents the precision of the link classification algorithm while k indirectly represents the corpus sizeto measure the effect of p and k we need to carry out a series of experiments where we systematically vary their valuesfor example as k increases for a given level of precision p for individual links we want to measure how this affects overall accuracy of the resulting groups of nodesthus we need to construct a series of data sets or graphs which represent different scenarios corresponding to a given combination of values of p and k to do this we construct a random graph by randomly assigning 50 nodes to the two possible orientationsbecause we do not have frequency and morphology information on these abstract nodes we cannot predict whether two nodes are of the same or different orientationrather we randomly assign links between nodes so that on average each node participates in k links and 100 x p of all links connect nodes of the same orientationthen we consider these links as identified by the link prediction algorithm as connecting two nodes with the same orientation this is equivalent to the baseline link classification method and provides a lower bound on the performance of the algorithm actually used in our system because of the lack of actual measurements such as frequency on these abstract nodes we also decouple the partitioning and labeling components of our system and score the partition found under the best matching conditions for the actual labelsthus the simulation measures only how well the system separates positive from negative adjectives not how well it determines which is whichhowever in all the experiments performed on real corpus data the system correctly found the labels of the groups any misclassifications came from misplacing an adjective in the wrong groupthe whole procedure of constructing the random graph and finding and scoring the groups is repeated 200 times for any given combination of p and k and the results are averaged thus avoiding accidentally evaluating our system on a graph that is not truly representative of graphs with the given p and k we observe that even for relatively low p our ability to correctly classify the nodes approaches very high levels with a modest number of linksfor p 08 we need only about 7 links per adjective for classification performance over 90 and only 12 links per adjective for performance over 996 the difference between low and high values of p is in the rate at which increasing data increases overall precisionthese results are somewhat more optimistic than those obtained with real data a difference which is probably due to the uniform distributional assumptions in the simulationnevertheless we expect the trends to be similar to the ones shown in figure 3 and the results of table 3 on real data support this expectationwe have proposed and verified from corpus data constraints on the semantic orientations of conjoined adjectiveswe used these constraints to automatically construct a loglinear regression model which combined with supplementary morphology rules predicts whether two conjoined adjectives are of same in each figure the last x coordinate indicates the d the dotted line shows the performance of a random or different orientation with 82 accuracywe then classified several sets of adjectives according to the links inferred in this way and labeled them as positive or negative obtaining 92 accuracy on the classification task for reasonably dense graphs and 100 accuracy on the labeling tasksimulation experiments establish that very high levels of performance can be obtained with a modest number of links per word even when the links themselves are not always correctly classifiedas part of our clustering algorithm output a quotgoodnessoffitquot measure for each word is computed based on rousseeuw silhouettesthis measure ranks the words according to how well they fit in their group and can thus be used as a quantitative measure of orientation refining the binary positivenegative distinctionby restricting the labeling decisions to words with high values of this measure we can also increase the precision of our system at the cost of sacrificing some coveragewe are currently combining the output of this system with a semantic group finding system so that we can automatically identify antonyms from the corpus without access to any semantic descriptionsthe learned semantic categorization of the adjectives can also be used in the reverse direction to help in interpreting the conjunctions they participatewe will also extend our analyses to nouns and verbsthis work was supported in part by the office of naval research under grant n000149510745 jointly by the office of naval research and the advanced research projects agency under grant n0001489j1782 by the national science foundation under grant ger9024069 and by the new york state center for advanced technology under contracts nysstfcat013 and nysstfcat013we thank ken church and the att bell laboratories for making the parts partofspeech tagger available to uswe also thank dragomir radev eric siegel and gregory sean mckinley who provided models for the categorization of the adjectives in our training and testing sets as positive and negative	P97-1023	predicting the semantic orientation of adjectiveswe identify and validate from a large corpus constraints from conjunctions on the positive or negative semantic orientation of the conjoined adjectivesa loglinear regression model uses these constraints to predict whether conjoined adjectives are of same or different orientations achieving 82 accuracy in this task when each conjunction is considered independentlycombining the constraints across many adjectives a clustering algorithm separates the adjectives into groups of different orientations and finally adjectives are labeled positive or negativeevaluations on real data and simulation experiments indicate high levels of performance classification precision is more than 90 for adjectives that occur in a modest number of conjunctions in the corpuswe cluster adjectives into and sets based on conjunction constructions weighted similarity graphs minimumcuts supervised learning and clustering
paradise a framework for evaluating spoken dialogue agents this paper presents paradise a general framework for evaluating spoken rlialogue agents the framework decouples task requirements from an agent dialogue behaviors supports comparisons among dialogue strategies enables the calculation of performance over subdialogues and whole dialogues specifies the relative contribution of various factors to performance and makes it possible to compare agents performing different tasks by normalizing for task complexity recent advances in dialogue modeling speech recognition and natural language processing have made it possible to build spoken dialogue agents for a wide variety of applicationspotential benefits of such agents include remote or handsfree access ease of use naturalness and greater efficiency of interactionhowever a critical obstacle to progress in this area is the lack of a general framework for evaluating and comparing the performance of different dialogue agentsone widely used approach to evaluation is based on the notion of a reference answer an agent responses to a query are compared with a predefined key of minimum and maximum reference answers performance is the proportion of responses that match the keythis approach has many widely acknowledged limitations eg although there may be many potential dialogue strategies for carrying out a task the key is tied to one particular dialogue strategyin contrast agents using different dialogue strategies can be compared with measures such as inappropriate utterance ratio turn correction ratio concept accuracy implicit recovery and transaction success consider a comparison of two train timetable information agents where agent a in dialogue 1 uses an explicit confirmation strategy while agent b in dialogue 2 uses an implicit confirmation strategy danieli and gerbino found that agent a had a higher transaction success rate and produced less inappropriate and repair utterances than agent b and thus concluded that agent a was more robust than agent bhowever one limitation of both this approach and the reference answer approach is the inability to generalize results to other tasks and environments such generalization requires the identification of factors that affect performance for example while danieli and gerbino found that agent a dialogue strategy produced dialogues that were approximately twice as long as agent b they had no way of determining whether agent a higher transaction success or agent b efficiency was more critical to performancein addition to agent factors such as dialogue strategy task factors such as database size and environmental factors such as background noise may also be relevant predictors of performancethese approaches are also limited in that they currently do not calculate performance over subdialogues as well as whole dialogues correlate performance with an external validation criterion or normalize performance for task complexitythis paper describes paradise a general framework for evaluating spoken dialogue agents that addresses these limitationsparadise supports comparisons among dialogue strategies by providing a task representation that decouples what an agent needs to achieve in terms of the task requirements from how the agent carries out the task via dialogueparadise uses a decisiontheoretic framework to specify the relative contribution of various factors to an agent overall performanceperformance is modeled as a weighted function of a taskbased success measure and dialoguebased cost measures where weights are computed by correlating user satisfaction with performancealso performance can be calculated for subdialogues as well as whole dialoguessince the goal of this paper is to explain and illustrate the application of the paradise framework for expository purposes the paper uses simplified domains with hypothetical data throughoutsection 2 describes paradise performance model and section 3 discusses its generality before concluding in section 4paradise uses methods from decision theory to combine a disparate set of performance measures into a single performance evaluation functionthe use of decision theory requires a specification of both the objectives of the decision problem and a set of measures for operationalizing the objectivesthe paradise model is based on the structure of objectives shown in figure 1the paradise model posits that performance can be correlated with a meaningful external criterion such as usability and thus that the overall goal of a spoken dialogue agent is to maximize an objective related to usabilityuser satisfaction ratings have been frequently used in the literature as an external indicator of the usability of a dialogue agentthe model further posits that two types of factors are potential relevant contributors to user satisfaction and that two types of factors are potential relevant contributors to costs in addition to the use of decision theory to create this objective structure other novel aspects of paradise include the use of the kappa coefficient to operationalize task success and the use of linear regression to quantify the relative contribution of the success and cost factors to user satisfactionthe remainder of this section explains the measures used to operationalize the set of objectives and the methodology for estimating a quantitative performance function that reflects the objective structuresection 21 describes paradise task representation which is needed to calculate the taskbased success measure described in section 22section 23 describes the cost measures considered in paradise which reflect both the efficiency and the naturalness of an agent dialogue behaviorssection 24 describes the use of linear regression and user satisfaction to estimate the relative contribution of the success and cost measures in a single performance functionfinally section 25 explains how performance can be calculated for subdialogues as well as whole dialogues while section 26 summarizes the methoda general evaluation framework requires a task representation that decouples what an agent and user accomplish from how the task is accomplished using dialogue strategieswe propose that an attribute value matrix can represent many dialogue tasksthis consists of the information that must be exchanged between the agent and the user during the dialogue represented as a set of ordered pairs of attributes and their possible values2 as a first illustrative example consider a simplification of the train timetable domain of dialogues 1 and 2 where the timetable only contains information about rushhour trains between four cities as shown in table 1this avm consists of four attributes 3 in table 1 these attributevalue pairs are annotated with the direction of information flow to represent who acquires the information although this information is not used for evaluationduring the dialogue the agent must acquire from the user the values of dc ac and dr while the user must acquire dtperformance evaluation for an agent requires a corpus of dialogues between users and the agent in which users execute a set of scenarioseach scenario execution has a corresponding avm instantiation indicating the task information requirements for the scenario where each attribute is paired with the attribute value obtained via the dialoguefor example assume that a scenario requires the user to find a train from torino to milano that leaves in the evening as in the longer versions of dialogues 1 and 2 in figures 2 and 34 table 2 contains an avm corresponding to a quotkeyquot for this scenarioall dialogues resulting from execution of this scenario in which the agent and the user correctly convey all attribute values would have the same avm as the scenario key in table 2the avms of the remaining dialogues would differ from the key by at least one valuethus even though the dialogue strategies in figures 2 and 3 are radically different the avm task representation for these dialogues is identical and the performance of the system for the same task can thus be assessed on the basis of the avm representationsuccess at the task for a whole dialogue is measured by how well the agent and user achieve the information requirements of the task by the end of the dialogue this section explains how paradise uses the kappa coefficient to operationalize the taskbased success measure in figure 1the kappa coefficient k is calculated from a confusion matrix that summarizes how well an agent achieves the information requirements of a particular task for a set of dialogues instantiating a set of scenarios5 for example tables 3 and 4 show two hypothetical confusion matrices that could have been generated in an evaluation of 100 complete dialogues with each of two train timetable agents a and b 6 the values in the matrix cells are based on comparisons between the dialogue and scenario key avmswhenever an attribute value in a dialogue avm matches the value in its scenario key the number in the appropriate diagonal cell of the matrix is incremented by 1the off diagonal cells represent misunderstandings that are not corrected in the dialoguenote that depending on the strategy that a spoken dialogue agent uses confusions across attributes are possible eg quotmilanoquot could be confused with quotmorningquot the effect of misunderstandings that are corrected during the course of the dialogue are reflected in the costs associated with the dialogue as will be discussed belowthe first matrix summarizes how the 100 avms representing each dialogue with agent a compare with the avms representing the relevant scenario keys while the second matrix summarizes the information exchange with agent b labels vi to v4 in each matrix represent the possible values of departcity shown in table 1 v5 to v8 are for arrivalcity etccolumns represent the key specifying which information values the agent and user were supposed to communicate to one another given a particular scenariorows represent the data collected from the dialogue corpus reflecting what attribute values were actually communicated between the agent and the usergiven a confusion matrix m success at achieving the information requirements of the task is measured with the kappa coefficient by chance7 when there is no agreement other than that which would be expected by chance n 0when there is total agreement k 1 ic is superior to other measures of success such as transaction success concept accuracy and percent agreement because n takes into account the inherent complexity of the task by correcting for chance expected agreementthus rc provides a basis for comparisons across agents that are performing different taskswhen the prior distribution of the categories is unknown p the expected chance agreement between the data and the key can be estimated from the distribution of the values in the keysthis can be calculated from confusion matrix m since the columns represent the values in the keysin particular p is the proportion of times that the avms for the actual set of dialogues agree with the avms for the scenario keys and p is the proportion of times that the avms for the dialogues and the keys are expected to agree 7k has been used to measure pairwise agreement among coders making category judgments thus the observed useragent interactions are modeled as a coder and the ideal interactions as an expert coder where ti is the sum of the frequencies in column i of m and t is the sum of the frequencies in m p the actual agreement between the data and the key is always computed from the confusion matrix m given the confusion matrices in tables 3 and 4 p 0079 for both agents8 for agent a p 0795 and frc 0777 while for agent b p 059 and c0555 suggesting that agent a is more successful than b in achieving the task goalsas shown in figure 1 performance is also a function of a combination of cost measuresintuitively cost measures should be calculated on the basis of any user or agent dialogue behaviors that should be minimizeda wide range of cost measures have been used in previous work these include pure efficiency measures such as the number of turns or elapsed time to complete the task as well as measures of qualitative phenomena such as inappropriate or repair utterances paradise represents each cost measure as a function ci that can be applied to any dialoguefirst consider the simplest case of calculating efficiency measures over a whole dialoguefor example let cl be the total number of utterancesfor the whole dialogue d1 in figure 2 c1 is 23 utterancesfor the whole dialogue d2 in figure 3 ci is 10 utterancesto calculate costs over subdialogues and for some of the qualitative measures it is necessary to be able to specify which information goals each utterance contributes toparadise uses its avm representation to link the information goals of the task to any arbitrary dialogue behavior by tagging the dialogue with the attributes for the task9 this makes it possible to evaluate any potential dialogue strategies for achieving the task as well as to evaluate dialogue strategies that operate at the level of dialogue subtasks consider the longer versions of dialogues 1 and 2 in figures 2 and 3each utterance in figures 2 and 3 has been tagged using one or more of the attribute abbreviations in table 1 according to the subtask the utterance contributes toas a convention of this type of tagging using a single confusion matrix for all attributes as in tables 3 and 4 inflates 1g when there are few crossattribute confusions by making p smallerin some cases it might be desirable to calculate ic first for identification of attributes and then for values within attributes or to average ic for each attribute to produce an overall ic for the task9this tagging can be hand generated or system generated and hand correctedpreliminary studies indicate that reliability for human tagging is higher for avm attribute tagging than for other types of discourse segment tagging utterances that contribute to the success of the whole dialogue such as greetings are tagged with all the attributessince the structure of a dialogue reflects the structure of the task the tagging of a dialogue by the avm attributes can be used to generate a hierarchical discourse structure such as that shown in figure 4 for dialogue 1 for example segment s2 in figure 4 is about both departcity and arrivalcity it contains segments s3 and s4 within it and consists of utterances ui u6tagging by avm attributes is required to calculate costs over subdialogues since for any subdialogue task attributes define the subdialoguefor subdialogue s4 in figure 4 which is about the attribute arrivalcity and consists of utterances a6 and u6 c is 2tagging by avm attributes is also required to calculate the cost of some of the qualitative measures such as number of repair utterances for example let c2 be the number of repair utterancesthe repair utterances in figure 2 are a3 through u6 thus c2 is 10 utterances and c2 is 2 utterancesthe repair utterance in figure 3 is u2 but note that according to the avm task tagging u2 simultaneously addresses the information goals for departrangein general if an utterance you contributes to the information goals of n different attributes each attribute accounts for un of any costs derivable from youthus c2 is 5given a set of ci it is necessary to combine the difwprevious work has shown that this can be done with high reliability ferent cost measures in order to determine their relative contribution to performancethe next section explains how to combine is with a set of ci to yield an overall performance measuregiven the definition of success and costs above and the model in figure 1 performance for any dialogue d is defined as followsquot here a is a weight on is the cost functions ci are weighted by wi and h is a z score normalization function the normalization function is used to overcome the problem that the values of ci are not on the same scale as k and that the cost measures ci may also be calculated over widely varying scales this problem is easily solved by normalizing each factor x to its z score cr where is the standard deviation for xagents a and b to illustrate the method for estimatinga performance function we will use a subset of the data from tables 3 and 4 shown in table 5table 5 represents the results quot we assume an additive performance function because it appears that k and the various cost factors ci are utility independent and additive independent it is possible however that user satisfaction data collected in future experiments would indicate otherwiseif so continuing use of an additive function might require a transformation of the data a reworking of the model shown in figure 1 or the inclusion of interaction terms in the model from a hypothetical experiment in which eight users were randomly assigned to communicate with agent a and eight users were randomly assigned to communicate with agent btable 5 shows user satisfaction ratings is number of utterances and number of repair utterances for each of these usersusers 5 and 11 correspond to the dialogues in figures 2 and 3 respectivelyto normalize ci for user 5 we determine that ft is 386 and cc is 189thus h is 083similarly 1v for user 11 is 151to estimate the performance function the weights a and wi must be solved forrecall that the claim implicit in figure 1 was that the relative contribution of task success and dialogue costs to performance should be calculated by considering their contribution to user satisfactionuser satisfaction is typically calculated with surveys that ask users to specify the degree to which they agree with one or more statements about the behavior or the performance of the systema single user satisfaction measure can be calculated from a single question or as the mean of a set of ratingsthe hypothetical user satisfaction ratings shown in table 5 range from a high of 6 to a low of 1given a set of dialogues for which user satisfaction is and the set of ci have been collected experimentally the weights a and wi can be solved for using multiple linear regressionmultiple linear regression produces a set of coefficients describing the relative contribution of each predictor factor in accounting for the variance in a predicted factorin this case on the basis of the model in figure 1 us is treated as the predicted factornormalization of the predictor factors to their z scores guarantees that the relative magnitude of the coefficients directly indicates the relative contribution of each factorregression on the table 5 data for both sets of users tests which factors k kat rep most strongly predicts usin this illustrative example the results of the regression with all factors included shows that only k and rep are significant in order to develop a performance function estimate that includes only significant factors and eliminates redundancies a second regression including only significant factors must then be donein this case a second regression yields the predictive equation ie a is 40 and w2 is 78the results also show rc is significant at p 0003 rep significant at p 0001 and the combination of is and rep account for 92 of the variance in us the external validation criterionthe factor utt was not a significant predictor of performance in part because utt and rep are highly redundantgiven these predictions about the relative contribution of different factors to performance it is then possible to return to the problem first introduced in section 1 given potentially conflicting performance criteria such as robustness and efficiency how can the performance of agent a and agent b be comparedgiven values for a and wi performance can be calculated for both agents using the equation abovethe mean performance of a is 44 and the mean performance of b is 44 suggesting that agent b may perform better than agent a overallthe evaluator must then however test these performance differences for statistical significancein this case a t test shows that differences are only significant at the p 07 level indicating a trend onlyin this case an evaluation over a larger subset of the user population would probably show significant differencessince both tc and ci can be calculated over subdialogues performance can also be calculated at the subdialogue level by using the values for a and wi as solved for abovethis assumes that the factors that are predictive of global performance based on us generalize as predictors of local performance ie within subdialogues defined by subtasks as defined by the attribute taggingi2 consider calculating the performance of the dialogue strategies used by train timetable agents a and b over the subdialogues that repair the value of departcitysegment s3 is an example of such a subdialogue with agent aas in the initial estimation of a performance function our analysis requires experimental data namely a set of values for and c and the application of the z score normalization function to this datahowever the values for rc and ci are now calculated at the subdialogue rather than the whole dialogue levelin addition only data from comparable strategies can be used to calculate the mean and standard deviation for normalizationinformally a comparable strategy is one which applies in the same state and has the same effectsfor example to calculate for agent a over the subdialogues that repair departcity p and p are computed using only the subpart of table 3 concerned with departcityfor agent a p 78 p 265 and frc 70then this value of is is normalized using data from comparable subdialogues with both agent a and agent bbased on the data in tables 3 and 4 the mean is 515 and a is 261 so that h for agent a is 71to calculate c2 for agent a assume that the average number of repair utterances for agent a subdialogues that repair departcity is 6 that the mean over all comparable repair subdialogues is 4 and the standard deviation is 279then h is 72let agent a repair dialogue strategy for subdialogues repairing departcity be ra and agent b repair strategy for departcity be rgthen using the performance equation above predicted performance for ra is for agent b using the appropriate subpart of table 4 to calculate lc assuming that the average number of departcity repair utterances is 138 and using similar i2this assumption has a sound basis in theories of dialogue structure but should be tested empirically calculations yields performance 40 71 78 94 045 thus the results of these experiments predict that when an agent needs to choose between the repair strategy that agent b uses and the repair strategy that agent a uses for repairing departcity it should use agent b strategy rb since the performance is predicted to be greater than the performancenote that the ability to calculate performance over subdialogues allows us to conduct experiments that simultaneously test multiple dialogue strategiesfor example suppose agents a and b had different strategies for presenting the value of departtime without the ability to calculate performance over subdialogues it would be impossible to test the effect of the different presentation strategies independently of the different confirmation strategieswe have presented the paradise framework and have used it to evaluate two hypothetical dialogue agents in a simplified train timetable task domainwe used paradise to derive a performance function for this task by estimating the relative contribution of a set of potential predictors to user satisfactionthe paradise methodology consists of the following steps note that all of these steps are required to develop the performance functionhowever once the weights in the performance function have been solved for user satisfaction ratings no longer need to be collectedinstead predictions about user satisfaction can be made on the basis of the predictor variables as illustrated in the application of paradise to subdialoguesgiven the current state of knowledge it is important to emphasize that researchers should be cautious about generalizing a derived performance function to other agents or tasksperformance function estimation should be done iteratively over many different tasks and dialogue strategies to see which factors generalizein this way the field can make progress on identifying the relationship between various factors and can move towards more predictive models of spoken dialogue agent performancein the previous section we used paradise to evaluate two confirmation strategies using as examples fairly simple information access dialogues in the train timetable domainin this section we demonstrate that paradise is applicable to a range of tasks domains and dialogues by presenting avms for two tasks involving more than information access and showing how additional dialogue phenomena can be tagged using avm attributesfirst consider an extension of the train timetable task where an agent can handle requests to reserve a seat or purchase a ticketthis task could be represented using the avm in table 6 where the agent must now acquire the value of the attribute requesttype in order to know what to do with the other information it has acquiredfigure 5 presents a hypothetical dialogue in this extended task domain and illustrates user utterance types and an agent dialogue strategy that are very different from those in figures 2 and 3first agent c in figure 5 uses a quotno confirmationquot dialogue strategy in contrast to the explicit and implicit confirmation strategies used in figures 2 and 3second figure 5 illustrates new types of user utterances that do not directly further the informational goals of the taskin u2 the user asks the agent a whquestion about the dr attribute itself rather than providing information about that attribute valuesince u2 satisfies a knowledge precondition related to answering cl u2 contributes to the dr goal and is tagged as suchin u3 the user similarly asks a yesno question that addresses a subgoal related to answering clfinally u5 illustrates a user request for an agent action and is tagged with the rt attributethe value of rt in the avm instantiation for the dialogue would be quotreservesecond consider the very different domain and task of diagnosing a fault and repairing a circuit figure 6 presents one dialogue from this domainsmith and gordon collected 144 dialogues for this task in which agent initiative was varied by using different dialogue strategies and tagged each dialogue according to the following subtask structure13 our informational analysis of this task results in the avm shown in table 7note that the attributes are almost identical to smith and gordon list of subtaskscircuitid corresponds to introduction correctcircuitbehavior and currentcircuitbehavior correspond to assessment faulttype corresponds to diagnosis faultcorrection corresponds to repair and test corresponds to testthe attribute names emphasize information exchange while the subtask names emphasize functionfigure 6 is tagged with the attributes from table 7smith and gordon tagging of this dialogue according to their subtask representation was as follows turns 14 were i turns 514 were a turns 1516 were d turns 1718 were r and turns 1935 were t note that there are only two differences between the dialogue structures yielded by the two tagging schemesfirst in our scheme the greetings are tagged with all the attributessecond smith and gordon single tag a corresponds to two attribute tags in table 7 which in our scheme defines an extra level of structure within assessment subdialoguesthis paper presented the paradise framework for evaluating spoken dialogue agentsparadise is a general framework for evaluating spoken dialogue agents that integrates and enhances previous workparadise supports comparisons among dialogue strategies with a task representation that decouples what an agent needs to achieve in terms of the task requirements from how the agent carries out the task via dialoguefurthermore this task representation supports the calculation of performance over subdialogues as well as whole dialoguesin addition because paradise success measure normalizes for task complexity it provides a basis for comparing agents performing different tasksthe paradise performance measure is a function of both task success and dialogue costs and has a number of advantagesfirst it allows us to evaluate performance at any level of a dialogue since k and ci can be calculated for any dialogue subtasksince performance can be measured over any subtask and since dialogue strategies can range over subdialogues or the whole dialogue we can associate performance with individual dialogue strategiessecond because our success measure k takes into account the complexity of the task comparisons can be made across dialogue tasksthird k allows us to measure partial success at achieving the taskfourth performance can combine both objective and subjective cost measures and specifies how to evaluate the relative contributions of those costs factors to overall performancefinally to our knowledge we are the first to propose using user satisfaction to determine weights on factors related to performancein addition this approach is broadly integrative incorporating aspects of transaction success concept accuracy multiple cost measures and user satisfactionin our framework transaction success is reflected in k corresponding to dialogues with a p of 1our performance measure also captures information similar to concept accuracy where low concept accuracy scores translate into either higher costs for acquiring information from the user or lower k scoresone limitation of the paradise approach is that the taskbased success measure does not reflect that some solutions might be better than othersfor example in the train timetable domain we might like our taskbased success measure to give higher ratings to agents that suggest express over local trains or that provide helpful information that was not explicitly requested especially since the better solutions might occur in dialogues with higher costsit might be possible to address this limitation by using the interval scaled data version of k another possibility is to simply substitute a domainspecific taskbased success measure in the performance model for k the evaluation model presented here has many applications in apoken dialogue processingwe believe that the framework is also applicable to other dialogue modalities and to humanhuman taskoriented dialoguesin addition while there are many proposals in the literature for algorithms for dialogue strategies that are cooperative collaborative or helpful to the user very few of these strategies have been evaluated as to whether they improve any measurable aspect of a dialogue interactionas we have demonstrated here any dialogue strategy can be evaluated so it should be possible to show that a cooperative response or other cooperative strategy actually improves task performance by reducing costs or increasing task successwe hope that this framework will be broadly applied in future dialogue researchwe would like to thank james allen jennifer chucarroll morena danieli wieland eckert giuseppe di fabbrizio don hindle julia hirschberg shri narayanan jay wilpon steve whittaker and three anonymous reviews for helpful discussion and comments on earlier versions of this paper	P97-1035	paradise a framework for evaluating spoken dialogue agentsthis paper presents paradise a general framework for evaluating spoken dialogue agentsthe framework decouples task requirements from an agent dialogue behaviors supports comparisons among dialogue strategies enables the calculation of performance over subdialogues and whole dialogues specifies the relative contribution of various factors to performance and makes it possible to compare agents performing different tasks by normalizing for task complexitywe identify three factors which carry an influence on the performance of sdss agent factors task factors and environmental factor eg factors related to the acoustic environment and the transmission channelwe aim to evaluate diaglogue agent strategies by relating overall user satisfaction to other metrics such as task success efficiency measure and qualitative measures
a trainable rulebased algorithm for word segmentation this paper presents a trainable rulebased algorithm for performing word segmentation the algorithm provides a simple languageindependent alternative to largescale lexicalbased segmenters requiring large amounts of knowledge engineering as a standalone segmenter we show our algorithm to produce high performance chinese segmentation in addition we show the transformationbased algorithm to be effective in improving the output of several existing word segmentation algorithms in three different languages this paper presents a trainable rulebased algorithm for performing word segmentationour algorithm is effective both as a highaccuracy standalone segmenter and as a postprocessor that improves the output of existing word segmentation algorithmsin the writing systems of many languages including chinese japanese and thai words are not delimited by spacesdetermining the word boundaries thus tokenizing the text is usually one of the first necessary processing steps making tasks such as partofspeech tagging and parsing possiblea variety of methods have recently been developed to perform word segmentation and the results have been published widelya major difficulty in evaluating segmentation algorithms is that there are no widelyaccepted guidelines as to what constitutes a word and there is therefore no agreement on how to quotcorrectlyquot segment a text in an unsegmented languageit is 1most published segmentation work has been done for chinesefor a discussion of recent chinese segmentation work see sproat et al frequently mentioned in segmentation papers that native speakers of a language do not always agree about the quotcorrectquot segmentation and that the same text could be segmented into several very different sets of words by different native speakerssproat et al and wu and fung give empirical results showing that an agreement rate between native speakers as low as 75 is commonconsequently an algorithm which scores extremely well compared to one native segmentation may score dismally compared to other equally quotcorrectquot segmentationswe will discuss some other issues in evaluating word segmentation in section 31one solution to the problem of multiple correct segmentations might be to establish specific guidelines for what is and is not a word in unsegmented languagesgiven these guidelines all corpora could theoretically be uniformly segmented according to the same conventions and we could directly compare existing methods on the same corporawhile this approach has been successful in driving progress in nlp tasks such as partofspeech tagging and parsing there are valid arguments against adopting it for word segmentationfor example since word segmentation is merely a preprocessing task for a wide variety of further tasks such as parsing information extraction and information retrieval different segmentations can be useful or even essential for the different tasksin this sense word segmentation is similar to speech recognition in which a system must be robust enough to adapt to and recognize the multiple speakerdependent quotcorrectquot pronunciations of wordsin some cases it may also be necessary to allow multiple quotcorrectquot segmentations of the same text depending on the requirements of further processing stepshowever many algorithms use extensive domainspecific word lists and intricate name recognition routines as well as hardcoded morphological analysis modules to produce a predetermined segmentation outputmodifying or retargeting an existing segmentation algorithm to produce a different segmentation can be difficult especially if it is not clear what and where the systematic differences in segmentation areit is widely reported in word segmentation papers2 that the greatest barrier to accurate word segmentation is in recognizing words that are not in the lexicon of the segmentersuch a problem is dependent both on the source of the lexicon as well as the correspondence between the text in question and the lexiconwu and fung demonstrate that segmentation accuracy is significantly higher when the lexicon is constructed using the same type of corpus as the corpus on which it is testedwe argue that rather than attempting to construct a single exhaustive lexicon or even a series of domainspecific lexica it is more practical to develop a robust trainable means of compensating for lexicon inadequaciesfurthermore developing such an algorithm will allow us to perform segmentation in many different languages without requiring extensive morphological resources and domainspecific lexica in any single languagefor these reasons we address the problem of word segmentation from a different directionwe introduce a rulebased algorithm which can produce an accurate segmentation of a text given a rudimentary initial approximation to the segmentationrecognizing the utility of multiple correct segmentations of the same text our algorithm also allows the output of a wide variety of existing segmentation algorithms to be adapted to different segmentation schemesin addition our rulebased algorithm can also be used to supplement the segmentation of an existing algorithm in order to compensate for an incomplete lexiconour algorithm is trainable and language independent so it can be used with any unsegmented hnguagethe key component of our trainable segmentation algorithm is transformationbased errordriven learning the corpusbased language processing method introduced by brill this technique provides a simple algorithm for learning a sequence of rules that can be applied to various nlp tasksit differs from other common corpusbased methods in several waysfor one it is weakly statistical but not probabilistic transformationbased approaches conseoiitly require far less training data than most stical approachesit is rulebased but relies on ee for example sproat et al machine learning to acquire the rules rather than expensive manual knowledge engineeringthe rules produced can be inspected which is useful for gaining insight into the nature of the rule sequence and for manual improvement and debugging of the sequencethe learning algorithm also considers the entire training set at all learning steps rather than decreasing the size of the training data as learning progresses such as is the case in decisiontree induction for a thorough discussion of transformationbased learning see ramshaw and marcus brill work provides a proof of viability of transformationbased techniques in the form of a number of processors including a partofspeech tagger a procedure for prepositional phrase attachment and a bracketing parser all of these provided performance comparable to or better than previous attemptstransformationbased learning has also been successfully applied to text chunking morphological disambiguation and phrase parsing word segmentation can easily be cast as a transformationbased problem which requires an initial model a goal state into which we wish to transform the initial model and a series of transformations to effect this improvementthe transformationbased algorithm involves applying and scoring all the possible rules to training data and determining which rule improves the model the mostthis rule is then applied to all applicable sentences and the process is repeated until no rule improves the score of the training datain this manner a sequence of rules is built for iteratively improving the initial modelevaluation of the rule sequence is carried out on a test set of data which is independent of the training dataif we treat the output of an existing segmentation algorithm as the initial state and the desired segmentation as the goal state we can perform a series of transformations on the initial state removing extraneous boundaries and inserting new boundaries to obtain a more accurate approximation of the goal statewe therefore need only define an appropriate rule syntax for transforming this initial approximathe quotexistingquot algorithm does not need to be a large or even accurate system the algorithm can be arbitrarily simple as long as it assigns some form of initial segmentation tion and prepare appropriate training datafor our experiments we obtained corpora which had been manually segmented by native or nearnative speakers of chinese and thaiwe divided the handsegmented data randomly into training and test setsroughly 80 of the data was used to train the segmentation algorithm and 20 was used as a blind test set to score the rules learned from the training datain addition to chinese and thai we also performed segmentation experiments using a large corpus of english in which all the spaces had been removed from the textsmost of our english experiments were performed using training and test sets with roughly the same 8020 ratio but in section 343 we discuss results of english experiments with different amounts of training dataunfortunately we could not repeat these experiments with chinese and thai due to the small amount of handsegmented data availablethere are three main types of transformations which can act on the current state of an imperfect segmentation in our syntax insert and delete transformations can be triggered by any two adjacent characters and one character to the left or right of the bigramslide transformations can be triggered by a sequence of one two or three characters over which the boundary is to be movedfigure 1 enumerates the 22 segmentation transformations we definewith the above algorithm in place we can use the training data to produce a rule sequence to augment an initial segmentation approximation in order to obtain a better approximation of the desired segmentationfurthermore since all the rules are purely characterbased a sequence can be learned for any character set and thus any languagewe used our rulebased algorithm to improve the word segmentation rate for several segmentation algorithms in three languagesdespite the number of papers on the topic the evaluation and comparison of existing segmentation algorithms is virtually impossiblein addition to the problem of multiple correct segmentations of the same texts the comparison of algorithms is difficult because of the lack of a single metric for reporting scorestwo common measures of performance are recall and precision where recall is defined as the percent of words in the handsegmented text identified by the segmentation algorithm and precision is defined as the percentage of words returned by the algorithm that also occurred in the handsegmented text in the same positionthe component recall and precision scores are then used to calculate an fmeasure where f priin this paper we will report all scores as a balanced fmeasure with 3 1 such that for our chinese experiments the training set consisted of 2000 sentences from a xinhua news agency corpus the test set was a separate set of 560 sentences from the same corpus5 we ran four experiments using this corpus with four different algorithms providing the starting point for the learning of the segmentation transformationsin each case the rule sequence learned from the training set resulted in a significant improvement in the segmentation of the test seta very simple initial segmentation for chinese is to consider each character a distinct wordsince the average word length is quite short in chinese with most words containing only 1 or 2 characters6 this characterasword segmentation correctly identified many onecharacter words and produced an initial segmentation score of f403while this is a low segmentation score this segmentation algorithm identifies enough words to provide a reasonable initial segmentation approximationin fact the caw algorithm alone has been shown to be adequate to be used successfully in chinese information retrievalour algorithm learned 5903 transformations from the 2000 sentence training setthe 5903 transformations applied to the test set improved the score from f403 to 781 a 633 reduction in the error ratethis is a very surprising and encouraging result in that from a very naive initial approximation using no lexicon except that implicit from the training data our rulebased algorithm is able to produce a series of transformations with a high segmentation accuracya common approach to word segmentation is to use a variation of the maximum matching algorithm frequently referred to as the quotgreedy algorithmquot the greedy algorithm starts at the first character in a text and using a word list for the language being segmented attempts to find the longest word in the list starting with that characterif a word is found the maximummatching algorithm marks a boundary at the end of the longest word then begins the same longest match search starting at the character following the matchif no match is found in the word list the greedy algorithm simply skips that character and begins the search starting at the next characterin this manner an initial segmentation can be obtained that is more informed than a simple characterasword approachwe applied the maximum matching algorithm to the test set using a list of 57472 chinese words from the nmsu chseg segmenter this greedy algorithm produced an initial score of f644a sequence of 2897 transformations was learned from the training set applied to the test set they improved the score from f644 to 849 a 578 error reductionfrom a simple chinese word list the rulebased algorithm was thus able to produce asegmentation score comparable to segmentation algorithms developed with a large amount of domain knowledge this score was improved further when combining the characterasword and the maximum matching algorithmsin the maximum matching algorithm described above when a sequence of characters occurred in the text and no subset of the sequence was present in the word list the entire sequence was treated as a single wordthis often resulted in words containing 10 or more characters which is very unlikely in chinesein this experiment when such a sequence of characters was encountered each of the characters was treated as a separate word as in the caw algorithm abovethis variation of the greedy algorithm using the same list of 57472 words produced an initial score of f829a sequence of 2450 transformations was learned from the training set applied to the test set they improved the score from f829 to 877 a 281 error reductionthe score produced using this variation of the maximum matching algorithm combined with a rule sequence is nearly equal to the score produced by the nmsu segmenter segmenter discussed in the next sectionthe previous three experiments showed that our rule sequence algorithm can produce excellent segmentation results given very simple initial segmentation algorithmshowever assisting in the adaptation of an existing algorithm to different segmentation schemes as discussed in section 1 would most likely be performed with an already accurate fullydeveloped algorithmin this experiment we demonstrate that our algorithm can also improve the output of such a systemthe chinese segmenter chseg developed at the computing research laboratory at new mexico state university is a complete system for highaccuracy chinese segmentation in addition to an initial segmentation module that finds words in a text based on a list of chinese words chseg additionally contains specific modules for recognizing idiomatic expressions derived words chinese person names and foreign proper namesthe accuracy of chseg on an 86mb corpus has been independently reported as f840 on our test set chseg produced a segmentation score of f879our rulebased algorithm learned a sequence of 1755 transformations from the training set applied to the test set they improved the score from 879 to 896 a 140 reduction in the error rateour rulebased algorithm is thus able to produce an improvement to an existing highperformance systemtable 1 shows a summary of the four chinese experimentswhile thai is also an unsegmented language the thai writing system is alphabetic and the average word length is greater than chinese7 we would therefore expect that our characterbased transformations would not work as well with thai since a context of more than one character is necessary in many cases to make many segmentation decisions in alphabetic languagesthe thai corpus consisted of texts from the thai news agency via nectec in thailandfor our experiment the training set consisted of 3367 sentences the test set was a separate set of 1245 sentences from the same corpusthe initial segmentation was performed using the maximum matching algorithm with a lexicon of 9933 thai words from the word separation filter in cite a thai language latex packagethis greedy algorithm gave an initial segmentation score of f482 on the test setour rulebased algorithm learned a sequence of 731 transformations which improved the score from 482 to 636 a 297 error reductionwhile the alphabetic system is obviously harder to segment we still see a significant reduction in the segmenter error rate using the transformationbased algorithmnevertheless it is doubtful that a segmentation with a score of 636 would be useful in too many applications and this result will need to be significantly improvedalthough english is not an unsegmented language the writing system is alphabetic like thai and the average word length is similarsince english language resources are more readily available it is instructive to experiment with a desegmented english corpus that is english texts in which the spaces have been removed and word boundaries are not explicitly indicatedthe following shows an example of an english sentence and its desegmented version about 20000 years ago the last ice age endedabout20000yearsagothelasticeageended the results of such experiments can help us determine which resources need to be compiled in order to develop a highaccuracy segmentation algorithm in unsegmented alphabetic languages such as thaiin addition we are also able to provide a more detailed error analysis of the english segmentation our english experiments were performed using a corpus of texts from the wall street journal the training set consisted of 2675 sentences in which all the spaces had been removed the test set was a separate set of 700 sentences from the same corpus for an initial experiment segmentation was performed using the maximum matching algorithm with a large lexicon of 34272 english words compiled from the wsjquot in contrast to the low initial thai score the greedy algorithm gave an initial english segmentation score of f732our rulebased algorithm learned a sequence of 800 transformations the average length of a word in our english data was 446 characters compared to 501 for thai and 160 for chinese10note that the portion of the wsj corpus used to compile the word list was independent of both the training and test sets used in the segmentation experiments which improved the score from 732 to 790 a 216 error reductionthe difference in the greedy scores for english and thai demonstrates the dependence on the word list in the greedy algorithmfor example an experiment in which we randomly removed half of the words from the english list reduced the performance of the greedy algorithm from 732 to 323 although this reduced english word list was nearly twice the size of the thai word list the longest match segmentation utilizing the list was much lower successive experiments in which we removed different random sets of half the words from the original list resulted in greedy algorithm performance of 392 351 and 355yet despite the disparity in initial segmentation scores the transformation sequences effect a significant error reduction in all cases which indicates that the transformation sequences are effectively able to compensate for weaknesses in the lexicontable 2 provides a summary of the results using the greedy algorithm for each of the three languagesas mentioned above lexical resources are more readily available for english than for thaiwe can use these resources to provide an informed initial segmentation approximation separate from the greedy algorithmusing our native knowledge of english as well as a short list of common english prefixes and suffixes we developed a simple algorithm for initial segmentation of english which placed boundaries after any of the suffixes and before any of the prefixes as well as segmenting punctuation charactersin most cases this simple approach was able to locate only one of the two necessary boundaries for recognizing full words and the initial score was understandably low f298nevertheless even from this flawed initial approximation our rulebased algorithm learned a sequence of 632 transformations which nearly doubled the word recall improving the score from 298 to 533 a 335 error reductionsince we had a large amount of english data we also performed a classic experiment to determine the effect the amount of training data had on the ability of the rule sequences to improve segmentationwe started with a training set only slightly larger than the test set 872 sentences and repeated the maximum matching experiment described in section 341we then incrementally increased the amount of training data and repeated the experimentthe results summarized in table 3 clearly indicate that more training sentences produce both a longer rule sequence and a larger error reduction in the test dataupon inspection of the english segmentation errors produced by both the maximum matching algorithm and the learned transformation sequences one major category of errors became clearmost apparent was the fact that the limited context transformations were unable to recover from many errors introduced by the naive maximum matching algorithmfor example because the greedy algorithm always looks for the longest string of characters which can be a word given the character sequence quoteconomicsituationquot the greedy algorithm first recognized quoteconomicsquot and several shorter words segmenting the sequence as quoteconomics it you at io nquotsince our transformations consider only a single character of context the learning algorithm was unable to patch the smaller segments back together to produce the desired output quoteconomic situationquotin some cases the transformations were able to recover some of the word but were rarely able to produce the full desired outputfor example in one case the greedy algorithm segmented quothumanactivityquot as quothumana c ti vi tyquotthe rule sequence was able to transform this into quothumana ctivityquot but was not able to produce the desired quothuman activityquotthis suggests that both the greedy algorithm and the transformation learning algorithm need to have a more global word model with the ability to recognize the impact of placing a boundary on the longer sequences of characters surrounding that pointthe results of these experiments demonstrate that a transformationbased rule sequence supplementing a rudimentary initial approximation can produce accurate segmentationin addition they are able to improve the performance of a wide range of segmentation algorithms without requiring expensive knowledge engineeringlearning the rule sequences can be achieved in a few hours and requires no languagespecific knowledgeas discussed in section 1 this simple algorithm could be used to adapt the output of an existing segmentation algorithm to different segmentation schemes as well as compensating for incomplete segmenter lexica without requiring modifications to segmenters themselvesthe rulebased algorithm we developed to improve word segmentation is very effective for segmenting chinese in fact the rule sequences combined with a very simple initial segmentation such as that from a maximum matching algorithm produce performance comparable to manuallydeveloped segmentersas demonstrated by the experiment with the nmsu segmenter the rule sequence algorithm can also be used to improve the output of an already highlyaccurate segmenter thus producing one of the best segmentation results reported in the literaturein addition to the excellent overall results in chinese segmentation we also showed the rule sequence algorithm to be very effective in improving segmentation in thai an alphabetic languagewhile the scores themselves were not as high as the chinese performance the error reduction was nevertheless very high which is encouraging considering the simple rule syntax usedthe current state of our algorithm in which only three characters are considered at a time will understandably perform better with a language like chinese than with an alphabetic language like thai where average word length is much greaterthe simple syntax described in section 22 can however be easily extended to consider larger contexts to the left and the right of boundaries this extension would necessarily come at a corresponding cost in learning speed since the size of the rule space searched during training would grow accordinglyin the future we plan to further investigate the application of our rulebased algorithm to alphabetic languagesacknowledgements this work would not have been possible without the assistance and encouragement of all the members of the mitre natural language groupthis paper benefited greatly from discussions with and comments from marc vilain lynette hirschman sam bayer and the anonymous reviewers	P97-1041	a trainable rulebased algorithm for word segmentationthis paper presents a trainable rulebased algorithm for performing word segmentationthe algorithm provides a simple languageindependent alternative to largescale lexicaibased segmenters requiring large amounts of knowledge engineeringas a standalone segmenter we show our algorithm to produce high performance chinese segmentationin addition we show the transformationbased algorithm to be effective in improving the output of several existing word segmentation algorithms in three different languagesour chinese segmenter makes use of only a manually segmented corpus without referring to any lexicon
a wordtoword model of translational equivalence many multilingual nlp applications need to translate words between different languages but cannot afford the computational expense of inducing or applying a full translation model for these applications we have designed a fast algorithm for estimating a partial translation model which accounts for translational equivalence only at the word level the model precisionrecall tradeoff can be directly controlled via one threshold parameter this feature makes the model more suitable for applications that are not fully statistical the model hidden parameters can be easily conditioned on information extrinsic to the model providing an easy way to integrate preexisting knowledge such as partofspeech dictionaries word order etc our model can link word tokens in parallel texts as well as other translation models in the literature unlike other translation models it can automatically produce dictionarysized translation lexicons and it can do so with over 99 accuracy over the past decade researchers at ibm have developed a series of increasingly sophisticated statistical models for machine translation however the ibm models which attempt to capture a broad range of translation phenomena are computationally expensive to applytable lookup using an explicit translation lexicon is sufficient and preferable for many multilingual nlp applications including quotcrummyquot mt on the world wide web certain machineassisted translation tools concordancing for bilingual lexicography computerassisted language learning corpus linguistics and crosslingual information retrieval in this paper we present a fast method for inducing accurate translation lexiconsthe method assumes that words are translated onetoonethis assumption reduces the explanatory power of our model in comparison to the ibm models but as shown in section 31 it helps us to avoid what we call indirect associations a major source of errors in other modelssection 31 also shows how the onetoone assumption enables us to use a new greedy competitive linking algorithm for reestimating the model parameters instead of more expensive algorithms that consider a much larger set of word correspondence possibilitiesthe model uses two hidden parameters to estimate the confidence of its own predictionsthe confidence estimates enable direct control of the balance between the model precision and recall via a simple thresholdthe hidden parameters can be conditioned on prior knowledge about the bitext to improve the model accuracywith the exception of previous methods for automatically constructing statistical translation models begin by looking at word cooccurrence frequencies in bitexts a bitext comprises a pair of texts in two languages where each text is a translation of the otherword cooccurrence can be defined in various waysthe most common way is to divide each half of the bitext into an equal number of segments and to align the segments so that each pair of segments si and ti are translations of each other then two word tokens are said to cooccur in theour translation model consists of the hidden parameters a and c and likelihood ratios lthe two hidden parameters are the probabilities of the model generating true and false positives in the datal represents the likelihood that you and v can be mutual translationsfor each cooccurring pair of word types you and v these likelihoods are initially set proportional to their cooccurrence frequency and inversely proportional to their marginal frequencies n and n 1 following 2when the l are reestimated the model hidden parameters come into playafter initialization the model induction algorithm iterates the competitive linking algorithm and its onetoone assumption are detailed in section 31section 31 explains how to reestimate the model parametersthe competitive linking algorithm is designed to overcome the problem of indirect associations illustrated in figure 1the sequences of you and v represent corresponding regions of a bitextif uk and vk cooccur much more often than expected by chance then any reasonable model will deem them likely to be mutual translationsif uk and vk are indeed mutual translations then their tendency to the cooccurrence frequency of a word type pair is simply the number of times the pair cooccurs in the corpushowever n ev n which is not the same as the frequency of you because each token of you can cooccur with several differentv2we could just as easily use other symmetric quotassociationquot measures such as 02 or the dice coefficient cooccur is called a direct associationnow suppose that uk and uki often cooccur within their languagethen vk and uki will also cooccur more often than expected by chancethe arrow connecting vk and uki in figure 1 represents an indirect association since the association between vk and uki arises only by virtue of the association between each of them and uk models of translational equivalence that are ignorant of indirect associations have quota tendency to be confused by collocatesquot fortunately indirect associations are usually not difficult to identify because they tend to be weaker than the direct associations on which they are based the majority of indirect associations can be filtered out by a simple competition heuristic whenever several word tokens ui in one half of the bitext cooccur with a particular word token v in the other half of the bitext the word that is most likely to be v translation is the one for which the likelihood l of translational equivalence is highestthe competitive linking algorithm implements this heuristic nb a and a need not sum to 1 because they are conditioned on different events would be the winners in any competitions involving you or v the competitive linking algorithm is more greedy than algorithms that try to find a set of link types that are jointly most probable over some segment of the bitextin practice our linking algorithm can be implemented so that its worstcase running time is 0 where 1 and m are the lengths of the aligned segmentsthe simplicity of the competitive linking algorithm depends on the onetoone assumption each word translates to at most one other wordcertainly there are cases where this assumption is falsewe prefer not to model those cases in order to achieve higher accuracy with less effort on the cases where the assumption is truethe purpose of the competitive linking algorithm is to help us reestimate the model parametersthe variables that we use in our estimation are summarized in figure 2the linking algorithm produces a set of links between word tokens in the bitextwe define a link token to be an ordered pair of word tokens one from each half of the bitexta link type is an ordered pair of word typeslet n be the cooccurrence frequency of you and v and k be the number of links between tokens of you and v3an note that k depends on the linking algorithm but n is a constant property of the bitext important property of the competitive linking algorithm is that the ratio kn tends to be very high if you and v are mutual translations and quite low if they are notthe bimodality of this ratio for several values of n is illustrated in figure 3this figure was plotted after the model first iteration over 300000 aligned sentence pairs from the canadian hansard bitextnote that the frequencies are plotted on a log scale the bimodality is quite sharpthe linking algorithm creates all the links of a given type independently of each other so the number k of links connecting word types you and v has a binomial distribution with parameters n and pif you and v are mutual translations then p tends to a relatively high probability which we will call aif you and v are not mutual translations then p tends to a very low probability which we will call a a and a correspond to the two peaks in the frequency distribution of kn in figure 2the two parameters can also be interpreted as the percentage of true and false positivesif the translation in the bitext is consistent and the model is accurate then a should be near 1 and ashould be near 0to find the most probable values of the hidden model parameters a and a we adopt the standard method of maximum likelihood estimation and find the values that maximize the probability of the link frequency distributionsthe onetoone assumption implies independence between different link types so that the factors on the righthand side of equation 1 can be written explicitly with the help of a mixture coefficientlet r be the probability that an arbitrary cooccurring pair of word types are mutual translationslet b denote the probability that k links are observed out of n cooccurrences where k has a binomial distribution with parameters n and p then the probability that you and v are linked k times out of n cooccurrences is a mixture of two binomials one more variable allows us to express t in terms of a and a let a be the probability that an arbitrary cooccuring pair of word tokens will be linked regardless of whether they are mutual translationssince t is constant over all word types it also represents the probability that an arbitrary cooccurring pair of word tokens are mutual translationstherefore a can also be estimated empiricallylet k be the total number of links in the bitext and let n be the total number of cooccuring word token pairs k equating the righthand sides of equations and and rearranging the terms we get since r is now a function of a and a only the latter two variables represent degrees of freedom in the modelthe probability function expressed by equations 1 and 2 has many local maximain practice these local maxima are like pebbles on a mountain invisible at low resolutionwe computed equation 1 over various combinations of a and a after the model first iteration over 300000 aligned sentence pairs from the canadian hansard bitextfigure 4 shows that the region of interest in the parameter space where 1 a a a 0 has only one clearly visible global maximumthis global maximum can be found by standard hillclimbing methods as long as the step size is large enough to avoid getting stuck on the pebblesgiven estimates for a and a we can compute b and bthese are probabilities that kuv links were generated by an algorithm that generates correct links and by an algorithm that generates incorrect links respectively out of n cooccurrencesthe ratio of these probabilities is the likelihood ratio in favor of you and v being mutual translations for all you and vin the basic wordtoword model the hidden parameters a and a depend only on the distributions of link frequencies generated by the competitive linking algorithmmore accurate models can be induced by taking into account various features of the linked tokensfor example frequent words are translated less consistently than rare words to account for this difference we can estimate separate values of a and a for different ranges of nsimilarly the hidden parameters can be conditioned on the linked parts of speechword order can be taken into account by conditioning the hidden parameters on the relative positions of linked word tokens in their respective sentencesjust as easily we can model links that coincide with entries in a preexisting translation lexicon separately from those that do notthis method of incorporating dictionary information seems simpler than the method proposed by brown et al for their models when the hidden parameters are conditioned on different link classes the estimation method does not change it is just repeated for each link classa wordtoword model of translational equivalence can be evaluated either over types or over tokensit is impossible to replicate the experiments used to evaluate other translation models in the literature because neither the models nor the programs that induce them are generally availablefor each kind of evaluation we have found one case where we can come closewe induced a twoclass wordtoword model of translational equivalence from 13 million words of the canadian hansards aligned using the method in one class represented contentword links and the other represented functionword links4link types with negative loglikelihood were discarded after each iterationboth classes parameters converged after six iterationsthe value of classbased models was demonstrated by the differences between the hidden parameters for the two classesc converged at for contentclass links and at for functionclass linksthe most direct way to evaluate the link types in a wordlevel model of translational equivalence is to treat each link type as a candidate translation lexicon entry and to measure precision and recallthis evaluation criterion carries much practical import because many of the applications mentioned in section 1 depend on accurate broadcoverage translation lexiconsmachine readable bilingual dictionaries even when they are available have only limited coverage and rarely include domainspecific terms we define the recall of a wordtoword translation model as the fraction of the bitext vocabulary represented in the modeltranslation model precision is a more thorny issue because people disagree about the degree to which context should play a role in judgements of translational equivalencewe handevaluated the precision of the link types in our model in the context of the bitext from which the model 4since function words can be identified by table lookup no postagger was involved was induced using a simple bilingual concordancera link type was considered correct if you and v ever cooccurred as direct translations of each otherwhere the onetoone assumption failed but a link type captured part of a correct translation it was judged quotincompletequot whether incomplete links are correct or incorrect depends on the applicationwe evaluated five random samples of 100 link types each at three levels of recallfor our bitext recall of 36 46 and 90 corresponded to translation lexicons containing 32274 43075 and 88633 words respectivelyfigure 5 shows the precision of the model with 95 confidence intervalsthe upper curve represents precision when incomplete links are considered correct and the lower when they are considered incorrecton the former metric our model can generate translation lexicons with precision and recall both exceeding 90 as well as dictionarysized translation lexicons that are over 99 correctthough some have tried it is not clear how to extract such accurate lexicons from other published translation modelspart of the difficulty stems from the implicit assumption in other models that each word has only one senseeach word is assigned the same unit of probability mass which the model distributes over all candidate translationsthe correct translations of a word that has several correct translations will be assigned a lower probability than the correct translation of a word that has only one correct translationthis imbalance foils thresholding strategies clever as they might be the likelihoods in the wordtoword model remain unnormalized so they do not competethe wordtoword model maintains high precision even given much less training dataresnik melamed report that the model produced translation lexicons with 94 precision and 30 recall when trained on frenchenglish software manuals totaling about 400000 wordsthe model was also used to induce a translation lexicon from a 6200word corpus of frenchenglish weather reportsnasr reported that the translation lexicon that our model induced from this tiny bitext accounted for 30 of the word types with precision between 84 and 90recall drops when there is less training data because the model refuses to make predictions that it cannot make with confidencefor many applications this is the desired behaviorthe most detailed evaluation of link tokens to date was performed by who trained brown et al model 2 on 74 million words of the canadian hansardsthese authors kindly provided us with the links generated by that model in 51 aligned sentences from a heldout test setwe generated links in the same 51 sentences using our twoclass wordtoword model and manually evaluated the contentword links from both modelsthe ibm models are directional ie they posit the english words that gave rise to each french word but ignore the distribution of the english wordstherefore we ignored english words that were linked to nothingthe errors are classified in table 1the quotwrong linkquot and quotmissing linkquot error categories should be selfexplanatoryquotpartial linksquot are those where one french word resulted from multiple english words but the model only links the french word to one of its english sourcesquotclass conflictquot errors resulted from our model refusal to link content words with function wordsusually this is the desired behavior but words like english auxiliary verbs are sometimes used as content words giving rise to content words in frenchsuch errors could be overcome by a model that classifies each word token for example using a partofspeech tagger instead of assigning the same class to all tokens of a given typethe bitext preprocessor for our wordtoword model split hyphenated words but macklovitch hannan preprocessor did notin some cases hyphenated words were easier to link correctly in other cases they were more difficultboth models made some errors because of this tokenization problem albeit in different placesthe quotparaphrasequot category covers all link errors that resulted from paraphrases in the translationneither ibm model 2 nor our model is capable of linking multiword sequences to multiword sequences and this was the biggest source of error for both modelsthe test sample contained only about 400 content words5 and the links for both models were evaluated posthoc by only one evaluatornevertheless it appears that our wordtoword model with only two link classes does not perform any worse than ibm model 2 even though the wordtoword model was trained on less than one fifth the amount of data that was used to train the ibm modelsince it does not store indirect associations our wordtoword model contained an average of 45 french words for every english wordsuch a compact model requires relatively little computational effort to induce and to applyin addition to the quantitative differences between the wordtoword model and the ibm model there is an important qualitative difference illustrated in figure 6as shown in table 1 the most common kind of error for the wordtoword model was a missing link whereas the most common error for ibm model 2 was a wrong linkmissing links are more informative they indicate where the model has failedthe level at which the model trusts its own judgement can be varied directly by changing the likelihood cutoff in step 1 of the competitive linking algorithmeach application of the wordtoword model can choose its own balance between link token precision and recallan application that calls on the wordtoword model to link words in a bitext could treat unlinked words differently from linked words and avoid basing subsequent decisions on uncertain inputsit is not clear how the precisionrecall tradeoff can be controlled in the ibm modelsone advantage that brown et al model 1 has over our wordtoword model is that their objective function has no local maximaby using the them algorithm they can guarantee convergence towards the globally optimum parameter setin contrast the dynamic nature of the competitive linking algorithm changes the pr in a nonmonotonic fashionwe have adopted the simple heuristic that the model quothas convergedquot when this probability stops increasingmany multilingual nlp applications need to translate words between different languages but cannot afford the computational expense of modeling the full range of translation phenomenafor these applications we have designed a fast algorithm for estimating wordtoword models of translational equivalencethe estimation method uses a pair of hidden parameters to measure the model uncertainty and avoids making decisions that it is not likely to make correctlythe hidden parameters can be conditioned on information extrinsic to the model providing an easy way to integrate preexisting knowledgeso far we have only implemented a twoclass model to exploit the differences in translation consistency between content words and function wordsthis relatively simple twoclass model linked word tokens in parallel texts as accurately as other translation models in the literature despite being trained on only one fifth as much dataunlike other translation models the wordtoword model can automatically produce dictionarysized translation lexicons and it can do so with over 99 accuracyeven better accuracy can be achieved with a more finegrained link class structurepromising features for classification include part of speech frequency of cooccurrence relative word position and translational entropy another interesting extension is to broaden the definition of a quotwordquot to include multiword lexical units if such units can be identified a priori their translations can be estimated without modifying the wordtoword modelin this manner the model can account for a wider range of translation phenomenathe frenchenglish software manuals were provided by gary adams of sun microsystems laboratoriesthe weather bitext was prepared at the university of montreal under the direction of richard kittredgethanks to alexis nasr for handevaluating the weather translation lexiconthanks also to mike collins george foster mitch marcus lyle ungar and three anonymous reviewers for helpful commentsthis research was supported by an equipment grant from sun microsystems and by arpa contract n6600194c6043	P97-1063	a wordtoword model of translational equivalencemany multilingual nlp applications need to translate words between different languages but cannot afford the computational expense of inducing or applying a full translation modelfor these applications we have designed a fast algorithm for estimating a partial translation model which accounts for translational equivalence only at the word level the model precisionrecall tradeoff can be directly controlled via one threshold parameterthis feature makes the model more suitable for applications that are not fully statisticalthe model hidden parameters can be easily conditioned on information extrinsic to the model providing an easy way to integrate preexisting knowledge such as partofspeech dictionaries word order etcour model can link word tokens in parallel texts as well as other translation models in the literatureunlike other translation models it can automatically produce dictionarysized translation lexicons and it can do so with over 99 accuracywe propose the competitive linking algorithm for linking word pairs and a method which calculates the optimized correspondence level between the word pairs by hill climbingone problem that arises in wordtoword alignment is as follows if e1 is the translation of f1 and f2 has a strong monolingual association with f1 e1 and f2 will also have a strong correlation
a memorybased approach to learning shallow natural language patterns recognizing shallow linguistic patterns such as basic syntactic relationships between words is a common task in applied natural language and text processing the common practice for approaching this task is by tedious manual definition of possible pattern structures often in the form of regular expressions or finite automata this paper presents a novel memorybased learning method that recognizes shallow patterns in new text based on a bracketed training corpus the training data are stored asis in efficient suffixtree data structures generalization is performed online at recognition time by comparing subsequences of the new text to positive and negative evidence in the corpus this way no information in the training is lost as can happen in other learning systems that construct a single generalized model at the time of training the paper presents experimental results for recognizing noun phrase subjectverb and verbobject patterns in english since the learning approach enables easy porting to new domains we plan to apply it to syntactic patterns in other languages and to sublanguage patterns for information extraction identifying local patterns of syntactic sequences and relationships is a fundamental task in natural language processing such patterns may correspond to syntactic phrases like noun phrases or to pairs of words that participate in a syntactic relationship like the heads of a verbobject relationsuch patterns have been found useful in various application areas including information extraction text summarization and bilingual alignmentsyntactic patterns are useful also for many basic computational linguistic tasks such as statistical word similarity and various disambiguation problemsone approach for detecting syntactic patterns is to obtain a full parse of a sentence and then extract the required patternshowever obtaining a complete parse tree for a sentence is difficult in many cases and may not be necessary at all for identifying most instances of local syntactic patternsan alternative approach is to avoid the complexity of full parsing and instead to rely only on local informationa variety of methods have been developed within this framework known as shallow parsing chunking local parsing etcthese works have shown that it is possible to identify most instances of local syntactic patterns by rules that examine only the pattern itself and its nearby contextoften the rules are applied to sentences that were tagged by partofspeech and are phrased by some form of regular expressions or finite state automatamanual writing of local syntactic rules has become a common practice for many applicationshowever writing rules is often tedious and time consumingfurthermore extending the rules to different languages or sublanguage domains can require substantial resources and expertise that are often not availableas in many areas of nlp a learning approach is appealingsurprisingly though rather little work has been devoted to learning local syntactic patterns mostly noun phrases this paper presents a novel general learning approach for recognizing local sequential patterns that may be perceived as falling within the memorybased learning paradigmthe method utilizes a partofspeech tagged training corpus in which all instances of the target pattern are marked the training data are stored asis in suffixtree data structures which enable linear time searching for subsequences in the corpusthe memorybased nature of the presented algorithm stems from its deduction strategy a new instance of the target pattern is recognized by examining the raw training corpus searching for positive and negative evidence with respect to the given test sequenceno model is created for the training corpus and the raw examples are not converted to any other representationconsider the following examplel suppose we want to decide whether the candidate sequence is a noun phrase by comparing it to the training corpusa good match would be if the entire sequence appears asis several times in the corpushowever due to data sparseness an exact match cannot always be expecteda somewhat weaker match may be obtained if we consider subparts of the candidate sequence for example suppose the corpus contains noun phrase instances with the following structures the first structure provides positive evidence that the sequence quotdt adj adj nmquot is a possible np prefix while the second structure provides evidence for quotadj nn nnpquot being an np suffixtogether these two training instances provide positive evidence that covers the entire candidateconsidering evidence for subparts of the pattern enables us to generalize over the exact structures that are present in the corpussimilarly we also consider the negative evidence for such subparts by noting where they occur in the corpus without being a corresponding part of a target instancethe proposed method as described in detail in the next section formalizes this type of reasoningit searches specialized data structures for both positive and negative evidence for subparts of the candidate structure and considers additional factors such as context and evidence overlapsection 3 presents experimental results for three target syntactic patterns in english and section 4 describes related workthe input to the memorybased sequence learning algorithm is a sentence represented as a sequence of pos tags and its output is a bracketed sentence indicating which subsequences of the sentence are to be considered instances of the target pattern mbsl determines the bracketing by first considering each subsequence of the sentence as a candidate to be a target instanceit computes a score for each candidate by comparing it to the training corpus which consists of a set of prebracketed sentencesthe algorithm then finds a consistent bracketing for the input sentence giving preference to high scoring subsequencesin the remainder of this section we describe the scoring and bracketing methods in more detailwe first describe the mechanism for scoring an individual candidatethe input is a candidate subsequence along with its context ie the other tags in the input sentencethe method is presented at two levels a general memorybased learning schema and a particular instantiation of itfurther instantiations of the schema are expected in future workthe mbsl scoring algorithm works by considering situated candidatesa situated candidate is a sentence containing one pair of brackets indicating a candidate to be a target instancethe portion of the sentence between the brackets is the candidate while the portion before and after the candidate is its contextthis subsection describes how to compute the score of a situated candidate from the training corpusthe idea of the mbsl scoring algorithm is to construct a tiling of subsequences of a situated candidate which covers the entire candidatewe consider as tiles subsequences of the situated candidate which contain a bracketeach tile is assigned a score based on its occurrence in the training memorysince brackets correspond to the boundaries of potential target instances it is important to consider how the bracket positions in the tile correspond to those in the training memoryfor example consider the training sentence nn vb adj nn nn adv pp nn we may now examine the occurrence in this sentence of several possible tiles vb adj nn occurs positively in the sentence and nn nn adv also occurs positively while nn nn adv occurs negatively in the training sentence since the bracket does not correspondthe positive evidence for a tile is measured by its positive count the number of times the tile occurs in the training memory with corresponding bracketssimilarly the negative evidence for a tile is measured by its negative count the number of times that the pos sequence of the tile occurs in the training memory with noncorresponding brackets the total count of a tile is its positive count plus its negative count that is the total count of the pos sequence of the tile regardless of bracket positionthe score f of a tile t is a function of its positive and negative countsthe overall score of a situated candidate is generally a function of the scores of all the tiles for the candidate as well as the relations between the tiles positionsthese relations include tile adjacency overlap between tiles the amount of context in a tile and so onin our instantiation of the mbsl schema we define the score f of a tile t as the ratio of its positive count pos and its total count total for a predefined threshold 0tiles with a score of 1 and so with sufficient positive evidence are called matching tileseach matching tile gives supporting evidence that a part of the candidate can be a part of a target instancein order to combine this evidence we try to cover the entire candidate by a set of matching tiles with no gapssuch a covering constitutes evidence that the entire candidate is a target instancefor example consider the matching tiles shown for the candidate in figure 1the set of matching tiles 2 4 and 5 covers the candidate as does the set of tiles 1 and 5also note that tile 1 constitutes a cover on its ownto make this precise we first say that a tile t1 connects to a tile 72 if 22 starts after t1 starts there is no gap between the end of t1 and the start of 72 and t2 ends after t1 for example tiles 2 and 4 in the figure connect while tiles 2 and 5 do not and neither do tiles 1 and 4 a cover for a situated candidate c is a sequence of matching tiles which collectively cover the entire candidate including the boundary brackets and possibly some context such that each tile connects to the following onea cover thus provides positive evidence for the entire sequence of tags in the candidatethe set of all the covers for a candidate summarizes all of the evidence for the candidate being a target instancewe therefore compute the score of a candidate as a function of some statistics of the set of all its coversfor example if a candidate has many different covers it is more likely to be a target instance since many different pieces of evidence can be brought to bearwe have empirically found several statistics of the cover set to be usefulthese include for each cover the number of tiles it contains the total number of context tags it contains and the number of positions which more than one tile covers we thus compute for the set of all covers of a candidate c the each of these items gives an indication regarding the overall strength of the coverbased evidence for the candidatethe score of the candidate is a linear function of its statistics if candidate c has no covers we set f 0note that minsize is weighted negatively since a cover with fewer tiles provides stronger evidence for the candidatein the current implementation the weights were chosen so as to give a lexicographic ordering preferring first candidates with more covers then those with covers containing fewer tiles then those with larger contexts and finally when all else is equal preferring candidates with more overlap between tileswe plan to investigate in the future a datadriven approach for optimal selection and weighting of statistical features of the scorewe compute a candidate statistics efficiently by performing a depthfirst traversal of the cover graph of the candidatethe cover graph is a directed acyclic graph whose nodes represent matching tiles of the candidate such that an arc exists between nodes n and n if tile n connects to na special start node is added as the root of the dag that connects to all of the nodes that contain an open bracketthere is a cover corresponding to each path from the start node to a node that contains a close bracketthus the statistics of all the covers may be efficiently computed by traversing the cover graphthe mbsl scoring algorithm searches the training corpus for each subsequence of the sentence in order to find matching tilesimplementing this search efficiently is therefore of prime importancewe do so by encoding the training corpus using suffix trees which provide string searching in time which is linear in the length of the searched stringinspired by satta we build two suffix trees for retrieving the positive and total counts for a tilethe first suffix tree holds all pattern instances from the training corpus surrounded by bracket symbols and a fixed amount of contextsearching a given tile in this tree yields the positive count for the tilethe second suffix tree holds an unbracketed version of the entire training corpusthis tree is used for searching the pos sequence of a tile with brackets omitted yielding the total count for the tile after the above procedure each situated candidate is assigned a scorein order to select a bracketing for the input sentence we assume that target instances are nonoverlapping we use a simple constraint propagation algorithm that finds the best choice of nonoverlapping candidates in an input sentence ber of patterns and average length in the training datawe have tested our algorithm in recognizing three syntactic patterns noun phrase sequences verbobject and subjectverb relationsthe np patterns were delimited by p and symbols at the borders of the phrasefor vo patterns we have put the starting delimiter before the main verb and the ending delimiter after the object head thus covering the whole noun phrase comprising the object for example investigators started to view the lower price levels 7 as attractive we used a similar policy for sv patterns defining the start of the pattern at the start of the subject noun phrase and the end at the first verb encountered for example argue that the yous should regulate the class the subject and object nounphrase borders were those specified by the annotators phrases which contain conjunctions or appositives were not further analyzedthe training and testing data were derived from the penn treebankwe used the np data prepared by ramshaw and marcus hereafter rm95the sv and vo data were obtained using t scripts2 table 1 summarizes the sizes of the training and test data sets and the number of examples in eachthe t scripts did not attempt to match dependencies over very complex structures since we are concerned with shallow or local patternstable 2 shows the distribution of pattern length in the train datawe also did not attempt to extract passivevoice vo relationsthe test procedure has two parameters maximum context size of a candidate which limits what queries are performed on the memory and the threshold 9 used for establishing a matching tile which determines how to make use of the query resultsrecall and precision figures were obtained for various parameter valuesfo a common measure in information retrieval was used as a singlefigure measure of performance we use 1 which gives no preference to either recall or precisiontable 3 summarizes the optimal parameter settings and results for np vo and sv on the test setin order to find the optimal values of the context size and threshold we tried 01 9 095 and maximum context sizes of 12 and 3our experiments used 5fold crossvalidation on the training data to determine the optimal parameter settingsin experimenting with the maximum context size parameter we found that the difference between the values of fo for context sizes of 2 and 3 is less than 05 for the optimal thresholdscores for a context size of 1 yielded fo values smaller by more than 1 than the values for the larger contextsfigure 2 shows recallprecision curves for the three data sets obtained by varying 9 while keeping the maximum context size at its optimal valuethe difference between fo1 values for different thresholds was always less than 2performance may be measured also on a wordby word basis counting as a success any word which was identified correctly as being part of the target patternthat method was employed along with recallprecision by rm95we preferred to measure performance by recall and precision for complete patternsmost errors involved identifications of slightly shifted shorter or longer sequencesgiven a pattern consisting of five words for example identifying only a fourword portion of this pattern would yield both a recall and precision errorstagassignment scoring on the other hand will give it a score of 80we hold the view that such an identification is an error rather than a partial successwe used the datasets created by rm95 for np learning their results are shown in table 33 the fo difference is small yet they use a richer feature set which incorporates lexical information as wellthe method of ramshaw and marcus makes a decision per word relying on predefined rule templatesthe method presented here makes decisions on sequences and uses sequences as its memory thereby attaining a dynamic perspective of the last line shows the results of ramshaw and marcus with the same traintest datathe optimal parameters were obtained by 5fold crossvalidation pattern structurewe aim to incorporate lexical information as well in the future it is still unclear whether that will improve the resultsfigure 3 shows the learning curves by amount of training examples and number of words in the training data for particular parameter settingstwo previous methods for learning local syntactic patterns follow the transformationbased paradigm introduced by brill vilain and day identify name phrases such as company names locations etcramshaw and marcus detect noun phrases by classifying each word as being inside a phrase outside or on the boundary between phrasesfinite state machines are a natural formalism for learning linear sequencesit was used for learning linguistic structures other than shallow syntaxgold showed that learning regular languages from positive examples is undecidable in the limitrecently however several learning methods have been proposed for restricted classes of fsmostia learns a subsequential transducer in the limitthis algorithm was used for naturallanguage tasks by vilar marzal and vidal for learning translation of a limiteddomain language as well as by gildea and jurafsky for learning phonological rulesahonen et al describe an algorithm for learning contextual regular languages which they use for learning the structure of sgml documentsapart from deterministic fsms there are a number of algorithms for learning stochastic models eg these algorithms differ mainly by their statemerging strategies used for generalizing from the training dataa major difference between the abovementioned learning methods and our memorybased approach is that the former employ generalized models that were created at training time while the latter uses the training corpus asis and generalizes only at recognition timemuch work aimed at learning models for full parsing ie learning hierarchical structureswe refer here only to the dop method which like the present work is a memorybased approachthis method constructs parse alternatives for a sentence based on combinations of subtrees in the training corpusthe mbsl approach may be viewed as a linear analogy to dop in that it constructs a cover for a candidate based on subsequences of training instancesother implementations of the memorybased paradigm for nlp tasks include daelemans et al for pos tagging cardie for syntactic and semantic tagging and stanfill and waltz for word pronunciationin all these works examples are represented as sets of features and the deduction is carried out by finding the most similar casesthe method presented here is radically different in that it makes use of the raw sequential form of the data and generalizes by reconstructing test examples from different pieces of the training datawe have presented a novel general schema and a particular instantiation of it for learning sequential patternsapplying the method to three syntactic patterns in english yielded positive results suggesting its applicability for recognizing local linguistic patternsin future work we plan to investigate a datadriven approach for optimal selection and weighting of statistical features of candidate scores as well as to apply the method to syntactic patterns of hebrew and to domainspecific patterns for information extractionthe authors wish to thank yoram singer for his collaboration in an earlier phase of this research project and giorgio satta for helpful discussionswe aiso thank the anonymous reviewers for their instructive commentsthis research was supported in part by grant 498951 from the israel science foundation and by grant 8560296 from the israeli ministry of science	P98-1010	a memorybased approach to learning shallow natural language patternsrecognizing shallow linguistic patterns such as basic syntactic relationships between words is a common task in applied natural language and text processingthe common practice for approaching this task is by tedious manual definition of possible pattern structures often in the form of regular expressions or finite automatathis paper presents a novel memorybased learning method that recognizes shallow patterns in new text based on a bracketed training corpusthe training data are stored asis in efficient suffixtree data structuresgeneralization is performed online at recognition time by comparing subsequences of the new text to positive and negative evidence in the corpusthis way no information in the training is lost as can happen in other learning systems that construct a single generalized model at the time of trainingthe paper presents experimental results for recognizing noun phrase subjectverb and verbobject patterns in englishsince the learning approach enables easy porting to new domains we plan to apply it to syntactic patterns in other languages and to sublanguage patterns for information extractionwe segment the pos sequence of a multiword into small pos titles count tile frequency in the new word and nonnewword on the training set respectively and detect new words using these counts
entitybased crossdocument core f erencing using the vector space model crossdocument coreference occurs when the same person place event or concept is discussed in more than one text source computer recognition of this phenomenon is important because it helps break quotthe document boundaryquot by allowing a user to examine information about a particular entity from multiple text sources at the same time in this paper we describe a crossdocument coreference resolution algorithm which uses the vector space model to resolve ambiguities between people having the same name in addition we also describe a scoring algorithm for evaluating the crossdocument coreference chains produced by our system and we compare our algorithm to the scoring algorithm used in the muc 6 coreference task crossdocument coreference occurs when the same person place event or concept is discussed in more than one text sourcecomputer recognition of this phenomenon is important because it helps break quotthe document boundaryquot by allowing a user to examine information about a particular entity from multiple text sources at the same timein particular resolving crossdocument coreferences allows a user to identify trends and dependencies across documentscrossdocument coreference can also be used as the central tool for producing summaries from multiple documents and for information fusion both of which have been identified as advanced areas of research by the tipster phase iii programcrossdocument coreference was also identified as one of the potential tasks for the sixth message understanding conference but was not included as a formal task because it was considered too ambitious in this paper we describe a highly successful crossdocument coreference resolution algorithm which uses the vector space model to resolve ambiguities between people having the same namein addition we also describe a scoring algorithm for evaluating the crossdocument coreference chains produced by our system and we compare our algorithm to the scoring algorithm used in the muc6 coreference taskcrossdocument coreference is a distinct technology from named entity recognizers like isoquest netowl and ibm textract because it attempts to determine whether name matches are actually the same individual neither netowl or textract have mechanisms which try to keep samenamed individuals distinct if they are different peoplecrossdocument coreference also differs in substantial ways from withindocument coreferencewithin a document there is a certain amount of consistency which cannot be expected across documentsin addition the problems encountered during within document coreference are compounded when looking for coreferences across documents because the underlying principles of linguistics and discourse context no longer apply across documentsbecause the underlying assumptions in crossdocument coreference are so distinct they require novel approachesfigure 1 shows the architecture of the crossdocument system developedthe system is built upon the university of pennsylvania within document coreference system camp which participated in the seventh message understanding conference within document coreference task our system takes as input the coreference processed documents output by campit then passes these documents through the sentenceextractor module which extracts for each document all the sentences relevant to a particular entity of interestthe vsmdisambiguate module then uses a vector space model algorithm to compute similarities between the sentences extracted for each pair of documentsoliver quotbiffquot kelly of weymouth succeeds john perry as president of the massachusetts golf associationquotwe will have continued growth in the futurequot said kelly who will serve for two yearsquotthere is been a lot of changes and there will be continued changes as we head into the year 2000quot details about each of the main steps of the crossdocument coreference algorithm are given below consider the two extracts in figures 2 and 4the coreference chains output by camp for the two extracts are shown in figures 3 and 5 tractor module produces a quotsummaryquot of the article with respect to the entity of interestthese summaries are a special case of the query sensitive techniques being developed at penn using camptherefore for doc36 since at least one of the three noun phrases in the coreference chain of interest appears in each of the three sentences in the extract the summary produced by sentenceextractor is the extract itselfon the other hand the summary produced by sentenceextractor for the coreference chain of interest in doc38 is only the first sentence of the extract because the only element of the coreference chain appears in this sentencethe university of pennsylvania camp system resolves within document coreferences for several different classes including pronouns and proper names it ranked among the top systems in the coreference task during the muc6 and the muc7 evaluationsthe coreference chains output by camp enable us to gather all the information about the entity of interest in an articlethis information about the entity is gathered by the sentenceextractor module and is used by the vsmdisambiguate module for disambiguation purposesconsider the extract for doc36 shown in figure 2we are able to include the fact that the john perry mentioned in this article was the president of the massachusetts golf association only because camp recognized that the quothequot in the second sentence is coreferent with quotjohn perryquot in the firstand it is this fact which actually helps vsmdisambiguate decide that the two john perrys in doc36 and doc38 are the same personthe vector space model used for disambiguating entities across documents is the standard vector space model used widely in information retrieval in this model each summary extracted by the sentenceextractor module is stored as a vector of termsthe terms in the vector are in their morphological root form and are filtered for stopwords if si and s2 are the vectors for the two summaries extracted from documents d1 and d21 then their similarity is computed as where tj is a term present in both si and s2 w1j is the weight of the term t3 in s1 and w23 is the weight of ti in s2the weight of a term tj in the vector st for a summary is given by where t f is the frequency of the term t3 in the summary n is the total number of documents in the collection being examined and df is the number of documents in the collection that the term tj occurs inmi 42 4n is the cosine normalization factor and is equal to the euclidean length of the vector sithe vsmdisambiguate module for each summary si computes the similarity of that summary with each of the other summariesif the similarity computed is above a predefined threshold then the entity of interest in the two summaries are considered to be coreferentthe crossdocument coreference system was tested on a highly ambiguous test set which consisted of 197 articles from 1996 and 1997 editions of the new york timesthe sole criteria for including an article in the test set was the presence or the absence of a string in the article which matched the quotjohnsmithr regular expressionin other words all of the articles either contained the name john smith or contained some variation with a middle initialnamethe system did not use any new york times data for training purposesthe answer keys regarding the crossdocument chains were manually created but the scoring was completely automatedthere were 35 different john smiths mentioned in the articlesof these 24 of them only had one article which mentioned themthe other 173 articles were regarding the 11 remaining john smithsthe background of these john smiths and the number of articles pertaining to each varied greatlydescriptions of a few of the john smiths are chairman and ceo of general motors assistant track coach at ucla the legendary explorer and the main character in disney pocahontas former president of the labor party of britainin order to score the crossdocument coreference chains output by the system we had to map the crossdocument coreference scoring problem to a withindocument coreference scoring problemthis was done by creating a meta document consisting of the file names of each of the documents that the system was run onassuming that each of the documents in the data set was about a single john smith the crossdocument coreference chains produced by the system could now be evaluated by scoring the corresponding withindocument coreference chains in the meta documentwe used two different scoring algorithms for scoring the outputthe first was the standard algorithm for withindocument coreference chains which was used for the evaluation of the systems participating in the muc6 and the muc7 coreference tasksthe shortcomings of the muc scoring algorithm when used for the crossdocument coreference task forced us to develop a second algorithmdetails about both these algorithms followthe muc algorithm computes precision and recall statistics by looking at the number of links identified by a system compared to the links in an answer keyin the modeltheoretic description of the algorithm that follows the term quotkeyquot refers to the manually annotated coreference chains while the term quotresponsequot refers to the coreference chains output by a systeman equivalence set is the transitive closure of a coreference chainthe algorithm developed by computes recall in the following wayfirst let s be an equivalence set generated by the key and let r1 r be equivalence classes generated by the responsethen we define the following functions over s fully reunite any components of the p partitionwe note that this is simply one fewer than the number of elements in the partition that is m i 1 looking in isolation at a single equivalence class in the key the recall error for that class is just the number of missing links divided by the number of correct links ie precision is computed by switching the roles of the key and response in the above formulationwhile the provides intuitive results for coreference scoring it however does not work as well in the context of evaluating cross document coreferencethere are two main reasons1the algorithm does not give any credit for separating out singletons from other chains which have been identifiedthis follows from the convention in coreference annotation of not identifying those entities that are markable as possibly coreferent with other entities in the textrather entities are only marked as being coreferent if they actually are coreferent with other entities in the textthis shortcoming could be easily enough overcome with different annotation conventions and with minor changes to the algorithm but it is worth noting2all errors are considered to be equalthe muc scoring algorithm penalizes the precision numbers equally for all types of errorsit is our position that for certain tasks some coreference errors do more damage than othersconsider the following examples suppose the truth contains two large coreference chains and one small one and suppose figures 7 and 8 show two different responseswe will explore two different precision errorsthe first error will connect one of the large coreference chains with the small one the second error occurs when the two large coreference chains are related by the errant coreferent link it is our position that the second error is more damaging because compared to the first error the second error makes more entities coreferent that should not bethis distinction is not reflected in the scorer which scores both responses as having a precision score of 90 imagine a scenario where a user recalls a collection of articles about john smith finds a single article about the particular john smith of interest and wants to see all the other articles about that individualin commercial systems with news data precision is typically the desired goal in such settingsas a result we wanted to model the accuracy of the system on a perdocument basis and then build a more global score based on the sum of the user experiencesconsider the case where the user selects document 6 in figure 8this a good outcome with all the relevant documents being found by the system and no extraneous documentsif the user selected document 1 then there are 5 irrelevant documents in the systems output precision is quite low thenthe goal of our scoring algorithm then is to model the precision and recall on average when looking for more documents about the same person based on selecting a single documentinstead of looking at the links produced by a system our algorithm looks at the presenceabsence of entities from the chains producedtherefore we compute the precision and recall numbers for each entity in the documentthe numbers computed with respect to each entity in the document are then combined to produce final precision and recall numbers for the entire outputfor an entity i we define the precision and recall with respect to that entity in figure 10the final precision and recall numbers are computed by the following two formulae final precision e wi precisioni final recall e wi recalls where n is the number of entities in the document and wi is the weight assigned to entity i in the documentfor all the examples and the experiments in this paper we assign equal weights to each entity ie wi 1nwe have also looked at the possibilities of using other weighting schemesfurther details about the bcubed algorithm including a model theoretic version of the algorithm can be found in consider the response shown in figure 7using the bcubed algorithm the precision for entity6 in the document equals 27 because the chain output for the entity contains 7 elements 2 of which are correct namely 67the recall for entity6 however is 22 because the chain output for the entity has 2 correct elements in it and the quottruthquot chain for the entity only contains those 2 elementsfigure 9 shows the final precision and recall numbers computed by the bcubed algorithm for the examples shown in figures 7 and 8the figure also shows the precision and recall numbers for each entity the bcubed algorithm does overcome the the two main shortcomings of the muc scoring algorithm discussed earlierit implicitly overcomes the first number of correct elements in the output chain containing entityi number of elements in the output chain containing entity number of correct elements in the output chain containing entity number of elements in the truth chain containing entity figure 10 definitions for precision and recall for an entity i shortcoming of the muc6 algorithm by calculating the precision and recall numbers for each entity in the document consider the responses shown in figures 7 and 8we had mentioned earlier that the error of linking the the two large chains in the second response is more damaging than the error of linking one of the large chains with the smaller chain in the first responseour scoring algorithm takes this into account and computes a final precision of 58 and 76 for the two responses respectivelyin comparison the muc algorithm computes a precision of 90 for both the responses figure 11 shows the precision recall and fmeasure using the bcubed scoring algorithmthe vector space model in this case constructed the space of terms only from the summaries extracted by sentenceextractorin comparison figure 12 shows the results when the vector space model constructed the space of terms from the articles input to the system the importance of using camp to extract summaries is verified by comparing the highest fmeasures achieved by the system for the two casesthe highest fmeasure for the former case is 846 while the highest fmeasure for the latter case is 780in comparison for this task namedentity tools like netowl and textract would mark all the john smiths the sametheir performance using our threshold figure 11 precision recall and fmeasure using the bcubed algorithm with training on the summaries scoring algorithm is 23 precision and 100 recallfigures 13 and 14 show the precision recall and fmeasure calculated using the muc scoring algorithmalso the baseline case when all the john smiths are considered to be the same person achieves 83 precision and 100 recallthe high initial precision is mainly due to the fact that the muc algorithm assumes that all errors are equalwe have also tested our system on other classes of crossdocument coreference like names of companies and eventsdetails about these experiments can be found in as a novel research problem cross document coreference provides an different perspective from related phenomenon like named entity recognition and within document coreferenceour system takes summaries about an entity of interest and uses various information retrieval metrics to rank the similarity of the summarieswe found it quite challenging to arrive at a scoring metric that satisfied our intuitions about what was good system output vs bad but we have developed a scoring algorithm that is an improvement for this class of data over other within document coreference scoring algorithmsour results are quite encouraging with potential performance being as good as 846 the first author was supported in part by a fellowship from ibm corporation and in part by the institute for research in cognitive science at the university of pennsylvania	P98-1012	entitybased crossdocument core f erencing using the vector space modelcrossdocument coreference occurs when the same person place event or concept is discussed in more than one text sourcecomputer recognition of this phenomenon is important because it helps break the document boundary by allowing a user to examine information about a particular entity from multiple text sources at the same timein this paper we describe a crossdocument coreference resolution algorithm which uses the vector space model to resolve ambiguities between people having the same namein addition we also describe a scoring algorithm for evaluating the crossdocument coreference chains produced by our system and we compare our algorithm to the scoring algorithm used in the muc6 coreference taskwe proposed entitybased crossdocument coreferencing which uses coreference chains of each document to generate its summary and then use the summary rather than the whole article to select informative words to be the features of the document
the berkeley framenet project is a nsfsupported project in corpusbased computational lexicography now in its second year the project key features are a commitment to corpus evidence for semantic and syntactic generalizations and the representation of the valences of its target words in which the semantic portion makes use of frame semantics the resulting database will contain descriptions of the semantic frames underlying the meanings of the words described and the valence representation of several thousand words and phrases each accompanied by a representative collection of annotated corpus attestations which jointly exemplify the observed linkings between quotframe elementsquot and their syntactic realizations this report will present the project goals and workflow and information about the computational tools that have been adapted or created inhouse for this work the berkeley framenet project is producing framesemantic descriptions of several thousand english lexical items and backing up these descriptions with semantically annotated attestations from contemporary english corpora2these descriptions are based on handtagged semantic annotations of example sentences extracted from large text corpora and systematic analysis of the semantic patterns they exemplify by lexicographers and linguiststhe primary emphasis of the project therefore is the encoding by humans of semantic knowledge in machinereadable formthe intuition of the lexicographers is guided by and constrained by the results of corpusbased research using highperformance software toolsthe semantic domains to be covered are health care chance perception communication transaction time space body motion life stages social context emotion and cognitionthe results of the project are a lexical resource called the framenet database3 and associated software toolsthe database has three major components links to the frame database and to other machinereadable resources such as wordnet and comlex marked up to exemplify the semantic and morphosyntactic properties of the lexical itemsthese sentences provide empirical support for the lexicographic analysis provided in the frame database and lexicon entriesthese three components form a highly relational and tightly integrated whole elements in each may point to elements in the other twothe database will also contain estimates of the relative frequency of senses and complementation patterns calculated by matching the senses and patterns in the handtagged examples against the entire bnc corpusthe framenet work is in some ways similar to efforts to describe the argument structures of lexical items in terms of caseroles or thetaroles5 but in framenet the role names are local to particular conceptual structures some of these are quite general while others are specific to a small family of lexical itemsfor example the transportation frame within the domain of motion provides movers means of transportation and paths6 6the semantic frames for individual lexical units are typically quotblendsquot of more than one basic frame from our point of view the socalled quotlinkingquot patterns proposed in lfg hpsg and construction grammar operate on higherlevel frames of action motion and location and experience etcin some but not all cases the assignment of syntactic correlates to frame elements could be mediated by mapping them to the roles of one of the more abstract frames6a detailed study of motion predicates would require a finergrained analysis of the path element separating out source and goal and perhaps direction and area but for a basic study of the transportation predicates such refined analysis is not necessaryin any case our subframes associated with individual words inherit all of these while possibly adding some of their ownfig1 shows some of the subframes as discussed belowthe driving frame for example specifies a driver a vehicle and potentially cargo or rider as secondary moversin this frame the driver initiates and controls the movement of the vehiclefor most verbs in this frame driver or vehicle can be realized as subjects vehicle rider or cargo can appear as direct objects and path and vehicle can appear as oblique complementssome combinations of frame elements or frame element groups for some real corpus sentences in the driving frame are shown in fig2a riding_1 frame has the primary mover role as rider and allows as vehicle those driven by others7 in grammatical realizations of this frame the rider can be the subject the vehicle can appear as a direct object or an oblique complement and the path is generally realized as an obliquethe framenet entry for each of these verbs will include a concise formula for all semanwork includes the separate analysis of the frame semantics of directional and locational expressions tic and syntactic combinatorial possibilities together with a collection of annotated corpus sentences in which each possibility is exemplifiedthe syntactic positions considered relevant for lexicographic description include those that are internal to the maximal projection of the target word and those that are external to the maximal projection under precise structural conditions the subject in the case of vp and the subject of support verbs in the case of ap and np8 used in nlp the framenet database should make it possible for a system which finds a valencebearing lexical item in a text to know where its individual arguments are likely to be foundfor example once a parser has found the verb drive and its direct object np the link to the driving frame will suggest some semantics for that np eg that a person as direct object probably represents the rider while a nonhuman proper noun is probably the vehiclefor practical lexicography the contribution of the framenet database will be its presentation of the full range of use possibilities for individual words documented with corpus data the model examples for each use and the statistical information on relative frequencythe computational side of the framenet project is directed at efficiently capturing human insights into semantic structurethe majority of the work involved is marking text with semantic tags specifying the structure of the frames to be treated and writing dictionarystyle entries based the results of annotation and a priori descriptionswith the exception of the example sentence extraction component all the software modules are highly interactive and have substantial user interface requirementsmost of this functionality is provided by wwwbased programs written in perlfour processing steps are required produce the framenet database of frame semantic representations generating initial descriptions of semantic and syntactic patterns for use in corpus queries and annotation extracting good example sentences marking the constituents of interest and building a database of lexical semantic representations based on the annotations and other data these are discussed briefly below and shown in fig3as work on the project has progressed we have defined several explicit roles which project participants play in the various steps these roles are referred to as vanguard annotators and rearguard these are purely functional designations the same person may play different roles at different times9 pares the initial descriptions of frames including lists of frames and frame elements and adds these to the frame database using the frame description tool the vanguard also selects the major vocabulary items for the frame and the syntactic patterns that need to be checked for each word which are entered in the lexical database by means of the lexical database tool 2subcorpus extractionbased on the vanguard work the subcorpus extraction tools produce a representative collection of sentences containing these wordsthis selection of examples is achieved through a hybrid process partially controlled by the preliminary lexical description of each lemmasentences containing the lemma are extracted from from a corpus and classified into subcorpora by syntactic pattern using a cascade filter representing a partial regularexpression grammar of english over partofspeech tags formatted for annotation and automatically sampled down to an appropriate number and the fegs extracted from them and builds both the entries for the lemmas in the lexical database and the frame descriptions in the frame database using the entry writing tools we are building a quotconstituent type identifierquot which will semiautomatically assign grammatical function and phrase type attributes to these femarked constituents eliminating the need for annotators to mark thesethe data structures described above are implemented in sgml11 each is described by a dtd and these dtds are structured to provide the necessary links between the componentsthe software suite currently supporting database development is an aggregate of existing software tools held together with perlcgibased quotgluequotin order to get the project started we have depended on offtheshelf software which in some cases is not ideal for our purposesnevertheless using these programs allowed us to get the project up and running within just a few monthswe describe below in approximate order of application the programs used and their state of completionquoteventually we plan to migrate to an xml data model which appears to provide more flexibility while reducing complexityalso the framenet software is being developed on unix but we plan to provide crossplatform capabilities by making our tool suite webbased and xmlcompatiblesgml files into html for convenient viewing on the web etc are being written in perlrcs maintains version control over most filesat the time of writing there is something in place for each of the major software components though in some cases these are little more than stubs or quottoyquot implementationsnearly 10000 sentences exemplifying just under 200 lemmas have been annotated there are over 20000 frame element tokens marked in these example sentencesabout a dozen frames have been specified which refer to 47 named frame elementsmost of these annotations have been accomplished in the last few months since the software for corpus extraction frame description and annotation became operationalwe expect the inventory to increase rapidlyif the proportions cited hold constant as the framenet database grows the final database of 5000 lexical units may contain 250000 annotated sentences and over half a million tokens of frame elements	P98-1013	the berkeley framenet projectframenet is a threeyear nsfsupported project in corpusbased computational lexicography now in its second year the project key features are a commitment to corpus evidence for semantic and syntactic generalizations and the representation of the valences of its target words in which the semantic portion makes use of frame semanticsthe resulting database will contain descriptions of the semantic frames underlying the meanings of the words described and the valence representation of several thousand words and phrases each accompanied by a representative collection of annotated corpus attestations which jointly exemplify the observed linkings between frame elements and their syntactic realizations this report will present the project goals and workflow and information about the computational tools that have been adapted or created inhouse for this workwe present the framenet project in which we havee been developing a framesemantic lexicon for the core vocabulary of english
classifier combination for improved lexical disambiguation one of the most exciting recent directions in machine learning is the discovery that the combination of multiple classifiers often results in significantly better performance than what can be achieved with a single classifier in this paper we first show that the errors made from three different state of the art part of speech taggers are strongly complementary next we show how this complementary behavior can be used to our advantage by using contextual cues to guide tagger combination we are able to derive a new tagger that achieves performance significantly greater than any of the individual taggers part of speech tagging has been a central problem in natural language processing for many yearssince the advent of manually tagged corpora such as the brown corpus and the penn treebank marcus the efficacy of machine learning for training a tagger has been demonstrated using a wide array of techniques including markov models decision trees connectionist machines transformations nearestneighbor algorithms and maximum entropy black schmid brilldaelemansratnaparkhiall of these methods seem to achieve roughly comparable accuracythe fact that most machinelearningbased taggers achieve comparable results could be attributed to a number of causesit is possible that the 8020 rule of engineering is applying a certain number of tagging instances are relatively simple to disambiguate and are therefore being successfully tagged by all approaches while another percentage is extremely difficult to disambiguate requiring deep linguistic knowledge thereby causing all taggers to erranother possibility could be that all of the different machine learning techniques are essentially doing the same thingwe know that the features used by the different algorithms are very similar typically the words and tags within a small window from the word being taggedtherefore it could be possible that they all end up learning the same information just in different formsin the field of machine learning there have been many recent results demonstrating the efficacy of combining classifiersin this paper we explore whether classifier combination can result in an overall improvement in lexical disambiguation accuracythe experiments described in this paper are based on four popular tagging algorithms all of which have readily available implementationsthese taggers are described belowthis is by far the simplest of tagging algorithmsevery word is simply assigned its most likely part of speech regardless of the context in which it appearssurprisingly this simple tagging method achieves fairly high accuracyaccuracies of 9094 are typicalin the unigram tagger used in our experiments for words that do not appear in the lexicon we use a i see dietterich for a good summary of these techniques collection of simple manuallyderived heuristics to guess the proper tag for the wordngram part of speech taggers church weischedel are perhaps the most widely used of tagging algorithmsthe basic model is that given a word sequence w we try to find the tag sequence t that maximizes pthis can be done using the viterbi algorithm to find the t that maximizes ppin our experiments we use a standard trigram tagger using deleted interpolation and used suffix information for handling unseen words in transformationbased tagging every word is first assigned an initial tag this tag is the most likely tag for a word if the word is known and is guessed based upon properties of the word if the word is not knownthen a sequence of rules are applied that change the tags of words based upon the contexts they appear inthese rules are applied deterministically in the order they appear in the listas a simple example if race appears in the corpus most frequently as a noun it will initially be mistagged as a noun in the sentence we can race all day longthe rule change a tag from noun to verb if the previous tag is a modal would be applied to the sentence resulting in the correct taggingthe environments used for changing a tag are the words and tags within a window of three wordsfor our experiments we used a publicly available implementation of transformationbased tagging2 retrained on our training setthe maximumentropy framework is a probabilistic framework where a model is found that is consistent with the observed data and is maximally agnostic with respect to all parameters for which no data existsit is a nice framework for combining multiple constraintswhereas the transformationbased tagger enforces multiple constraints by having multiple rules fire the maximumentropy tagger can have all of these constraints play a role at setting the probability estimates for the model parametersin ratnaparkhi a maximum entropy tagger is presentedthe tagger uses essentially the same parameters as the transformationbased tagger but employs them in a different modelfor our experiments we used a publicly available implementation of maximumentropy tagging3 retrained on our training setall experiments presented in this paper were run on the penn treebank wall street journal corpus the corpus was divided into approximately 80 training and 20 testing giving us approximately 11 million words of training data and 265000 words of test datathe test set was not used in any way in training so the test set does contain unknown wordsin figure 1 we show the relative accuracies of the four taggersin parentheses we include tagger accuracy when only ambiguous and unknown words are considered4 next we examine just how different the errors of the taggers arewe define the complementary rate of taggers a and b as ambiguity resolution over all words including words that are unambiguouscorrectly tagging words that only can have one label contributes to the accuracywe see in figure 1 that when accuracy is measured on truly ambiguous words the numbers are lowerin this paper we stick to the convention of giving results for all words including unambiguous ones of errors in a only in other words comp measures the percentage of time when tagger a is wrong that tagger b is correctin figure 2 we show the complementary rates between the different taggersfor instance when the maximum entropy tagger is wrong the transformationbased tagger is right 377 of the time and when the transformationbased tagger is wrong the maximum entropy tagger is right 417 of the timethe complementary rates are quite high which is encouraging since this sets the upper bound on how well we can do in combining the different classifiersif all taggers made the same errors or if the errors that loweraccuracy taggers made were merely a superset of higheraccuracy tagger errors then combination would be futilein addition a tagger is much more likely to have misclassified the tag for a word in instances where there is disagreement with at least one of the other classifiers than in the case where all classifiers agreein figure 3 we see for instance that while the overall error rate for the maximum entropy tagger is 317 in cases where there is disagreement between the four taggers the maximum entropy tagger error rate jumps to 271and discarding the unigram tagger which is significantly less accurate than the others when there is disagreement between the maximum entropy transformationbased and trigram taggers the maximum entropy tagger error rate jumps up to 437these cases account for 58 of the total errors the maximum entropy tagger makes next we check whether tagger complementarity is additivein figure 4 the first row shows the additive error rate an oracle could achieve on the test set if the oracle could pick between the different outputs of the taggersfor example when the oracle can examine the output of the maximum entropy transformationbased and trigram taggers it could achieve an error rate of 162the second row shows the additive error rate reduction the oracle could achieveif the oracle is allowed to choose between all four taggers a 555 error rate reduction is obtained over the maximum entropy tagger error rateif the unigram output is discarded the oracle improvement drops down to 488 over maximum entropy tagger error ratefrom these results we can conclude that there is at least hope that improvments can be gained by combining the output of different taggerswe can also conclude that the improvements we expect are somewhat additive meaning the more taggers we combine the better results we should expectthe fact that the errors the taggers make are strongly complementary is very encouragingif all taggers made the exact same errors there would obviously be no chance of improving accuracy through classifier combinationhowever note that the high complementary rate between tagger errors in itself does not necessarily imply that there is anything to be gained by classifier combinationwe ran experiments to determine whether the outputs of the different taggers could be effectively combinedwe first explored combination via simple majoritywins votingnext we attempted to automatically acquire contextual cues that learned both which tagger to believe in which contexts and what tags are indicated by different patterns of tagger outputsboth the word environments and the tagger outputs for the word being tagged and its neighbors are used as cues for predicting the proper tagthe simplest combination scheme is to have the classifiers votethe part of speech that appeared as the choice of the largest number of classifiers is picked as the answer with some method being specified for breaking tieswe tried simple voting using the maximum entropy transformationbased and trigram taggersin case of ties the maximum entropy tagger output is chosen since this tagger had the highest overall accuracy the results are shown in figure 5simple voting gives a net reduction in error of 69 over the best of the three taggersthis difference is significant at a 99 confidence levelnext we try to exploit the idiosyncracies of the different taggersalthough the maximum entropy transformationbased and trigram taggers use essentially the same types of contextual information for disambiguation this information is exploited differently in each caseour hope is that there is some regularity to these differences which would then allow us to learn what conditions suggest that we should trust one tagger output over anotherwe used a version of examplebased learning to determine whether these tagger differences could be exploited5 to determine 5 examplebased learning has also been applied succesfully in building a single part of speech tagger the tag of a word we use the previous word current word next word and the output of each tagger for the previous current and next wordsee figure 6for each such context in the training set we store the probabilities of what correct tags appeared in that contextwhen the tag distribution for a context has low entropy it is a very good predictor of the correct tag when the identical environment occurs in unseen datathe problem is that these environments are very specific and will have low overall recall in a novel corpusto account for this we must back off to more general contexts when we encounter an environment in the test set that did not occur in the training setthis is done by specifying an order in which fields should be ignored until a match is foundthe backoff ordering is learned automaticallywe ran two variants of this experimentin the first case given an instance in the test set we find the most specific matching example in the training set using the prespecified backoff ordering and see what the most probable tag was in the training set for that environmentthis is then chosen as the tag for the wordnote that this method is capable of learning to assign a tag that none of the taggers assignedfor instance it could be the case that when the unigram tagger thinks the tag should be x and the trigram and maximum entropy taggers think it should be y then the true tag is most frequently zin the second experiment we use contexts to specify which tagger to trust rather than which tag to outputagain the most specific context is found but here we check which tagger has the highest probability of being correct in this particular contextfor instance we may learn that the trigram tagger is most accurate at tagging the word up or that the unigrarn tagger does best at tagging the word race when the word that follows is andthe results are given in figure 7we see that while simple voting achieves an error reduction of 69 using contexts to choose a tag gives an error reduction of 98 and using contexts to choose a tagger gives an error reduction of 104in this paper we showed that the error distributions for three popular state of the art part of speech taggers are highly complementarynext we described experiments that demonstrated that we can exploit this complementarity to build a tagger that attains significantly higher accuracy than any of the individual taggersin the future we plan to expand our repertoire of base taggers to determine whether performance continues to improve as we add additional systemswe also plan to explore different methods for combining classifier outputswe suspect that the features we have chosen to use for combination are not the optimal set of featureswe need to carefully study the different algorithms to find possible cues that can indicate where a particular tagger performs wellwe hope that by following these general directions we can further exploit differences in classifiers to improve accuracy in lexical disambiguation	P98-1029	classifier combination for improved lexical disambiguationone of the most exciting recent directions in machine learning is the discovery that the combination of multiple classifiers often results in significantly better performance than what can be achieved with a single classifierin this paper we first show that the errors made from three different state of the art part of speech taggers are strongly complementarynext we show how this complementary behavior can be used to our advantageby using contextual cues to guide tagger combination we are able to derive a new tagger that achieves performance significantly greater than any of the individual taggerswe define the complementarity between two learners in order to quantify the percentage of time when one system is wrong while another system is correct therefore providing an upper bound on combination accuracy
errordriven pruning of treebank grammars for base noun phrase identification finding simple nonrecursive base noun phrases is an important subtask for many natural language processing applications while previous empirical methods for base np identification have been rather complex this paper instead proposes a very simple algorithm that is tailored to the relative simplicity of the task in particular we present a corpusbased approach for finding base nps by matching partofspeech tag sequences the training phase of the algorithm is based on two successful techniques first the base np grammar is read from a quottreebankquot corpus then the grammar is improved by selecting rules with high quotbenefitquot scores using this simple algorithm with a naive heuristic for matching rules we achieve surprising accuracy in an evaluation on the finding base noun phrases is a sensible first step for many natural language processing tasks accurate identification of base noun phrases is arguably the most critical component of any partial parser in addition information retrieval systems rely on base noun phrases as the main source of multiword indexing terms furthermore the psycholinguistic studies of gee and grosjean indicate that text chunks like base noun phrases play an important role in human language processingin this work we define base nps to be simple nonrecursive noun phrases noun phrases that do not contain other noun phrase descendantsthe bracketed portions of figure 1 for example show the base nps in one sentence from the penn treebank wall street journal corpus thus the string the sunny confines of resort towns like boca raton and hot springs is too complex to be a base np instead it contains four simpler noun phrases each of which is considered a base np the sunny confines resort towns boca raton and hot springsprevious empirical research has addressed the problem of base np identificationseveral algorithms identify quotterminological phrasesquot certain when it is time for their biannual powwow the nation manufacturing titans typically jet off to the sunny confines of resort towns like boca raton and hot springs base noun phrases with initial determiners and modifiers removed justeson katz look for repeated phrases bourigault uses a handcrafted noun phrase grammar in conjunction with heuristics for finding maximal length noun phrases voutilainen nptool uses a handcrafted lexicon and constraint grammar to find terminological noun phrases that include phrasefinal prepositional phraseschurch parts program on the other hand uses a probabilistic model automatically trained on the brown corpus to locate core noun phrases as well as to assign parts of speechmore recently ramshaw marcus apply transformationbased learning to the problemunfortunately it is difficult to directly compare approacheseach method uses a slightly different definition of base npeach is evaluated on a different corpusmost approaches have been evaluated by hand on a small test set rather than by automatic comparison to a large test corpus annotated by an impartial third partya notable exception is the ramshaw marcus work which evaluates their transformationbased learning approach on a base np corpus derived from the penn treebank wsj and achieves precision and recall levels of approximately 93this paper presents a new algorithm for identifying base nps in an arbitrary textlike some of the earlier work on base np identification ours is a trainable corpusbased algorithmin contrast to other corpusbased approaches however we hypothesized that the relatively simple nature of base nps would permit their accurate identification using correspondingly simple methodsassume for example that we use the annotated text of figure 1 as our training corpusto identify base nps in an unseen text we could simply search for all occurrences of the base nps seen during training it time their biannual powwow hot springs and mark them as base nps in the new texthowever this method would certainly suffer from data sparsenessinstead we use a similar approach but back off from lexical items to parts of speech we identify as a base np any string having the same partofspeech tag sequence as a base np from the training corpusthe training phase of the algorithm employs two previously successful techniques like charniak statistical parser our initial base np grammar is read from a quottreebankquot corpus then the grammar is improved by selecting rules with high quotbenefitquot scoresour benefit measure is identical to that used in transformationbased learning to select an ordered set of useful transformations using this simple algorithm with a naive heuristic for matching rules we achieve surprising accuracy in an evaluation on two base np corpora of varying complexity both derived from the penn treebank wsjthe first base np corpus is that used in the ramshaw marcus workthe second espouses a slightly simpler definition of base np that conforms to the base nps used in our empire sentence analyzerthese simpler phrases appear to be a good starting point for partial parsers that purposely delay all complex attachment decisions to later phases of processingoverall results for the approach are promisingfor the empire corpus our base np finder achieves 94 precision and recall for the ramshaw marcus corpus it obtains 91 precision and recall which is 2 less than the best published resultsramshaw marcus however provide the learning algorithm with wordlevel information in addition to the partofspeech information used in our base np finderby controlling for this disparity in available knowledge sources we find that our base np algorithm performs comparably achieving slightly worse precision and slightly better recall than the ramshaw marcus approachmoreover our approach offers many important advantages that make it appropriate for many nlp tasks note also that the treebank approach to base np identification obtains good results in spite of a very simple algorithm for quotparsingquot base npsthis is extremely encouraging and our evaluation suggests at least two areas for immediate improvementfirst by replacing the naive match heuristic with a probabilistic base np parser that incorporates lexical preferences we would expect a nontrivial increase in recall and precisionsecond many of the remaining base np errors tend to follow simple patterns these might be corrected using localized learnable repair rulesthe remainder of the paper describes the specifics of the approach and its evaluationthe next section presents the training and application phases of the treebank approach to base np identification in more detailsection 3 describes our general approach for pruning the base np grammar as well as two instantiations of that approachthe evaluation and a discussion of the results appear in section 4 along with techniques for reducing training time and an initial investigation into the use of local repair heuristicsfigure 2 depicts the treebank approach to base np identificationfor training the algorithm requires a corpus that has been annotated with base npsmore specifically we assume that the training corpus is a sequence of words wi w2 along with a set of base np annotations b where b indicates that the np brackets words i through j nr wi wjthe goal of the training phase is to create a base np grammar from this training corpus the resulting quotgrammarquot can then be used to identify base nps in a novel textnot this yearnational association of manufacturers settled on the hoosier capital of indianapolis for its next meetingand the city decided to treat its guests more like royalty or rock stars than factory ownersnot this yearnational association of manufacturers settled on the hoosier capital of indianapolis for i its next meetingand the city decided to treat its guests more like royalty or frock stars than factory ownerstraining corpus when it is time for their biannual powwow the nation manufacturing titans typically jet off to the sunny confines of resort towns like boca raton and hot springs3if there are multiple rules that match beginning at ti use the longest matching rule r add the new base noun phrase t to the set of base npscontinue matching at titriwith the rules stored in an appropriate data structure this greedy quotparsingquot of base nps is very fastin our implementation for example we store the rules in a decision tree which permits base np identification in time linear in the length of the tagged input text when using the longest match heuristicunfortunately there is an obvious problem with the algorithm described abovethere will be many unhelpful rules in the rule set extracted from the training corpusthese quotbadquot rules arise from four sources bracketing errors in the corpus tagging errors unusual or irregular linguistic constructs and inherent ambiguities in the base nps in spite of their simplicityfor example the rule which was extracted from manufacturingvbg titansnns in the example text is ambiguous and will cause erroneous bracketing in sentences such as the execs squeezed in a few meetings before boardingvbg busesnns againin order to have a viable mechanism for identifying base nps using this algorithm the grammar must be improved by removing problematic rulesthe next section presents two such methods for automatically pruning the base np grammaras described above our goal is to use the base np corpus to extract and select a set of noun phrase rules that can be used to accurately identify base nps in novel textour general pruning procedure is shown in figure 3first we divide the base np corpus into two parts a training corpus and a pruning corpusthe initial base np grammar is extracted from the training corpus as described in section 2next the pruning corpus is used to evaluate the set of rules and produce a ranking of the rules in terms of their utility in identifying base npsmore specifically we use the rule set and the longest match heuristic to find all base nps in the pruning corpusperformance of the rule set is measured in terms of labeled precision we then assign to each rule a score that denotes the quotnet benefitquot achieved by using the rule during np parsing of the improvement corpusthe benefit of rule r is given by br cr er where cr is the number of nps correctly identified by r and er is the number of precision errors for which r is responsible1 a rule is considered responsible for an error if it was the first rule to bracket part of a reference np ie an np in the base np training corpusthus rules that form erroneous bracketings are not penalized if another rule previously bracketed part of the same reference npfor example suppose the fragment containing base nps boca raton hot springs and palm beach is bracketed as shown below resort towns like np bocannp ratonnnp hotnnp np springsnnp and np palmnnp beachnnp1 rule brackets npi brackets np2 and brackets np3rule incorrectly identifies boca raton hot as a noun phrase so its score is 1rule incorrectly identifies springs but it is not held responsible for the error because of the previous error by on the same original np hot springs so its score is 0finally rule receives a score of 1 for correctly identifying palm beach as a base npthe benefit scores from evaluation on the pruning corpus are used to rank the rules in the grammarwith such a ranking we can improve the rule set by discarding the worst rulesthus far we have investigated two iterative approaches for discarding rules a thresholding approach and an incremental approachwe describe each in turn in the subsections belowthis same benefit measure is also used in the rm study but it is used to rank transformations rather than to rank np rulesgiven a ranking on the rule set the threshold algorithm simply discards rules whose score is less than a predefined threshold r for all of our experiments we set are 1 to select rules that propose more correct bracketings than incorrectthe process of evaluating ranking and discarding rules is repeated until no rules have a score less than r for our evaluation on the wsj corpus this typically requires only four to five iterationsthresholding provides a very coarse mechanism for pruning the np grammarin particular because of interactions between the rules during bracketing thresholding discards rules whose score might increase in the absence of other rules that are also being discardedconsider for example the boca raton fragments given earlierin the absence of the rule would have received a score of three for correctly identifying all three npsas a result we explored a more finegrained method of discarding rules each iteration of incremental pruning discards the n worst rules rather than all rules whose rank is less than some thresholdin all of our experiments we set n 10as with thresholding the process of evaluating ranking and discarding rules is repeated this time until precision of the current rule set on the pruning corpus begins to dropthe rule set that maximized precision becomes the final rule setin the experiments below we compare the thresholding and incremental methods for pruning the np grammar to a rule set that was pruned by handwhen the training corpus is large exhaustive review of the extracted rules is not practicalthis is the case for our initial rule set culled from the wsj corpus which contains approximately 4500 base np rulesrather than identifying and discarding individual problematic rules our reviewer identified problematic classes of rules that could be removed from the grammar automaticallyin particular the goal of the human reviewer was to discard rules that introduced ambiguity or corresponded to overly complex base npswithin our partial parsing framework these nps are better identified by more informed components of the nlp systemour reviewer identified the following classes of rules as possibly troublesome rules that contain a preposition period or colon rules that contain wh tags rules that beginend with a verb or adverb rules that contain pronouns with any other tags rules that contain misplaced commas or quotes rules that end with adjectivesrules covered under any of these classes were omitted from the humanpruned rule sets used in the experiments of section 4to evaluate the treebank approach to base np identification we created two base np corporaeach is derived from the penn treebank wsjthe first corpus attempts to duplicate the base nps used the ramshaw marcus studythe second corpus contains slightly less complicated base nps base nps that are better suited for use with our sentence analyzer empireby evaluating on both corpora we can measure the effect of noun phrase complexity on the treebank approach to base np identificationin particular we hypothesize that the treebank approach will be most appropriate when the base nps are sufficiently simplefor all experiments we derived the training pruning and testing sets from the 25 sections of wall street journal distributed with the penn treebank iiall experiments employ 5fold cross validationmore specifically in each of five runs a different fold is used for testing the final pruned rule set three of the remaining folds comprise the training corpus and the final partition is the pruning corpus all results are averages across the five foldsperformance is measured in terms of precision and recallprecision was described earlier it is a standard measure of accuracyrecall on the other hand is an attempt to measure coverage throughout the table we see the effects of base np complexity the base nps of the rm corpus are substantially more difficult for our approach to identify than the simpler nps of the empire corpusfor the rm corpus we lag the best published results by approximately 3this straightforward comparison however is not entirely appropriateramshaw marcus allow their learning algorithm to access wordlevel information in addition to partofspeech tagsthe treebank approach on the other hand makes use only of partofspeech tagstable 2 compares ramshaw marcus results with and without lexical knowledgethe first column reports their performance when using lexical templates the second when lexical templates are not used the third again shows the treebank approach using incremental pruningthe treebank approach and the rm approach without lecial templates are shown to perform comparably lexicalization of our base np finder will be addressed in section 41finally note the relatively small difference between the threshold and incremental pruning methods in table 1for some applications this minor drop in performance may be worth the decrease in training timeanother effective technique to speed up training is motivated by charniak observation that the benefit of using rules that only occurred once in training is marginalby discarding these rules before pruning we reduce the size of the initial grammar and the time for incremental pruning by 60 with a performance drop of only 03p01rit is informative to consider the kinds of errors made by the treebank approach to bracketingin particular the errors may indicate options for incorporating lexical information into the base np findergiven the increases in performance achieved by ramshaw marcus by including wordlevel cues we would hope to see similar improvements by exploiting lexical information in the treebank approachfor each corpus we examined the first 100 or so errors and found that certain linguistic constructs consistently because trouble of nps in the annotated text table 1 summarizes the performance of the treebank approach to base np identification on the rm and empire corpora using the initial and pruned rule setsthe first column of results shows the performance of the initial unpruned base np grammarthe next two columns show the performance of the automatically pruned rule setsthe final column indicates the performance of rule sets that had been pruned using the handcrafted pruning heuristicsas expected the initial rule set performs quite poorlyboth automated approaches provide significant increases in both recall and precisionin addition they outperform the rule set pruned using handcrafted pruning heuristicsmany errors appear to stem from four underlying causesfirst close to 20 can be attributed to errors in the treebank and in the base np corpus bringing the effective performance of the algorithm to 942p959r and 915p927r for the empire and rm corpora respectivelyfor example neither corpus includes whphrases as base npswhen the bracketer correctly recognizes these nps they are counted as errorspartofspeech tagging errors are a second becausethird many nps are missed by the bracketer because it lacks the appropriate rulefor example household products business is bracketed as householdnn productsnns businessnmfourth idiomatic and specialized expressions especially time date money and numeric phrases also account for a substantial portion of the errorsthese last two categories of errors can often be detected because they produce either recognizable patterns or unlikely linguistic constructsconsecutive nps for example usually denote bracketing errors as in householdnn productsnns businessnmmerging consecutive nps in the correct contexts would fix many such errorsidiomatic and specialized expressions might be corrected by similarly local repair heuristicstypical examples might include changing effectivejj mondaynnp to effective monday changing thedt balancenn duejj to the balance due and changing werevbp intrb thedt onlyrb losersnns to were nt the only losersgiven these observations we implemented three local repair heuristicsthe first merges consecutive nps unless either might be a time expressionthe second identifies two simple date expressionsthe third looks for quantifiers preceding of npthe first heuristic for example merges household products business to form household products business but leaves increased 15 last friday untouchedthe second heuristic merges june 5 1995 into june 5 1995 and june 1995 into june 1995the third finds examples like some of the companies and produces some of the companiesthese heuristics represent an initial exploration into the effectiveness of employing lexical information in a postprocessing phase rather than during grammar induction and bracketingwhile we are investigating the latter in current work local repair heuristics have the advantage of keeping the training and bracketing algorithms both simple and fastthe effect of these heuristics on recall and precision is shown in table 3we see consistent improvements for both corpora and both pruning methods achieving approximately 94pr for the empire corpus and approximately 91pr for the rm corpusnote that these are the final results reported in the introduction and conclusionalthough these experiments represent only an initial investigation into the usefulness of local repair heuristics we are very encouraged by the resultsthe heuristics uniformly boost precision without harming recall they help the rm corpus even though they were designed in response to errors in the empire corpusin addition these three heuristics alone recover 12 to 13 of the improvements we can expect to obtain from lexicalization based on the rm resultsthis paper presented a new method for identifying base npsour treebank approach uses the simple technique of matching partofspeech tag sequences with the intention of capturing the simplicity of the corresponding syntactic structureit employs two existing corpusbased techniques the initial noun phrase grammar is extracted directly from an annotated corpus and a benefit score calculated from errors on an improvement corpus selects the best subset of rules via a coarse or finegrained pruning algorithmthe overall results are surprisingly good especially considering the simplicity of the methodit achieves 94 precision and recall on simple base npsit achieves 91 precision and recall on the more complex nps of the rainshaw marcus corpuswe believe however that the base np finder can be improved furtherfirst the longestmatch heuristic of the noun phrase bracketer could be replaced by more sophisticated parsing methods that account for lexical preferencesrule application for example could be disambiguated statistically using distributions induced during trainingwe are currently investigating such extensionsone approach closely related to ours weighted finitestate transducers might provide a principled way to do thiswe could then consider applying our errordriven pruning strategy to rules encoded as transducerssecond we have only recently begun to explore the use of local repair heuristicswhile initial results are promising the full impact of such heuristics on overall performance can be determined only if they are systematically learned and tested using available training datafuture work will concentrate on the corpusbased acquisition of local repair heuristicsin conclusion the treebank approach to base nps provides an accurate and fast bracketing method running in time linear in the length of the tagged textthe approach is simple to understand implement and trainthe learned grammar is easily modified for use with new corpora as rules can be added or deleted with minimal interaction problemsfinally the approach provides a general framework for developing other treebank grammars in addition to these for base npsacknowledgmentsthis work was supported in part by nsf grants iri9624639 and ger9454149we thank mitre for providing their partofspeech tagger	P98-1034	errordriven pruning of treebank grammars for base noun phrase identificationfinding simple nonrecursive base noun phrases is an important subtask for many natural language processing applicationswhile previous empirical methods for base np identification have been rather complex this paper instead proposes a very simple algorithm that is tailored to the relative simplicity of the taskin particular we present a corpusbased approach for finding base nps by matching partofspeech tag sequencesthe training phase of the algorithm is based on two successful techniques first the base np grammar is read from a treebank corpus then the grammar is improved by selecting rules with high benefit scoresusing this simple algorithm with a naive heuristic for matching rules we achieve surprising accuracy in an evaluation on the penn treebank wall street journalwe store pos tag sequences that make up complete chunks and use these sequences as rules for classifying unseen data
exploiting syntactic structure for language modeling the paper presents a language model that develops syntactic structure and uses it to extract meaningful information from the word history thus enabling the use of long distance dependencies the model assigns probability to every joint sequence of wordsbinaryparsestructure with headword annotation and operates in a lefttoright manner therefore usable for automatic speech recognition the model its probabilistic parameterization and a set of experiments meant to evaluate its predictive power are presented an improvement over standard trigram modeling is achieved the main goal of the present work is to develop a language model that uses syntactic structure to model longdistance dependenciesduring the summer96 dod workshop a similar attempt was made by the dependency modeling groupthe model we present is closely related to the one investigated in however different in a few important aspects our model operates in a lefttoright manner allowing the decoding of word lattices as opposed to the one referred to previously where only whole sentences could be processed thus reducing its applicability to nbest list rescoring the syntactic structure is developed as a model component our model is a factored version of the one in thus enabling the calculation of the joint probability of words and parse structure this was not possible in the previous case due to the huge computational complexity of the modelour model develops syntactic structure incrementally while traversing the sentence from left to rightthis is the main difference between our approach and other approaches to statistical natural language parsingour parsing strategy is similar to the incremental syntax ones proposed relatively recently in the linguistic community the probabilistic model its parameterization and a few experiments that are meant to evaluate its potential for speech recognition are presentedconsider predicting the word after in the sentence the contract ended with a loss of 7 cents after trading as low as 89 centsa 3gram approach would predict after from whereas it is intuitively clear that the strongest predictor would be ended which is outside the reach of even 7gramsour assumption is that what enables humans to make a good prediction of after is the syntactic structure in the pastthe linguistically correct partial parse of the word history when predicting after is shown in figure 1the word ended is called the headword of the constituent and ended is an exposed headword when predicting after topmost headword in the largest constituent that contains itthe syntactic structure in the past filters out irrelevant words and points to the important ones thus enabling the use of long distance information when predicting the next wordour model will attempt to build the syntactic structure incrementally while traversing the sentence lefttorightthe model will assign a probability p to every sentence w with every possible postag assignment binary branching parse nonterminal label and headword annotation for every constituent of t let w be a sentence of length n words to which we have prepended and appended so that wo and w1 let wk be the word kprefix wo wk of the sentence and wk tk the wordparse kprefixto stress this point a wordparse kprefix contains for a given parse only those binary subtrees whose span is completely included in the word kprefix excluding wo single words along with their postag can be regarded as rootonly treesfigure 2 shows a wordparse kprefix h_o h_m are the exposed heads each head being a pair or in the case of a rootonly treea complete parse figure 3 is any binary parse of the sequence with the restriction that is the only allowed headnote that need not be a constituent but for the parses where it is there is no restriction on which of its words is the headword or what is the nonterminal label that accompanies the headwordthe model will operate by means of three modules or until it passes control to the predictor by taking a null transitionntlabel is the nonterminal label assigned to the newly built constituent and left right specifies where the new headword is inherited fromthe operations performed by the parser are illustrated in figures 46 and they ensure that all possible binary branching parses with all possible headword and nonterminal label assignments for the w1 wk word sequence can be generatedthe following algorithm formalizes the above description of the sequential generation of a sentence with a complete parsethe unary transition is allowed only when the most recent exposed head is a leaf of the tree a regular word along with its postag hence it can be taken at most once at a given position in the input word stringthe second subtree in figure 2 provides an example of a unary transition followed by a null transitionit is easy to see that any given word sequence with a possible parse and headword annotation is generated by a unique sequence of model actionsthis will prove very useful in initializing our model parameters from a treebank see section 35the probability p of a word sequence w and a complete parse t can be broken into where as can be seen is one of the nk wordparse kprefixes wktk at position k in the sentence i i nkto ensure a proper probabilistic model we have to make sure that and are well defined conditional probabilities and that the model halts with probability oneconsequently certain parser and wordpredictor probabilities must be given specific values pwktk 1 if h_o and h_1 word 0 ensure that the parse generated by our model is consistent with the definition of a complete parse the wordpredictor model predicts the next word based on the preceding 2 exposed heads thus making the following equivalence classification after experimenting with several equivalence classifications of the wordparse prefix for the tagger model the conditioning part of model was reduced to using the word to be tagged and the tags of the two most recent exposed heads model assigns probability to different parses of the word kprefix by chaining the elementary operations described abovethe workings of the parser module are similar to those of spatter the equivalence classification of the wktk wordparse we used for the parser model was the same as the one used in it is worth noting that if the binary branching structure developed by the parser were always rightbranching and we mapped the postag and nonterminal label vocabularies to a single type then our model would be equivalent to a trigram language modelall model components wordpredictor tagger parser are conditional probabilistic models of the type p where yx1x2xn belong to a mixed bag of words postags nonterminal labels and parser operations for simplicity the modeling method we chose was deleted interpolation among relative frequency estimates of different orders f using a recursive mixing scheme as can be seen the context mixing scheme discards items in the context in righttoleft orderthe a coefficients are tied based on the range of the count cthe approach is a standard one which does not require an extensive description given the literature available on it since the number of parses for a given word prefix wk grows exponentially with kitk1 0 the state space of our model is huge even for relatively short sentences so we had to use a search strategy that prunes itour choice was a synchronous multistack search algorithm which is very similar to a beam searcheach stack contains hypotheses partial parses that have been constructed by the same number of predictor and the same number of parser operationsthe hypotheses in each stack are ranked according to the ln score highest on topthe width of the search is controlled by two parameters above pruning strategy proved to be insufficient so we chose to also discard all hypotheses whose score is more than the logprobability threshold below the score of the topmost hypothesisthis additional pruning step is performed after all hypotheses in stage k have been extended with the null parser transition and thus prepared for scanning a new wordthe conditional perplexity calculated by assigning to a whole sentence the probability where t argmaxtp is not valid because it is not causal when predicting wki we use t which was determined by looking at the entire sentenceto be able to compare the perplexity of our model with that resulting from the standard trigram approach we need to factor in the entropy of guessing the correct parse i before predicting wki based solely on the word prefix wkthe probability assignment for the word at position k 1 in the input sentence is made using which ensures a proper probability over strings w where sk is the set of all parses present in our stacks at the current stage k another possibility for evaluating the word level perplexity of our model is to approximate the probability of a whole sentence where t is one of the quotnbestquot in the sense defined by our search parses for w this is a deficient probability assignment however useful for justifying the model parameter reestimationthe two estimates and are both consistent in the sense that if the sums are carried over all possible parses we get the correct value for the word level perplexity of our modelthe major problem we face when trying to reestimate the model parameters is the huge state space of the model and the fact that dynamic programming techniques similar to those used in hmm parameter reestimation cannot be used with our modelour solution is inspired by an hmm reestimation technique that works on pruned nbest trelliseslet k 1 n be the set of hypotheses that survived our pruning strategy until the end of the parsing process for sentence w each of them was produced by a sequence of model actions chained together as described in section 2 let us call the sequence of model actions that produced a given the derivationlet an elementary event in the derivation the probability associated with each model action is determined as described in section 31 based on counts c x one set for each model componentassuming that the deleted interpolation coefficients and the count ranges used for tying them stay fixed these counts are the only parameters to be reestimated in an eventual reestimation procedure indeed once a set of counts c x is specified for a given model m we can easily calculate in ky xn for all context orders n 0 maximumorder this is all we need for calculating the probability of an elementary event and then the probability of an entire derivationone training iteration of the reestimation procedure we propose is described by the following algorithm nbest parse development data countsei prepare countse for each model component c gather_counts development model_c in the parsing stage we retain for each quotnbestquot hypothesis k 1 n only the quantity 0 p nn1 p kw7 t and its derivationwe then scan all the derivations in the quotdevelopment setquot and for each occurrence of the elementary event x in derivation we accumulate the value 0 in the c x counter to be used in the next iterationthe intuition behind this procedure is that 0 is an approximation to the pw probability which places all its mass on the parses that survived the parsing process the above procedure simply accumulates the expected values of the counts om x under the 0 conditional distributionas explained previously the om x counts are the parameters defining our model making our procedure similar to a rigorous them approach a particular and very interesting case is that of events which had count zero but get a nonzero count in the next iteration caused by the quotnbestquot nature of the reestimation processconsider a given sentence in our quotdevelopmentquot setthe quotnbestquot derivations for this sentence are trajectories through the state space of our modelthey will change from one iteration to the other due to the smoothing involved in the probability estimation and the change of the parameters event counts defining our model thus allowing new events to appear and discarding others through purging low probability events from the stacksthe higher the number of trajectories per sentence the more dynamic this change is expected to bethe results we obtained are presented in the experiments sectionall the perplexity evaluations were done using the lefttoright formula for which the perplexity on the quotdevelopment setquot is not guaranteed to decrease from one iteration to anotherhowever we believe that our reestimation method should not increase the approximation to perplexity based on again on the quotdevelopment setquot we rely on the consistency property outlined at the end of section 33 to correlate the desired decrease in l2rpeople with that in sumpeopleno claim can be made about the change in either l2rpeople or sumpeople on test dataeach model component wordpredictor tagger parser is trained initially from a set of parsed sentences after each parse tree undergoes these are the initial parameters used with the reestimation procedure described in the previous sectionin order to get initial statistics for our model components we needed to binarize the upenn treebank parse trees and percolate headwordsthe procedure we used was to first percolate headwords using a contextfree rulebased approach and then binarize the parses by using a rulebased approach againthe headword of a phrase is the word that best represents the phrase all the other words in the phrase being modifiers of the headwordstatistically speaking we were satisfied with the output of an enhanced version of the procedure described in also known under the name quotmagerman sz black headword percolation rulesquotonce the position of the headword within a constituent equivalent with a cf production of the type z 4 y1 y2 where z yr are nonterminal labels or postags is identified to be k we binarize the constituent as follows depending on the z identity a fixed rule is used to decide which of the two binarization schemes in figure 8 to applythe intermediate nodes created by the above binarization schemes receive the nonterminal label zdue to the low speed of the parser 200 wdsmin for stack depth 10 and logprobability threshold 691 nats we could carry out the reestimation technique described in section 34 on only 1 mwds of training datafor convenience we chose to work on the upenn treebank corpusthe vocabulary sizes were the training data was split into quotdevelopmentquot set 929564wds and quotcheck setquot 73760wds the test set size was 82430wds the quotcheckquot set has been used for estimating the interpolation weights and tuning the search parameters the quotdevelopmentquot set has been used for gatheringestimating counts the test set has been used strictly for evaluating model performancetable 1 shows the results of the reestimation technique presented in section 34we achieved a reduction in testdata perplexity bringing an improvement over a deleted interpolation trigram model whose perplexity was 16714 on the same trainingtest data the reduction is statistically significant according to a sign testsimple linear interpolation between our model and the trigram model yielded a further improvement in people as shown in table 2the interpolation weight was estimated on check data to be a 036an overall relative reduction of 11 over the trigram model has been achievedthe large difference between the perplexity of our model calculated on the quotdevelopmentquot set used for model parameter estimation and quottestquot set unseen data shows that the initial point we choose for the parameter values has already captured a lot of information from the training datathe same problem is encountered in standard ngram language modeling however our approach has more flexibility in dealing with it due to the possibility of reestimating the model parameterswe believe that the above experiments show the potential of our approach for improved language modelsour future plans includethis research has been funded by the nsf iri19618874 grant the authors would like to thank sanjeev khudanpur for his insightful suggestionsalso to harry printz eric ristad andreas stolcke dekai wu and all the other members of the dependency modeling group at the summer96 dod workshop for useful comments on the model programming support and an extremely creative environmentalso thanks to eric brill sanjeev khudanpur david yarowsky radu florian lidia mangu and jun wu for useful input during the meetings of the people working on our stimulate grant	P98-1035	exploiting syntactic structure for language modelingthe paper presents a language model that develops syntactic structure and uses it to extract meaningful information from the word history thus enabling the use of long distance dependenciesthe model assigns probability to every joint sequence of wordsbinaryparsestructure with headword annotation and operates in a lefttoright manner therefore usable for automatic speech recognitionthe model its probabilistic parameterization and a set of experiments meant to evaluate its predictive power are presented an improvement over standard trigram modeling is achievedwe choose the lexical heads of the two previous constituents as determined by a shiftreduce parser and find that this works better than a trigram modelwe condition on linguistically relevant words by assigning partial phrase structure to the history and percolating headwords
investigating regular sense extensions based on intersective levin classes in this paper we specifically address questions of polysemy with respect to verbs and how regular extensions of meaning can be achieved through the adjunction of particular syntactic phrases we see verb classes as the key to making generalizations about regular extensions of meaning current approaches to english classification levin classes and wordnet have limitations in their applicability that impede their utility as general classification schemes we present a refinement of levin classes intersective sets which are a more finegrained classification and have more coherent sets of syntactic frames and associated semantic components we have preliminary indications that the membership of our intersective sets will be more compatible with wordnet than the original levin classes we also have begun to examine related classes in portuguese and find that these verbs demonstrate similarly coherent syntactic and semantic properties the difficulty of achieving adequate handcrafted semantic representations has limited the field of natural language processing to applications that can be contained within welldefined subdomainsthe only escape from this limitation will be through the use of automated or semiautomated methods of lexical acquisitionhowever the field has yet to develop a clear consensus on guidelines for a computational lexicon that could provide a springboard for such methods although attempts are being made the authors would like to acknowledge the support of darpa grant n6600194c6043 aro grant daah0494g0426 and capes grant 0914952one of the most controversial areas has to do with polysemywhat constitutes a clear separation into senses for any one verb and how can these senses be computationally characterized and distinguishedthe answer to this question is the key to breaking the bottleneck of semantic representation that is currently the single greatest limitation on the general application of natural language processing techniquesin this paper we specifically address questions of polysemy with respect to verbs and how regular extensions of meaning can be achieved through the adjunction of particular syntactic phraseswe base these regular extensions on a finegrained variation on levin classes intersective levin classes as a source of semantic components associated with specific adjunctswe also examine similar classes in portuguese and the predictive powers of alternations in this language with respect to the same semantic componentsthe difficulty of determining a suitable lexical representation becomes multiplied when more than one language is involved and attempts are made to map between thempreliminary investigations have indicated that a straightforward translation of levin classes into other languages is not feasible however we have found interesting parallels in how portuguese and english treat regular sense extensionstwo current approaches to english verb classifications are wordnet and levin classes wordnet is an online lexical database of english that currently contains approximately 120000 sets of noun verb adjective and adverb synonyms each representing a lexicalized concepta synset contains besides all the word forms that can refer to a given concept a definitional gloss and in most cases an example sentencewords and synsets are interrelated by means of lexical and semanticconceptual links respectivelyantonymy or semantic opposition links individual words while the supersubordinate relation links entire synsetswordnet was designed principally as a semantic network and contains little syntactic informationlevin verb classes are based on the ability of a verb to occur or not occur in pairs of syntactic frames that are in some sense meaning preserving the distribution of syntactic frames in which a verb can appear determines its class membershipthe fundamental assumption is that the syntactic frames are a direct reflection of the underlying semanticslevin classes are supposed to provide specific sets of syntactic frames that are associated with the individual classesthe sets of syntactic frames associated with a particular levin class are not intended to be arbitrary and they are supposed to reflect underlying semantic components that constrain allowable argumentsfor example break verbs and cut verbs are similar in that they can all participate in the transitive and in the middle construction john broke the window glass breaks easily john cut the bread this loaf cuts easilyhowever only break verbs can also occur in the simple intransitive the window broke the bread cutin addition cut verbs can occur in the conative john valiantly cuthacked at the frozen loaf but his knife was too dull to make a dent in it whereas break verbs cannot john broke at the windowthe explanation given is that cut describes a series of actions directed at achieving the goal of separating some object into piecesit is possible for these actions to be performed without the end result being achieved but where the cutting manner can still be recognized ie john cut at the loafwhere break is concerned the only thing specified is the resulting change of state where the object becomes separated into piecesif the result is not achieved there are no attempted breaking actions that can still be recognizedit is not clear how much wordnet synsets should be expected to overlap with levin classes and preliminary indications are that there is a wide discrepancy however it would be useful for the wordnet senses to have access to the detailed syntactic information that the levin classes contain and it would be equally useful to have more guidance as to when membership in a levin class does in fact indicate shared semantic componentsof course some levin classes such as braid are clearly not intended to be synonymous which at least partly explains the lack of overlap between levin and wordnetthe association of sets of syntactic frames with individual verbs in each class is not as straightforward as one might supposefor instance carry verbs are described as not taking the conative the mother carried at the baby and yet many of the verbs in the carry class are also listed in the pushpull class which does take the conativethis listing of a verb in more than one class is left open to interpretation in levindoes it indicate that more than one sense of the verb is involved or is one sense primary and the alternations for that class should take precedence over the alternations for the other classes in which the verb is listedthe grounds for deciding that a verb belongs in a particular class because of the alternations that it does not take are elusive at bestwe augmented the existing database of levin semantic classes with a set of intersective classes which were created by grouping together subsets of existing classes with overlapping membersall subsets were included which shared a minimum of three membersif only one or two verbs were shared between two classes we assumed this might be due to homophony an idiosyncrasy involving individual verbs rather than a systematic relationship involving coherent sets of verbsthis filter allowed us to reject the potential intersective class that would have resulted from combining the remove verbs with the scribble verbs for examplethe sole member of this intersection is the verb mantic classes such that ci n n cal c where e is a relevance cutoffwe then reclassified the verbs in the database as followsa verb was assigned membership in an intersective class if it was listed in each of the existing classes that were combined to form the new intersective classsimultaneously the verb was removed from the membership lists of those existing classessome of the large levin classes comprise verbs that exhibit a wide range of possible semantic components and could be divided into smaller subclassesthe split verbs do not obviously form a homogeneous semantic classinstead in their use as split verbs each verb manifests an extended sense that can be paraphrased as quotseparate by vingquot where quotvquot is the basic meaning of that verb many of the verbs that do not have an inherent semantic component of quotseparatingquot belong to this class because of the component of force in their meaningthey are interpretable as verbs of splitting or separating only in particular syntactic frames the branch but not i pulled the twig and the branchthe adjunction of the apart adverb adds a change of state semantic component with respect to the object which is not present otherwisethese fringe split verbs appear in several other intersective classes that highlight the force aspect of their meaningfigure 2 depicts the intersection of split carry and pushpullfigure 2 intersective class formed from levin carry pushpull and split verbs verbs in are not listed by levin in all the intersecting classes but participate in all the alternations the intersection between the pushpull verbs of exerting force the carry verbs and the split verbs illustrates how the force semantic component of a verb can also be used to extend its meaning so that one can infer a causation of accompanied motiondepending on the particular syntactic frame in which they appear members of this intersective class can be used to exemplify any one of the the component levin classes for each levin verb class is not always complete so to check if a particular verb belongs to a class it is better to check that the verb exhibits all the alternations that define the classsince intersective classes were built using membership lists rather than the set of defining alternations they were similarly incompletethis is an obvious shortcoming of the current implementation of intersective classes and might affect the choice of 3 as a relevance cutoff in later implementations although the levin classes that make up an intersective class may have conflicting alternations this does not invalidate the semantic regularity of the intersective classas a verb of exerting force push can appear in the conative alternation which emphasizes its force semantic component and ability to express an quotattemptedquot action where any result that might be associated with the verb is not necessarily achieved as a carry verb push cannot take the conative alternation which would conflict with the core meaning of the carry verb class the critical point is that while the verb meaning can be extended to either quotattemptedquot action or directed motion these two extensions cannot cooccur they are mutually exclusivehowever the simultaneous potential of mutually exclusive extensions is not a problemit is exactly those verbs that are triplelisted in the splitpushcarry intersective class that can take the conativethe carry verbs that are not in the intersective class are more quotpurequot examples of the carry class and always imply the achievement of causation of motionthus they cannot take the conative alternationeven though the levin verb classes are defined by their syntactic behavior many reflect semantic distinctions made by wordnet a classification hierarchy defined in terms of purely semantic word relations when examining in detail the intersective classes just described which emphasize not only the individual classes but also their relation to other classes we see a rich semantic lattice much like wordnetthis is exemplified by the levin cut verbs and the intersective class formed by the cut verbs and split verbsthe original intersective class exhibits alternations of both parent classes and has been augmented with chip clip slash snip since these cut verbs also display the syntactic properties of split verbswordnet distinguishes two subclasses of cut differentiated by the type of result this distinction appears in the secondorder levin classes as membership vs nonmembership in the intersective class with splitlevin verb classes are based on an underlying lattice of partial semantic descriptions which are manifested indirectly in diathesis alternationswhereas high level semantic relations are represented directly in wordnet they can sometimes be inferred from the intersection between levin verb classes as with the cutsplit classhowever other intersective classes such as the splitpushcarry class are no more consistent with wordnet than the original levin classesthe most specific hypernym common to all the verbs in this intersective class is move displace which is also a hypernym for other carry verbs not in the intersectionin addition only one verb has a wordnet sense corresponding to the change of state separation semantic component associated with the split classthe fact that the split sense for these verbs does not appear explicitly in wordnet is not surprising since it is only an extended sense of the verbs and separation is inferred only when the verb occurs with an appropriate adjunct such as aparthowever apart can also be used with other classes of verbs including many verbs of motionto explicitly list separation as a possible sense for all these verbs would be extravagant when this sense can be generated from the combination of the adjunct with the force or motion semantic component of the verbwordnet does not currently provide a consistent treatment of regular sense extension it would be straightforward to augment it with pointers indicating which senses are basic to a class of verbs and which can be generated automatically and include corresponding syntactic informationfigure 3 shows intersective classes involving two classes of verbs of manner of motion and a class of verbs of existence roll and run verbs have semantic components describing a manner of motion that typically though not necessarily involves change of locationin the absence of a goal or path adjunct they do not specify any direction of motion and in some cases require the adjunct to explicitly specify any displacement at allthe two classes differ in that roll verbs relate to manners of motion characteristic of inanimate entities while run verbs describe manners in which animate entities can movesome manner of motion verbs allow a transitive alternation in addition to the basic intransitivewhen a roll verb occurs in the transitive the subject physically causes the object to move whereas the subject of a transitive run verb merely induces the object to move some verbs can be used to describe motion of both animate and inanimate objects and thus appear in both roll and run verb classesthe slide class partitions this rollrun intersection into verbs that can take the transitive alternation and verbs that cannot verbs in the sliderollrun intersection are also allowed to appear in the dative alternation in which the sense of change of location is extended to change of possessionwhen used intransitively with a path prepositional phrase some of the manner of motion verbs can take on a sense of pseudomotional existence in which the subject does not actually move but has a shape that could describe a path for the verb these verbs are listed in the intersective classes with meander verbs of existencethe portuguese verbs we examined behaved much more similarly to their english counterparts than we expectedmany of the verbs participate in alternations that are direct translations of the english alternationshowever there are some interesting differences in which sense extensions are allowedwe have made a preliminary study of the portuguese translation of the carry verb classas in english these verbs seem to take different alternations and the ability of each to participate in an alternation is related to its semantic contenttable 1 shows how these portuguese verbs naturally cluster into two different subclasses based on their ability to take the conative and apart alternations as well as path prepositionsthese subclasses correspond very well to the english subclasses created by the intersective classthe conative alternation in portuguese is mainly contra and the apart alternation is mainly separando for example eu puxei o ramo e o galho separandoos and ele empurrou contra a parede we also investigated the portuguese translation of some intersective classes of motion verbswe selected the sliderollrun meanderroll and rollrun intersective classesmost verbs have more than one translation into portuguese so we chose the translation that best described the meaning or that had the same type of arguments as described in levin verb classesthe elements of the sliderollrun class are rebater flutuar rolar and deslizar the resultative in portuguese cannot be expressed in the same way as in englishit takes a gerund plus a reflexive as in a porta deslizou abrindose transitivity is also not always preserved in the translationsfor example flutuar does not take a direct object so some of the alternations that are related to its transitive meaning are not presentfor these verbs we have the induced action alternation by using the light verb fazer before the verb as in maria fez o barco flutuar as can be seen in table 2 the alternations for the portuguese translations of the verbs in this intersective class indicate that they share similar properties with the english verbs including the causativeinchoativethe exception to this as just noted is flutuar the result of this is that flutuar should move out of the slide class which puts it with derivar and planar in the closely related rollrun classas in english derivar and planar are not externally controllable actions and thus do not take the causativeinchoative alternation common to other verbs in the roll classplanar does not take a direct object in portuguese and it shows the induced action alternation the same way as flutuar derivar is usually said as quotestar a derivaquot showing its noncontrollable action more explicitlywe have presented a refinement of levin classes intersective classes and discussed the potential for mapping them to wordnet senseswhereas each wordnet synset is hierarchicalized according to only one aspect levin recognizes that verbs in a class may share many different semantic features without designating one as primaryintersective levin sets partition these classes according to more coherent subsets of features in effect highlighting a lattice of semantic features that determine the sense of a verbgiven the incompleteness of the list of members of levin classes each verb must be examined to see whether it exhibits all the alternations of a classthis might be approximated by automatically extracting the syntactic frames in which the verb occurs in corpus data rather than manual analysis of each verb as was done in this studywe have also examined a mapping between the english verbs that we have discussed and their portuguese translations which have several of the same properties as the corresponding verbs in englishmost of these verbs take the same alternations as in english and by virtue of these alternations achieve the same regular sense extensionsthere are still many questions that require further investigationfirst since our experiment was based on a translation from english to portuguese we can expect that other verbs in portuguese would share the same alternations so the classes in portuguese should by no means be considered completewe will be using resources such as dictionaries and online corpora to investigate potential additional members of our classessecond since the translation mappings may often be manytomany the alternarebater flutuar rolar deslizar derivar planar dative yes yes yes conative no no no causinch yes yes yes middle yes yes yes accept coref yes yes yes causinch yes yes yes yes yes yes resultative yes yes yes yee yes yee adject part yes yes yes ind action yes yes yes yes no yes locat invers yes yes yes yes yes yes measure yee yes yes yee yes yes adj perf no no no no no no cogn object no no no no no no zero nom yes yes no yes yes yes tions may depend on which translation is chosen potentially giving us different clusters but it is uncertain to what extent this is a factor and it also requires further investigationin this experiment we have tried to choose the portuguese verb that is most closely related to the description of the english verb in the levin classwe expect these crosslinguistic features to be useful for capturing translation generalizations between languages as discussed in the literature in pursuing this goal we are currently implementing features for motion verbs in the english treeadjoining grammar tag tags have also been applied to portuguese in previous work resulting in a small portuguese grammar we intend to extend this grammar building a more robust tag grammar for portuguese that will allow us to build an englishportuguese transfer lexicon using these features	P98-1046	investigating regular sense extensions based on intersective levin classesin this paper we specifically address questions of polysemy with respect to verbs and how regular extensions of meaning can be achieved through the adjunction of particular syntactic phraseswe see verb classes as the key to making generalizations about regular extensions of meaningcurrent approaches to english classification levin classes and wordnet have limitations in their applicability that impede their utility as general classification schemeswe present a refinement of levin classes intersective sets which are a more finegrained classification and have more coherent sets of syntactic frames and associated semantic componentswe have preliminary indications that the membership of our intersective sets will be more compatible with wordnet than the original levin classeswe also have begun to examine related classes in portuguese and find that these verbs demonstrate similarly coherent syntactic and semantic propertieswe show that multiple listings could in some cases be interpreted as regular sense extensions and defined intersective levin classes which are a more finegrained syntactically and semantically coherent refinement of basic levin classeswe argue that the use of syntactic frames and verb classes can simplify the definition of different verb senses
an ir approach for translating new words from nonparallel comparable texts in recent years there is a phenomenal growth in the amount of online text material available from the greatest information repository known as the world wide webvarious traditional information retrieval techniques combined with natural language processing techniques have been retargeted to enable efficient access of the wwwsearch engines indexing relevance feedback query term and keyword weighting document analysis document classification etcmost of these techniques aim at efficient online search for information already on the webmeanwhile the corpus linguistic community regards the www as a vast potential of corpus resourcesit is now possible to download a large amount of texts with automatic tools when one needs to compute for example a list of synonyms or download domainspecific monolingual texts by specifying a keyword to the search engine and then use this text to extract domainspecific termsit remains to be seen how we can also make use of the multilingual texts as nlp resourcesin the years since the appearance of the first papers on using statistical models for bilingual lexicon compilation and machine translation large amount of human effort and time has been invested in collecting parallel corpora of translated textsour goal is to alleviate this effort and enlarge the scope of corpus resources by looking into monolingual comparable textsthis type of texts are known as nonparallel corporasuch nonparallel monolingual texts should be much more prevalent than parallel textshowever previous attempts at using nonparallel corpora for terminology translation were constrained by the inadequate availability of samedomain comparable texts in electronic formthe type of nonparallel texts obtained from the ldc or university libraries were often restricted and were usually outofdate as soon as they became availablefor new word translation the timeliness of corpus resources is a prerequisite so is the continuous and automatic availability of nonparallel comparable texts in electronic formdata collection effort should not inhibit the actual translation effortfortunately nowadays the world wide web provides us with a daily increase of fresh uptodate multilingual material together with the archived versions all easily downloadable by software tools running in the backgroundit is possible to specify the url of the online site of a newspaper and the start and end dates and automatically download all the daily newspaper materials between those datesin this paper we describe a new method which combines ir and nlp techniques to extract new word translation from automatically downloaded englishchinese nonparallel newspaper textsto improve the performance of a machine translation system it is often necessary to update its bilingual lexicon either by human lexicographers or statistical methods using large corporaup until recently statistical bilingual lexicon compilation relies largely on parallel corporathis is an undesirable constraint at timesin using a broadcoverage englishchinese mt system to translate some text recently we discovered that it is unable to translate mel liougan which occurs very frequently in the textother words which the system cannot find in its 20000entry lexicon include proper names such as the taiwanese president lee tenghui and the hong kong chief executive tung cheehwato our disappointment we cannot locate any parallel texts which include such words since they only start to appear frequently in recent monthsa quick search on the web turned up archives of multiple local newspapers in english and chineseour challenge is to find the translation of 1133 i liougan and other words from this online nonparallel comparable corpus of newspaper materialswe choose to use issues of the english newspaper hong kong standard and the chinese newspaper mingpao from dec1297 to dec3197 as our corpusthe english text contains about 3 mb of text whereas the chinese text contains 88 mb of 2 byte character textsso both texts are comparable in sizesince they are both local mainstream newspapers it is reasonable to assume that their contents are comparable as wellunlike in parallel texts the position of a word in a text does not give us information about its translation in the other language suggest that a content word is closely associated with some words in its contextas a tutorial example we postulate that the words which appear in the context ofmet iliougan should be similar to the words appearing in the context of its english translation fluwe can form a vector space model of a word in terms of its context word indices similar to the vector space model of a text in terms of its constituent word indices the value of the ith dimension of a word vector w is f if the ith word in the lexicon appears f times in the same sentences as w left columns in table 1 and table 2 show the list of content words which appear most frequently in the context of flu and africa respectivelythe right column shows those which occur most frequently in the context of bvgwe can see that the context of at is more similar to that of flu than to that of africaso the first clue to the similarity between a word and its translation number of common words in their contextsin a bilingual corpus the quotcommon wordquot is actually a bilingual word pairwe use the lexicon of the mt system to quotbridgequot all bilingual word pairs in the corporathese word pairs are used as seed wordswe found that the contexts of flu and mi iliougan share 233 quotcommonquot context words whereas the contexts of africa and wiff iliougan share only 121 common words even though the context of flu has 491 unique words and the context of africa has 328 wordsin the vector space model wflu and wliougan has 233 overlapping dimensions whereas there are 121 overlapping dimensions between wflu and w africathe flu example illustrates that the actual ranking of the context word frequencies provides a second clue to the similarity between a bilingual word pairfor example virus ranks very high for both flu and me iliougan and is a strong quotbridgequot between this bilingual word pairthis leads us to use the term frequency measurethe tf of a context word is defined as the frequency of the word in the context of w however the tf of a word is not independent of its general usage frequencyin an extreme case the function word the appears most frequently in english texts and would have the highest tf in the context of any w in our hkstandardmingpao corpus hong kong is the most frequent content word which appears everywhereso in the flu example we would like to reduce the significance of hong kong tf while keeping that of virusa common way to account for this difference is by using the inverse document frequencyamong the variants of idf we choose the following representation from next a ranking algorithm is needed to match the unknown word vectors to their counterparts in the other languagea ranking algorithm selects the best target language candidate for a source language word according to direct comparison of some similarity measures we modify the similarity measure proposed by into the following so where maxn ni the maximum frequency of any word in the corpus the total number of occurrences of word i in the corpus so where wic wie veliwic2 x eliwie2 tfic tfie the idf of virus is 181 and that of hong kong is 123 in the english textthe idf of ws is 192 and that of hong kong is 083 in chineseso in both cases virus is a stronger quotbridgequot for biglliougan than hong konghence for every context seed word i we assign a word weighting factor w tfiw x idfi where tfiw is the tf of word i in the context of word w the updated vector space model of word w has wi in its ith dimensionthe ranking of the 20 words in the contexts of wf3 iliougan is rearranged by this weighting factor as shown in table3variants of similarity measures such as the above have been used extensively in the ir community they are mostly based on the cosine measure of two vectorsfor different tasks the weighting factor might varyfor example if we add the idf into the weighting factor we get the following measure si where wic tfic x idf wie tfie x idfi in addition the dice and jaccard coefficients are also suitable similarity measures for document comparison we also implement the dice coefficient into similarity measure s2 where wic tfie x idfi si is often used in comparing a short query with a document text whereas s2 is used in comparing two document textsreasoning that our objective falls somewhere in betweenwe are comparing segments of a document we also multiply the above two measures into a third similarity measure s3in using bilingual seed words such asnet virus as quotbridgesquot for terminology translation the quality of the bilingual seed lexicon naturally affects the system outputin the case of european language pairs such as frenchenglish we can envision using words sharing common cognates as these quotbridgesquotmost importantly we can assume that the word boundaries are similar in french and englishhowever the situation is messier with english and chinesefirst segmentation of the chinese text into words already introduces some ambiguity of the seed word identitiessecondly englishchinese translations are complicated by the fact that the two languages share very little stemming properties or partofspeech set or word orderthis property causes every english word to have many chinese translations and vice versain a sourcetarget language translation scenario the translated text can be quotrearrangedquot and cleaned up by a monolingual language model in the target languagehowever the lexicon is not very reliable in establishing quotbridgesquot between nonparallel englishchinese textsto compensate for this ambiguity in the seed lexicon we introduce a confidence weighting to each bilingual word pair used as seed wordsif a word ie is the kth candidate for word ic then wite wite kithe similarity scores then become s4 and s5 and s6 s4 x s5 we also experiment with other combinations of the similarity scores such as s7 so x s5all similarity measures s3 s7 are used in the experiment for finding a translation for me in order to apply the above algorithm to find the translation for blgt i liougan from the hkstandardmingpao corpus we first use a script to select the 118 english content words which are not in the lexicon as possible candidatesusing similarity measures s3 s7 the highest ranking candidates of ms are shown in table 6s6 and s7 appear to be the best similarity measureswe then test the algorithm with s7 on more chinese words which are not found in the lexicon but which occur frequently enough in the mingpao textsa statistical new word extraction tool can be used to find these wordsthe unknown chinese words and their english counterparts as well as the occurrence frequencies of these words in hkstandardmingpao are shown in table 4frequency numbers with a indicates that this word does not occur frequent enough to be foundchinese words with a indicates that it is a word with segmentation and translation ambiguitiesfor example 44 could be a family name or part of another word meaning forestwhen it is used as a family name it could be transliterated into lam in cantonese or lin in mandarindisregarding all entries with a in the above table we apply the algorithm to the rest of the chinese unknown words and the 118 english unknown words from hkstandardthe output is ranked by the similarity scoresthe highest ranking translated pairs are shown in table 5the only chinese unknown words which are not correctly translated in the above list are ik6 fl lunar and mr yeltsin 1 tungcheehwa is a pair of collocates which is actually the full name of the chief executivepoultry in chinese is closely related to flu because the chinese name for bird flu is poultry fluin fact almost all unambiguous chinese new words find their translations in the first 100 of the ranked listsix of the chinese words have correct translation as their first candidateusing vector space model and similarity measures for ranking is a common approach in ir for querytext and texttext comparisons this approach has also been used by for sense disambiguation between multiple usages of the same wordsome of the early statistical terminology translation methods are these algorithms all require parallel translated texts as inputattempts at exploring nonparallel corpora for terminology translation are very few among these proposes that the association between a word and its close collocate is preserved in any language and suggests that the associations between a word and many seed words are also preserved in another languagein this paper we have demonstrated that the associations between a word and its context seed words are wellpreserved in nonparallel comparable texts of different languagesour algorithm is the first to have generated a collocation bilingual lexicon albeit small from a nonparallel comparable corpuswe have shown that the algorithm has good precision but the recall is low due to the difficulty in extracting unambiguous chinese and english wordsbetter results can be obtained when the following changes are made we will test the precision and recall of the algorithm on a larger set of unknown wordswe have devised an algorithm using context seed word tfidf for extracting bilingual lexicon from nonparallel comparable corpus in englishchinesethis algorithm takes into account the reliability of bilingual seed words and is language independentthis algorithm can be applied to other language pairs such as englishfrench or englishgermanin these cases since the languages are more similar linguistically and the seed word lexicon is more reliable the algorithm should yield better resultsthis algorithm can also be applied in an iterative fashion where highranking bilingual word pairs can be added to the seed word list which in turn can yield more new bilingual word pairs	P98-1069	an ir approach for translating new words from nonparallel comparable textswe demonstrate that the associations between a word and its context seed words are preserved in comparable texts of different languageswe propose to represent the contexts of a word or phrase with a realvalued vector which one element corresponds to one word in the contexts
improving data driven wordclass tagging by system combination in this paper we examine how the differences in modelling between different data driven systems performing the same nlp task can be exploited to yield a higher accuracy than the best individual system we do this by means of an experiment involving the task of morphosyntactic wordclass tagging four wellknown tagger generators are trained on the same corpus data after comparison their outputs are combined using several voting strategies and second stage classifiers all combination taggers outperform their best component with the best combination showing a 191 lower error rate than the best individual tagger in all natural language processing systems we find one or more language models which are used to predict classify andor interpret language related observationstraditionally these models were categorized as either rulebasedsymbolic or corpusbasedprobabilisticrecent work has demonstrated clearly that this categorization is in fact a mixup of two distinct categorization systems on the one hand there is the representation used for the language model and on the other hand the manner in which the model is constructed data driven methods appear to be the more popularthis can be explained by the fact that in general hand crafting an explicit model is rather difficult especially since what is being modelled natural language is not wellunderstoodwhen a data driven method is used a model is automatically learned from the implicit structure of an annotated training corpusthis is much easier and can quickly lead to a model which produces results with a reasonably good qualityobviously reasonably good quality is not the ultimate goalunfortunately the quality that can be reached for a given task is limited and not merely by the potential of the learning method usedother limiting factors are the power of the hard and software used to implement the learning method and the availability of training materialbecause of these limitations we find that for most tasks we are faced with a ceiling to the quality that can be reached with any available machine learning systemhowever the fact that any given system cannot go beyond this ceiling does not mean that machine learning as a whole is similarly limiteda potential loophole is that each type of learning method brings its own inductive bias to the task and therefore different methods will tend to produce different errorsin this paper we are concerned with the question whether these differences between models can indeed be exploited to yield a data driven model with superior performancein the machine learning literature this approach is known as ensemble stacked or combined classifiersit has been shown that when the errors are uncorrelated to a sufficient degree the resulting combined classifier will often perform better than all the individual systems the underlying assumption is twofoldfirst the combined votes will make the system more robust to the quirks of each learner particular biasalso the use of information about each individual method behaviour in principle even admits the possibility to fix collective errorswe will execute our investigation by means of an experimentthe nlp task used in the experiment is morphosyntactic wordclass taggingthe reasons for this choice are severalfirst of all tagging is a widely researched and wellunderstood task 1998second current performance levels on this task still leave room for improvement tate of the art performance for data driven automatic wordclass taggers is 9697 correctly tagged wordsfinally a number of rather different methods are available that generate a fully functional tagging system from annotated textin 1992 van halteren combined a number of taggers by way of a straightforward majority vote since the component taggers all used ngram statistics to model context probabilities and the knowledge representation was hence fundamentally the same in each component the results were limitednow there are more varied systems available a variety which we hope will lead to better combination effectsfor this experiment we have selected four systems primarily on the basis of availabilityeach of these uses different features of the text to be tagged and each has a completely different representation of the language modelthe first and oldest system uses a traditional trigram model based on context statistics p and lexical statistics p directly estimated from relative corpus frequenciesthe viterbi algorithm is used to determine the most probable tag sequencesince this model has no facilities for handling unknown words a memorybased system is used to propose distributions of potential tags for words not in the lexiconthe second system is the transformation based learning system as described by brill this 1 brill system is available as a collection of c programs and perl scripts at ftp ftp cs jhu edupubbrillprograms rule_based_tagger_v 1 14 tarz system starts with a basic corpus annotation and then searches through a space of transformation rules in order to reduce the discrepancy between its current annotation and the correct one during tagging these rules are applied in sequence to new textof all the four systems this one has access to the most information contextual information as well as lexical information however the actual use of this information is severely limited in that the individual information items can only be combined according to the patterns laid down in the rule templatesthe third system uses memorybased learning as described by daelemans et al during the training phase cases containing information about the word the context and the correct tag are stored in memoryduring tagging the case most similar to that of the focus word is retrieved from the memory which is indexed on the basis of the information gain of each feature and the accompanying tag is selectedthe system used here has access to information about the focus word and the two positions before and after at least for known wordsfor unknown words the single position before and after three suffix letters and information about capitalization and presence of a hyphen or a digit are usedthe fourth and final system is the mxpost system as described by ratnaparkhi it uses a number of word and context features rather similar to system m and trains a maximum entropy model that assigns a weighting parameter to each featurevalue and combination of features that is relevant to the estimation of the probability pa beam search is then used to find the highest probability tag sequenceboth this system and brill system are used with the default settings that are suggested in their documentationthe data we use for our experiment consists of the tagged lob corpus the corpus comprises about one million words divided over 500 samples of 2000 words from 15 text typesits tagging which was manually checked and corrected is generally accepted to be quite accuratehere we use a slight adaptation of the tagsetthe changes are mainly cosmetic eg nonalphabetic characters such as quot8quot in tag names have been replacedhowever there has also been some retokenization genitive markers have been split off and the negative marker quotntquot has been reattachedan example sentence tagged with the resulting tagset is the ati singular or plural lord npt article major npt singular titular extended vbd noun an at singular titular invitation nn noun to in past tense of verb all abn singular article the ati singular common parliamentary jj noun candidates nns preposition sper prequantifier singular or plural article adjective plural common noun period the tagset consists of 170 different tags and has an average ambiguity of 269 tags per wordformthe difficulty of the tagging task can be judged by the two baseline measurements in table 2 below representing a completely random choice from the potential tags for each token and selection of the lexically most likely tag for our experiment we divide the corpus into three partsthe first part called train consists of 80 of the data constructed 3ditto tags are used for the components of multitoken units eg if quotas well asquot is taken to be a coordination conjunction it is tagged quotas_cc1 well_cc2 as_cc3quot using three related but different ditto tags by taking the first eight utterances of every tenthis part is used to train the individual taggersthe second part tune consists of 10 of the data and is used to select the best tagger parameters where applicable and to develop the combination methodsthe third and final part test consists of the remaining 10 and is used for the final performance measurements of all taggersboth tune and test contain around 25 new tokens and a further 02 known tokens with new tagsthe data in train and tune is to be the only information used in tagger construction all components of all taggers are to be entirely data driven and no manual adjustments are to be donethe data in test is never to be inspected in detail but only used as a benchmark tagging for quality measurementquotin order to see whether combination of the component taggers is likely to lead to improvements of tagging quality we first examine the results of the individual taggers when applied to tuneas far as we know this is also one of the first rigorous measurements of the relative quality of different tagger generators using a single tagset and dataset and identical circumstancesthe quality of the individual taggers certainly still leaves room for improvement although tagger e surprises us with an accuracy well above any results reported so far and makes us less confident about the gain to be accomplished with combinationhowever that there is room for improvement is not enoughas explained above for combination to lead to improvement the component taggers must differ in the errors that they makethat this is indeed the case can be seen in table 1it shows that for 9922 of tune at least one tagger selects the correct taghowever it is unlikely that we will be able to identify this terns between the brackets give the distribution of correctincorrect tags over the systems tag in each casewe should rather aim for optimal selection in those cases where the correct tag is not outvoted which would ideally lead to correct tagging of 9821 of the words there are many ways in which the results of the component taggers can be combined selecting a single tag from the set proposed by these taggersin this and the following sections we examine a number of themthe accuracy measurements for all of them are listed in table 25 the most straightforward selection method is an nway voteeach tagger is allowed to vote for the tag of its choice and the tag with the highest number of votes is selected6 the question is how large a vote we allow each taggerthe most democratic option is to give each tagger one vote however it appears more useful to give more weight to taggers which have proved their qualitythis can be general quality eg each tagger votes its overall precision or quality in relation to the current situation eg each tagger votes its precision on the suggested tag the information about each tagger quality is derived from an inspection of its results on tune5for any tag x precision measures which percentage of the tokens tagged x by the tagger are also tagged x in the benchmark and recall measures which percentage of the tokens tagged x in the benchmark are also tagged x by the taggerwhen abstracting away from individual tags precision and recall are equal and measure how many tokens are tagged correctly in this case we also use the more generic term accuracy6in our experiment a random selection from among the winning tags is made whenever there is a tiebut we have even more information on how well the taggers performwe not only know whether we should believe what they propose but also know how often they fail to recognize the correct tag this information can be used by forcing each tagger also to add to the vote for tags suggested by the opposition by an amount equal to 1 minus the recall on the opposing tag as it turns out all voting systems outperform the best single tagger e7 also the best voting system is the one in which the most specific information is used precisionrecallhowever specific information is not always superior for totprecision scores higher than tagprecisionthis might be explained by the fact that recall information is missing even the worst combinator majority is significantly better than e using mcnemar chisquare p0so far we have only used information on the performance of individual taggersa next step is to examine them in pairswe can investigate all situations where one tagger suggests t1 and the other t2 and estimate the probability that in this situation the tag should actually be tx eg if e suggests dt and t suggests cs the probabilities for the appropriate tag are cs subordinating conjunction 03276 dt determiner 06207 ql quantifier 00172 wpr whpronoun 00345 when combining the taggers every tagger pair is taken in turn and allowed to vote for each possible tag ie not just the ones suggested by the component taggersif a tag pair t1t2 has never been observed in tune we fall back on information on the individual taggers viz the probability of each tag tx given that the tagger suggested tag tinote that with this method a tag suggested by a minority of the taggers still has a chance to winin principle this could remove the restriction of gain only in 22 and 1111 casesin practice the chance to beat a majority is very slight indeed and we should not get our hopes up too high that this should happen very oftenwhen used on test the pairwise voting strategy clearly outperforms the other voting strategies8 but does not yet approach the level where all tying majority votes are handled correctly from the measurements so far it appears that the use of more detailed information leads to a better accuracy improvementit ought therefore to be advantageous to step away from the underlying mechanism of voting and to model the situations observed in tune more closelythe practice of feeding the outputs of a number of classifiers as features for a next learner is usually called stacking the second stage can be provided with the first level outputs and with additional information eg about the original input patternthe first choice for this is to use a memorybased second level learnerin the basic version each case consists of the tags suggested by the component taggers and the correct tagin the more advanced versions we also add information about the word in question and the tags suggested by all taggers for the previous and the next position for the first two the similarity metric used during tagging is a straightforward overlap count for the third we need to use an information gain weighting surprisingly none of the memorybased based methods reaches the quality of tagpair9 the explanation for this can be found when we examine the differences within the memorybased general strategy the more feature information is stored the higher the accuracy on tune but the lower the accuracy on testthis is most likely an overtraining effect tune is probably too small to collect case bases which can leverage the stacking effect convincingly especially since only 751 of the second stage material shows disagreement between the featured tagsto examine if the overtraining effects are specific to this particular second level classifier we also used the c50 system a commercial version of the wellknown program c45 for the induction of decision trees on the same training materia119 because c50 prunes the decision tree the overfitting of training material is less than with memorybased learning but the results on test are also worsewe conjecture that pruning is not beneficial when the interesting cases are very rareto realise the benefits of stacking either more data is needed or a second stage classifier that is better suited to this type of problemthe relation between the accuracy of combinations and that of the individual taggers is shown in table 3the most important observation is that every combination outperforms the combination of any strict subset of its componentsalso of note is the improvement yielded by the best combinationthe pairwise voting system using all four individual taggers scores 9792 correct on test a 191 reduction in error rate over the best individual system viz the maximum entropy tagger a major factor in the quality of the combination results is obviously the quality of the best component all combinations with e score higher than those without e after that the decisive factor appears to be the difference in language model t is generally a better combiner than m and r12 even though it has the lowest accuracy when operating alonea possible criticism of the proposed combi11by a margin at the edge of significance p0060812although not significantly better eg the differences within the group meeret are not significant nation scheme is the fact that for the most successful combination schemes one has to reserve a nontrivial portion of the annotated data to set the parameters for the combinationto see whether this is in fact a good way to spend the extra data we also trained the two best individual systems on a concatenation of train and tune so that they had access to every piece of data that the combination had seenit turns out that the increase in the individual taggers is quite limited when compared to combinationthe more extensively trained e scored 9751 correct on test and m 9707 our experiment shows that at least for the task at hand combination of several different systems allows us to raise the performance ceiling for data driven systemsobviously there is still room for a closer examination of the differences between the combination methods eg the question whether memorybased combination would have performed better if we had provided more training data than just tune and of the remaining errors eg the effects of inconsistency in the data regardless of such closer investigation we feel that our results are encouraging enough to extend our investigation of combination starting with additional component taggers and selection strategies and going on to shifts to other tagsets andor languagesbut the investigation need not be limited to wordclass tagging for we expect that there are many other nlp tasks where combination could lead to worthwhile improvementsour thanks go to the creators of the tagger generators used here for making their systems available	P98-1081	improving data driven wordclass tagging by system combinationin this paper we examine how the differences in modelling between different data driven systems performing the same nlp task can be exploited to yield a higher accuracy than the best individual systemwe do this by means of an experiment involving the task of morphosyntactic wordclass taggingfour wellknown tagger generators are trained on the same corpus dataafter comparison their outputs are combined using several voting strategies and second stage classifiersall combination taggers outperform their best component with the best combination showing a 191 lower error rate than the best individual taggerwe suggest three voting strategies equal vote where each classifier vote is weighted equally overall accuracy where the weight depends on the overall accuracy of a classifier and pairwise voting
pseudoprojectivity a polynomially parsable nonprojective dependency grammar dependency grammar has a long tradition in syntactic theory dating back to at least tesniere work from the thirtiesrecently it has gained renewed attention as empirical methods in parsing are discovering the importance of relations between words which is what dependency grammars model explicitly do but contextfree phrasestructure grammars do notone problem that has posed an impediment to more widespread acceptance of dependency grammars is the fact that there is no computationally tractable version of dependency grammar which is not restricted to projective analyseshowever it is well known that there are some syntactic phenomena that require nonprojective analysesin this paper we present a form of projectivity which we call pseudoprojectivity and we present a generative stringrewriting formalism that can generate pseudoprojective analyses and which is polynomially parsablethe paper is structured as followsin section 2 we introduce our notion of pseudoprojectivitywe briefly review a previously proposed formalization of projective dependency grammars in section 3in section 4 we extend this formalism to handle pseudoprojectivitywe informally present a parser in section 5we will use the following terminology and notation in this paperthe hierarchical order between the nodes of a tree t will be represented with the symbol t and whenever they are unambiguous the notations and will be usedwhen x y we will say that x is a descendent of y and y an ancestor of xthe projection of a node x belonging to a tree t is the set of the nodes y of t such that y xan arc between two nodes y and x of a tree t directed from y to x will be noted either or the node x will be referred to as the dependent and y as the governorthe latter will be noted when convenient xt the notations t and x are unambiguous because a node x has at most one governor in a treeas usual an ordered tree is a tree enriched with a linear order over the set of its nodesfinally if 1 is an arc of an ordered tree t then supp represents the support of 1 ie the set of the nodes of t situated between the extremities of 1 extremities includedwe will say that the elements of supp are covered by 1the notion of projectivity was introduced by and has received several different definitions since thenthe definition given here is borrowed from and definition an arc t is projective if and only if for every y covered by y x a tree t is projective if and only if every arc of t is projective a projective tree has been represented in figure 1a projective dependency tree can be associated with a phrase structure tree whose constituents are the projections of the nodes of the dependency treeprojectivity is therefore equivalent in phrase structure markers to continuity of constituentthe strong constraints introduced by the projectivity property on the relationship between hierarchical order and linear order allow us to describe word order of a projective dependency tree at a local level in order to describe the linear position of a node it is sufficient to describe its position towards its governor and sister nodesthe domain of locality of the linear order rules is therefore limited to a subtree of depth equal to oneit can be noted that this domain of locality is equal to the domain of locality of subcategorization rulesboth rules can therefore be represented together as in or separately as will be proposed in 3although most linguistic structures can be represented as projective trees it is well known that projectivity is too strong a constraint for dependency trees as shown by the example of figure 2 which includes a nonprojective arc who do you think she invited the non projective structures found in linguistics represent a small subset of the potential non projective structureswe will define a property weaker than projectivity called pseudoprojectivity which describes a subset of the set of ordered dependency trees containing the nonprojective linguistic structuresin order to define pseudoprojectivity we introduce an operation on dependency trees called liftingwhen applied to a tree this operation leads to the creation of a second tree a lift of the first onean ordered tree t is a lift of the ordered tree t if and only if t and t have the same nodes in the same order and for every node x xt txt we will say that the node x has been lifted from xt to xti recall that the linear position of a node in a projective tree can be defined relative to its governor and its sistersin order to define the linear order in a non projective tree we will use a projective lift of the treein this case the position of a node can be defined only with regards to its governor and sisters in the lift ie its linear governor and sistersdefinition an ordered tree t is said pseudoprojective if there exists a lift t of tree t which is projectiveif there is no restriction on the lifting the previous definition is not very interesting since we can in fact take any nonprojective tree and lift all nodes to the root node and obtain a projective treewe will therefore constrain the lifting by a set of rules called lifting rulesconsider a set of categoriesthe following definitions make sense only for trees whose nodes are labeled with categoriesthe lifting rules are of the following form this rule says that a node of category ld can be lifted from its syntactic governor of category sg to its linear governor of category lg through a path consisting of nodes of category c1 cn where the string belongs to levery set of lifting rules defines a particular property of pseudoprojectivity by imposing particular constraints on the liftinga 21t is possible to define pseudoprojectivity purely structurally for example we can impose that each node x is lifted to the highest ancestor of x covered by t the resulting pseudoprojectivity is a fairly weak extension to projectivity which nevertheless covers major nonprojective linguistic structureshowever we do not pursue a purely structural definition of pseudoprojectivity in this paper linguistic example of lifting rule is given in section 4the idea of building a projective tree by means of lifting appears in and is used by and this idea can also be compared to the notion of word order domain to the slash feature of gpsg and hpsg to the functional uncertainty of lfg and to the movea of gb theorywe define a projective dependency grammar as a stringrewriting system3 by giving a set of categories such as n v and adv4 a set of distinguished start categories a mapping from strings to categories and two types of rules dependency rules which state hierarchical order and lp rules which state linear orderthe dependency rules are further subdivided into subcategorization rules and modification rules here are some sample srules lp rules are represented as regular expressions associated with each categorywe use the hash sign to denote the position of the governor for example 3we follow throughout this paper by modeling a dependency grammar with a stringrewriting systemhowever we will identify a derivation with its representation as a tree and we will sometimes refer to symbols introduced in a rewrite step as quotdependent nodesquotfor a model of a dg based on treerewriting see we will call this system generative dependency grammar or gdg for shortderivations in gdg are defined as followsin a rewrite step we choose a multiset of dependency rules which contains exactly one srule and zero or more mrulesthe lefthand side nonterminal is the same as that we want to rewritecall this multiset the rewritemultisetin the rewriting operation we introduce a multiset of new nonterminals and exactly one terminal symbol the rewriting operation then must meet the following three conditions as an example consider a grammar containing the three dependency rules di d2 and d3 as well as the lp rule pi in addition we have some lexical mappings and the start symbol is vfinitea sample derivation is shown in figure 3 with the sentential form representation on top and the corresponding tree representation belowusing this kind of representation we can derive a bottomup parser in the following straightforward manner5 since syntactic and linear governors coincide we can derive deterministic finitestate machines which capture both the dependency and the lp rules for a given governor categorywe will refer to these fsms as rulefsms and if the governor is of category c we will refer to a crulefsmin a rulefsm the transitions are labeled by categories and the transition corresponding to the governor labeled by its category and a special mark this transition is called the quothead transitionquotthe entries in the parse matrix m are of the form where in is a rulefsm and q a state of it except for the entries in squares m 1 j n which also contain category labelslet wo wn be the input wordwe initialize the parse matrix as followslet c be a category of word wifirst we add c to mthen we add to m every pair such that m is a rulefsm with a transition labeled c from a start state and q the state reached after that transition6 embedded in the usual three loops on i j k we add an entry to m if is in m is in m q2 is a final state of m2 m2 is a crulefsm and mi transitions from qi to q on c there is a special case for the head transitions in mi if k i 1 c is in m mi is a crulefsm and there is a head transition from qi to q in ml then we add to mthe time complexity of the algorithm is 0 where g is the number of rulefsms derived from the dependency and lp rules in the grammar and qmax is the maximum number of states in any of the rulefsmsrecall that in a pseudoprojective tree we make a distinction between a syntactic governor and a linear governora node can be quotliftedquot along a lifting path from being a dependent of its syntactic governor to being a dependent of its linear this type of parser has been proposed previouslysee for example who also discuss earlystyle parsers for projective dependency grammarswe can use precomputed topdown prediction to limit the number of pairs added governor which must be an ancestor of the governorin defining a formal rewriting system for pseudoprojective trees we will not attempt to model the quotliftingquot as a transformational step in the derivationrather we will directly derive the quotliftedquot version of the tree where a node is dependent of its linear governorthus the derived structure resembles more a unistratal dependency representation like those used by than the multistratal representations of for example however from a formal point of view the distinction is not significantin order to capture pseudoprojectivity we will interpret rules of the form and as introducing syntactic dependents which may lift to a higher linear governoran lp rule of the form orders all linear dependents of the linear governor no matter whose syntactic dependents they arein addition we need a third type of rule namely a lifting rule or 1rule the 1rule can be rewrited on the following form this rule resembles normal dependency rules but instead of introducing syntactic dependents of a category it introduces a lifted dependentbesides introducing a linear dependent ld a 1rule should make sure that the syntactic governor of ld will be introduced at a later stage of the derivation and prevent it to introduce ld as its syntactic dependent otherwise non projective nodes would be introduced twice a first time by their linear governor and a second time by their syntactic governorthis condition is represented in the rule by means of a constraint on the categories found along the lifting paththis condition which we call the lifting condition is represented by the regular expression lg w sgthe regular expression representing the lifting condition is enriched with a dot separating on its left the part of the lifting path which has already been introduced during the rewriting and on its right the part which is still to be introduced for the rewriting to be validthe dot is an unperfect way of representing the current state in a finite state automaton equivalent to the regular expressionwe can further notice that the lifting condition ends with a repetition of ld for reasons which will be made clear when discussing the rewriting processa sentential form contains terminal strings and categories paired with a multiset of lifting conditions called the lift multisetthe lift multiset associated to a category c contains transiting lifting conditions introduced by ancestors of c and passing across c three cases must be distinguished when rewriting a category c and its lifting multiset lm lm contains a single lifting condition which dot is situated to its right lg w sg cin such a case c must be rewritten by the empty stringthe situation of the dot at the right of the lifting condition indicates that c has been introduced by its syntactic governor although it has already been introduced by its linear governor earlier in the rewriting processthis is the reason why c has been added at the end of the lifting condition of the 1rules used in the rewriting operation3the lifting conditions contained in the lift multiset of all the newly introduced dependents d should be compatible with d with the dot advanced appropriatelyin addition we require that when we rewrite a category as a terminal the lift multiset is emptylet us consider an examplesuppose we have have a grammar containing the dependency rules di d2 and d3 the lp rule pi and p2 this rule says that an objective whnoun with feature top which depends on a verb with no further restrictions can raise to any verb that dominates its immediate governor as long as the raising paths contains only verb with feature bridge ie bridge verbsa sample derivation is shown in figure 4 with the sentential form representation on top and the corresponding tree representation belowwe start our derivation with the start symbol klause and rewrite it using dependency rules d2 and d3 and the lifting rule 1 which introduces an objective np argumentthe lifting condition of i is passed to the v dependent but the dot remains at the left of vbridge because of the kleene starwhen we rewrite the embedded v we choose to rewrite again with klause and the lifting condition is passed on to the next verbthis verb is a vrans which requires a nobjthe lifting condition is passed to nobj and the dot is moved to the right of the regular expression therefore nobj is rewritten as the empty stringin this section we show that pseudoprojective dependency grammars as defined in section 23 are polynomially parsablewe can extend the bottomup parser for gdg to a parser for ppgdg in the following mannerin ppgdg syntactic and linear governors do not necessarily coincide and we must keep track separately of linear precedence and of lifting the entries in the parse matrix m are of the form where m is a rulefsm q a state of m and lm is a multiset of lifting conditions as defined in section 4an entry in a square m of the parse matrix means that the subword wi wj of the entry can be analyzed by in up to state q but that nodes corresponding to the lifting rules in lm are being lifted from the subtrees spanning wz wjput differently in this bottomup view lm represents the set of nodes which have a syntactic governor in the subtree spanning id w3 and a lifting rule but are still looking for a linear governorsuppose we have an entry in the parse matrix m of the form as we traverse the crulefsm m we recognize one by one the linear dependents of a node of category c call this governor n the action of adding a new entry to the parse matrix corresponds to adding a single new linear dependent to n each new dependent 71 brings with it a multiset of nodes being lifted from the subtree it is the root ofcall this multiset lmthe new entry will be therefore when we have reached the final state of a rulefsm we must connect up all nodes and lifting conditions before we can proceed to put an entry in the parse matrixthis involves these steps 1for every lifting condition in lm we ensure that it is compatible with the category of 77this is done by moving the dot leftwards in accordance with the category of the obvious special provisions deal with the kleene star and optional elementsif the category matches a catgeory with kleene start in the lifting condition we do not move the dotif the category matches a category which is to the left of an optional category or to the left of category with kleene star then we can move the dot to the left of that categoryif the dot cannot be placed in accordance with the category of ij then no new entry is made in the parse matrix for n 2we then choose a multiset of s m and 1rules whose lefthand side is the category of n for every dependent of n introduced by an 1rule the dependent must be compatible with an instance of a lifting condition in lm the lifting condition is then removed from l 3if after the above repositioning of the dot and the linking up of all linear dependents to lifting conditions there are still lifting conditions in lm such that the dot is at the beginning of the lifting condition then no new entry is made in the parse matrix for n 4for every syntactic dependent of rh we determine if it is a linear dependent of n which has not yet been identified as liftedfor each syntactic dependents which is not also a linear dependent we check whether there is an applicable lifting ruleif not no entry is made in the parse matrix for 77if yes we add the lifting rule to lmthis procedure determines a new multiset lm so we can add entry in the parse matrixthe parse is complete if there is an entry in square m of the parse matrix where in is a crulefsm for a start category and qm is a final state of m if we keep backpointers at each step in the algorithm we have a compact representation of the parse forestthe maximum number of entries in each square of the parse matrix is 0 where g is the number of rulefsms corresponding to lp rules in the grammar q is the maximum number of states in any of the rulefsms and l is the maximum number of states that the lifting rules can be in note that the exponent is a grammar constant but this number can be rather small since the lifting rules are not lexicalized they are constructionspecific not lexemespecificthe time complexity of the algorithm is therefore 0	P98-1106	pseudoprojectivity a polynomially parsable nonprojective dependency grammarthe pseudoprojective grammar we propose can be parsed in polynomial time and captures nonlocal dependencies through a form of gapthreading but the structures generated by the grammar are strictly projective
role of verbs in document analysis representations of context for the detection and corof malapropisms an electronic lexical database and some of its applications we present techniques to characterize document type and event by using semantic classification of verbsthe intuition motivating our research is illustrated by an examination of the role of nouns and verbs in documentsthe listing below shows the ontological categories which express the fundamental conceptual components of propositions using the framework of jackendoff each category permits the formation of a whquestion eg for thing quotwhat did you buyquot can be answered by the noun quota fishquotthe whquestions for action and event can only be answered by verbal constructions eg in the question quotwhat did you doquot where the response must be a verb eg jog write fall etcthing direction action place manner event amount the distinction in the ontological categories of nouns and verbs is reflected in information extraction systemsfor example given the noun phrases fares and us air that occur within a particular article the reader will know what the story is about ie fares and us airhowever the reader will not know the event ie what happened to the fares or to us airdid airfare prices rise fall or stabilizethese are the verbs most typically applicable to prices and which embody the eventmany natural language analysis systems focus on nouns and noun phrases in order to identify information on who what and wherefor example in summarization barzilay and elhadad and lin and hovy focus on multiword noun phrasesfor information extraction tasks such as the darpasponsored message understanding conferences only a few projects use verb phrases egappelt et al lin in contrast the named entity task which identifies nouns and noun phrases has generated numerous projects as evidenced by a host of papers in recent conferences although rich information on nominal participants actors and other entities is provided the named entity task provides no information on what happened in the document ie the event or actionless progress has been made on ways to utilize verbal information efficientlyin earlier systems with stemming many of the verbal and nominal forms were conflated sometimes erroneouslywith the development of more sophisticated tools such as part of speech taggers more accurate verb phrase identification is possiblewe present in this paper an effective way to utilize verbal information for document type discriminationour initial observations suggested that both occurrence and distribution of verbs in news articles provide meaningful insights into both article type and contentexploratory analysis of parsed wall street journal data2 suggested that articles characterized by movement verbs such as drop plunge or fall have a different event profile from articles with a high percentage of communication verbs such as report say comment or complainhowever without associated nominal arguments it is impossible to know whether the thing that drops refers to airfare prices or projected earningsin this paper we assume that the set of verbs in a document when considered as a whole can be viewed as part of the conceptual map of the events and action in a document in the same way that the set of nouns has been used as a concept map for entitiesthis paper reports on two methods using verbs to determine an event profile of the document while also reliably categorizing documents by typeintuitively the event profile refers to the classification of an article by the kind of eventfor example the article could be a discussion event a reporting event or an argument eventto illustrate consider a sample article from wsj of average length with a high percentage of communication verbsthe profile of the article shows that there are 19 verbs 11 are communication verbs including add report say and tellother verbs include be skeptical carry produce and closerepresentative nouns include polaroid corp michael ellmann wertheim schroder ey co prudentialbache savings operating results gain revenue cuts profit loss sales analyst and spokesmanin this case the verbs clearly contribute information that this article is a report with more opinions than new factsthe preponderance of communication verbs coupled with proper noun subjects and human nouns suggest a discussion articleif verbs are ignored this fact would be overlookedmatches on frequent nouns like gain and loss do not discriminate this article from one which announces a gain or loss as breaking news indeed according to our results a breaking news article would feature a higher percentage of motion verbs rather than verbs of communicationverbs are an important factor in providing an event profile which in turn might be used in categorizing articles into different genresturning to the literature in genre classification biber outlines five dimensions which can be used to characterize genreproperties for distinguishing dimensions include verbal features such as tense agentless passives and infinitivesbiber also refers to three verb classes private public and suasive verbskarlgren and cutting take a computationally tractable set of these properties and use them to compute a score to recognize text genre using discriminant analysisthe only verbal feature used in their study is presenttense verb countas karlgren and cutting show their techniques are effective in genre categorization but they do not claim to show how genres differkessler et al discuss some of the complexities in automatic detection of genre using a set of computationally efficient cues such as punctuation abbreviations or presence of latinate suffixesthe taxonomy of genres and facets developed in kessler et al is useful for a wide range of types such as found in the brown corpusalthough some of their discriminators could be useful for news articles the indicators do not appear to be directly applicable to a finer classification of news articlesnews articles can be divided into several standard categories typically addressed in journalism textbookswe base our article category ontology shown in lowercase on hill and breen in uppercase the goal of our research is to identify the role of verbs keeping in mind that event profile is but one of many factors in determining text typein our study we explored the contribution of verbs as one factor in document type discrimination we show how article types can be successfully classified within the news domain using verb semantic classeswe initially considered two specific categories of verbs in the corpus communication verbs and support verbsin the wsj corpus the two most common main verbs are say a communication verb and be a support verbin addition to say other high frequency communication verbs include report announce and statein journalistic prose as seen by the statistics in table 1 at least 20 of the sentences contain communication verbs such as say and announce these sentences report point of view or indicate an attributed commentin these cases the subordinated complement represents the main event eg in quotadvisors announced that ibm stock rose 36 points over a three year periodquot there are two actions announce and risein sentences with a communication verb as main verb we considered both the main and the subordinate verb this decision augmented our verb count an additional 20 and even more importantly further captured information on the actual event in an article not just the communication eventas shown in table 1 support verbs such as go or get constitute 30 and other content verbs such as fall adapt recognize or vow make up the remaining 50if we exclude all support type verbs 70 of the verbs yield information in answering the question quotwhat happenedquot or quotwhat did x doquotsince our first intuition of the data suggested that articles with a preponderance of verbs of communication say announce support have get go remainder abuse claim offer table 1 approximate frequency of verbs by type from the wall street journal a certain semantic type might reveal aspects of document type we tested the hypothesis that verbs could be used as a predictor in providing an event profilewe developed two algorithms to explore wordnet to cluster related verbs and build a set of verb chains in a document much as morris and hirst used roget thesaurus or like hirst and st onge used wordnet to build noun chains classify verbs according to a semantic classification system in this case using levin english verb classes and alternations as a basisfor source material we used the manuallyparsed linguistic data consortium wall street journal corpus from which we extracted main and complement of communication verbs to test the algorithms onusing wordnetour first technique was to use wordnet to build links between verbs and to provide a semantic profile of the documentwordnet is a general lexical resource in which words are organized into synonym sets each representing one underlying lexical concept these synonym sets or synsets are connected by different semantic relationships such as hypernymy synonymy antonymy and others the determination of relatedness via taxonomic relations has a rich history the premise is that words with similar meanings will be located relatively close to each other in the hierarchyfigure 1 shows the verbs cite and post which are related via a common ancestor inform let knowthe wnverber toolwe used the hypernym relationship in wordnet because of its high coveragewe counted the number of edges needed to find a common ancestor for a pair of verbsgiven the hierarchical structure of wordnet the lower the edge count in principle the closer the verbs are semanticallybecause wordnet common ancestor inform__ let know testify to indicate announce abduce cite attest report post sound allows individual words to be the descendent of possibly more than one ancestor two words can often be related by more than one common ancestor via different paths possibly with the same relationship results from wnverberwe ran all articles longer than 10 sentences in the wsj corpus through wnverberoutput showed that several verbs eg go take and say participate in a very large percentage of the high frequency synsets this is due to the width of the verb forest in wordnet top level verb synsets tend to have a large number of descendants which are arranged in fewer generations resulting in a flat and bushy tree structurefor example a top level verb synset inform give information let know has over 40 children whereas a similar top level noun synset entity only has 15 childrenas a result using fewer than two levels resulted in groupings that were too limited to aggregate verbs effectivelythus for our system we allowed up to two edges to intervene between a common ancestor synset and each of the verbs respective synsets as in figure 2in addition to the problem of the flat nature of the verb hierarchy our results from wnverber are degraded by ambiguity similar effects have been reported for nounsverbs with differences in high versus low frequency senses caused certain verbs to be incorrectly related for example have and drop are related by the synset meaning quotto give birthquot although this sense of drop is rare in wsjthe results of wnverber in table 2 reflect the effects of bushiness and ambiguitythe five most frequent synsets are given in column 1 column 2 shows some typical verbs which participate in the clustering column 3 shows the type of article which tends to contain these synsetsmost articles end up in the top five nodesthis illustrates the ineffectiveness of these most frequent wordnet synset to discriminate between article typesevaluation using kendall tauwe sought independent confirmation to assess the correlation between two variables rank for wnverber resultsto evaluate the effects of one synset frequency on another we used kendall tau rank order statistic for example was it the case that verbs under the synset act tend not to occur with verbs under the synset thinkif so do articles with this property fit a particular profilein our results we have information about synset frequency where each of the 1236 articles in the corpus constitutes a sampletable 3 shows the results of calculating kendall t with considerations for ranking ties for all 45 pairing combinations of the top 10 most frequently occurring synsetscorrelations can range from 10 reflecting inverse correlation to 10 showing direct correlation ie the presence of one class increases as the presence of the correlated verb class increasesa r value of 0 would show that the two variables values are independent of each otherresults show a significant positive correlation between the synsetsthe range of correlation is from 850 between the communication verb synset and the act verb synset to 238 between the think verb synset and the change state verb synset these correlations show that frequent synsets do not behave independently of each other and thus confirm that the wordnet results are not an effective way to achieve document discriminationalthough the wordnet results were not discriminatory we were still convinced that our initial hypothesis on the role of verbs in determining event profile was worth pursuingwe believe that these results are a byproduct of lexical ambiguity and of the richness of the wordnet hierarchywe thus decided to pursue a new approach to test our hypothesis one which turned out to provide us with clearer and more robust resultsutilizing evcaa different approach to test the hypothesis was to use another semantic categorization method we chose the semantic classes of levin evca as a basis for our next analysis3 levin seminal work is based on the timehonored observation that verbs which participate in similar syntactic alternations tend to share semantic propertiesthus the behavior of a verb with respect to the expression and interpretation of its arguments can be said to be in large part determined by its meaninglevin has meticulously set out a list of syntactic tests which predict membership in no less than 48 classes each of which is divided into numerous subclassesthe rigor and thoroughness of levin study permitted us to encode our algorithm evcaverber on a subset of the evca classes ones which were frequent in our corpusfirst we manually categorized the 100 most frequent verbs as well as 50 additional verbs which covers 56 of the verbs by token in the corpuswe subjected each verb to a set of strict linguistic tests as shown in table 4 and verified primary verb usage against the corpusresults from evcaverberin order to be able to compare article types and emphasize their differences we selected articles that had the highest percentage of a particular verb class from each of the ten verb classes we chose five articles from each evca class yielding a total of 50 articles for analysis from the full set of 1236 articleswe observed that each class discriminated between different article types as shown in table 5in contrast to table 2 the article types are well discriminated by verb classfor example a concentration of communication class verbs indicated that the article type was a general announcement of short or medium length or a longer feature article with many opinions in the textarticles high in motion verbs were also announcements but differed from the communication ones in that they were commonly postings of company earnings reaching a new high or dropping from last quarteragreement and argument verbs appeared in many of the same articles involving issues of some controversyhowever we noted that articles with agreement verbs were a superset of the argument ones in that in our corpus argument verbs did not appear in articles concerning joint ventures and mergersarticles marked by causative class verbs tended to be a bit longer possibly reflecting prose on both the because and effect of a particular actionwe also used evcaverber to investigate articles marked by the absence of members of each verb class such as articles lacking any verbs in the motion verb classhowever we found that absence of a verb class was not discriminatoryevaluation of evca verb classesto strengthen the observations that articles dominated by verbs of one class reflect distinct article types we verified that the verb classes behaved independently of each othercorrelations for evca classes are shown in table 6these show a markedly lower level of correlation between verb classes than the results for wordnet synsets the range being from 265 between motion and aspectual verbs to 026 for motion verbs and agreement verbsthese low values of 7 for pairs of verb classes reflects the independence of the classesfor example the communication and experience verb classes are weakly correlated this we surmise may be due to the different ways opinions can be expressed ie as factual quotes using communication class verbs or as beliefs using experience class verbssonthis paper reports results from two approaches one using wordnet and other based on evca classeshowever the basis for comparison must be made explicitin the case of wordnet all verb tokens were considered in all senses whereas in the case of evca a subset of less ambiguous verbs were manually selectedas reported above we covered 56 of the verbs by tokenindeed when we attempted to add more verbs to evca categories at the 59 mark we reached a point of difficulty in adding new verbs due to ambiguity eg verbs such as getthus although our results using evca are revealing in important ways it must be emphasized that the comparison has some imbalance which puts wordnet in an unnaturally negative lightin order to accurately compare the two approaches we would need to process either the same less ambiguous verb subset with wordnet or the full set of all verbs in all senses with evcaalthough the results reported in this paper permitted the validation of our hypothesis unless a fair comparison between resources is performed conclusions about wordnet as a resource versus evca class distinctions should not be inferredverb patternsin addition to considering verb type frequencies in texts we have observed that verb distribution and patterns might also reveal subtle information in textverb class distribution within the document and within particular subsections also carry meaningfor example we have observed that when sentences with movement verbs such as rise or fall are followed by sentences with because and then a telic aspectual verb such as reach this indicates that a value rose to a certain point due to the actions of some entityidentification of such sequences will enable us to assign functions to particular sections of contiguous text in an article in much the same way that text segmentation program seeks identify topics from distributional vocabulary we can also use specific sequences of verbs to help in determining methods for performing semantic aggregation of individual clauses in text generation for summarizationfuture workour plans are to extend the current research in terms of verb coverage and in terms of article coveragefor verbs we plan to increase the verbs that we cover to include phrasal verbs increase coverage of verbs by categorizing additional high frequency verbs into evca classes examine the effects of increased coverage on determining article typefor articles we plan to explore a general parser so we can test our hypothesis on additional texts and examine how our conclusions scale upfinally we would like to combine our techniques with other indicators to form a more robust system such as that envisioned in biber or suggested in kessler et al conclusionwe have outlined a novel approach to document analysis for news articles which permits discrimination of the event profile of news articlesthe goal of this research is to determine the role of verbs in document analysis keeping in mind that event profile is one of many factors in determining text typeour results show that levin evca verb classes provide reliable indicators of article type within the news domainwe have applied the algorithm to wsj data and have discriminated articles with five evca semantic classes into categories such as features opinions and announcementsthis approach to document type classification using verbs has not been explored previously in the literatureour results on verb analysis coupled with what is already known about np identification convinces us that future combinations of information will be even more successful in categorization of documentsresults such as these are useful in applications such as passage retrieval summarization and information extraction	P98-1112	role of verbs in document analysiswe present results of two methods for assessing the event profile of news articles as a function of verb typethe unique contribution of this research is the focus on the role of verbs rather than nounstwo algorithms are presented and evaluated one of which is shown to accurately discriminate documents by type and semantic properties ie the event profilethe initial method using wordnet produced multiple crossclassification of articles primarily due to the bushy nature of the verb tree coupled with the sense disambiguation problem our second approach using english verb classes and alternations levin showed that monosemous categorization of the frequent verbs in wsj made it possible to usefully discriminate documents for example our results show that articles in which communication verbs predominate tend to be opinion pieces whereas articles with a high percentage of agreement verbs tend to be about mergers or legal casesan evaluation is performed on the results using kendall tauwe present convincing evidence for using verb semantic classes as a discriminant in document classificationwe demonstrate that document type is correlated with the presence of many verbs of a certain evca class
automatic retrieval and clustering of similar words bootstrapping semantics from text is one of the greatest challenges in natural language learning we first define a word similarity measure based on the distributional pattern of words the similarity measure allows us to construct a thesaurus using a parsed corpus we then present a new evaluation methodology for the automatically constructed thesaurus the evaluation results show that the thesaurus is significantly closer to wordnet than roget thesaurus is the meaning of an unknown word can often be inferred from its contextconsider the following example in everyone likes tezgiiinotezgiiino makes you drunkwe make tezgiiino out of cornthe contexts in which the word tezgiiino is used suggest that tezgiiino may be a kind of alcoholic beverage made from corn mashbootstrapping semantics from text is one of the greatest challenges in natural language learningit has been argued that similarity plays an important role in word acquisition identifying similar words is an initial step in learning the definition of a wordthis paper presents a method for making this first stepfor example given a corpus that includes the sentences in our goal is to be able to infer that tezgiiino is similar to quotbeerquot quotwinequot quotvodkaquot etcin addition to the longterm goal of bootstrapping semantics from text automatic identification of similar words has many immediate applicationsthe most obvious one is thesaurus constructionan automatically created thesaurus offers many advantages over manually constructed thesaurifirstly the terms can be corpus or genrespecificmanually constructed generalpurpose dictionaries and thesauri include many usages that are very infrequent in a particular corpus or genre of documentsfor example one of the 8 senses of quotcompanyquot in wordnet 15 is a quotvisitorvisitantquot which is a hyponym of quotpersonquotthis usage of the word is practically never used in newspaper articleshowever its existance may prevent a coreference recognizer to rule out the possiblity for personal pronouns to refer to quotcompanyquotsecondly certain word usages may be particular to a period of time which are unlikely to be captured by manually compiled lexiconsfor example among 274 occurrences of the word quotwesternerquot in a 45 million word san jose mercury corpus 55 of them refer to hostagesif one needs to search hostagerelated articles quotwesternerquot may well be a good search termanother application of automatically extracted similar words is to help solve the problem of data sparseness in statistical natural language processing when the frequency of a word does not warrant reliable maximum likelihood estimation its probability can be computed as a weighted sum of the probabilities of words that are similar to itit was shown in that a similaritybased smoothing method achieved much better results than backoff smoothing methods in word sense disambiguationthe remainder of the paper is organized as followsthe next section is concerned with similarities between words based on their distributional patternsthe similarity measure can then be used to create a thesaurusin section 3 we evaluate the constructed thesauri by computing the similarity between their entries and entries in manually created thesaurisection 4 briefly discuss future work in clustering similar wordsfinally section 5 reviews related work and summarize our contributionsour similarity measure is based on a proposal in where the similarity between two objects is defined to be the amount of information contained in the commonality between the objects divided by the amount of information in the descriptions of the objectswe use a broadcoverage parser to extract dependency triples from the text corpusa dependency triple consists of two words and the grammatical relationship between them in the input sentencefor example the triples extracted from the sentence quoti have a brown dogquot are we use the notation 11w r w ii to denote the frequency count of the dependency triple in the parsed corpuswhen w r or w is the wild card the frequency counts of all the dependency triples that matches the rest of the pattern are summed upfor example kook obj 11 is the total occurrences of cookobject relationships in the parsed corpus and 1111 is the total number of dependency triples extracted from the parsed corpusthe description of a word w consists of the frequency counts of all the dependency triples that matches the pattern the commonality between two words consists of the dependency triples that appear in the descriptions of both wordsfor example is the the description of the word quotcellquotii cell objof contain114 objof decorate112 icell nmod bacteriall3 i cell nmod blood vessell11 i cell nmod body112 icell nmod bone marrow 112 icell nmod burial 111 j cell nmod chameleon111 assuming that the frequency counts of the dependency triples are independent of each other the information contained in the description of a word is the sum of the information contained in each individual frequency countto measure the information contained in the statement ilw r ilc we first measure the amount of information in the statement that a randomly selected dependency triple is when we do not know the value of 11w r willwe then measure the amount of information in the same statement when we do know the value of11w r w iithe difference between these two amounts is taken to be the information contained in 11w r w ii c an occurrence of a dependency triple can be regarded as the cooccurrence of three events a a randomly selected word is w b a randomly selected dependency type is r c a randomly selected word is w1when the value of ii w r w ii is unknown we assume that a and c are conditionally independent given bthe probability of a b and c cooccurring is estimated by where pale is the maximum likelihood estimation of a probability distribution and ii cell obj of bludgeon111 licell objof cal11111 cell objof come from113 when the value of ii w r w ii is known we can obtain pale directly let denote the amount information contained in 11w r w1 iicits value can be cornit is worth noting that 1 is equal to the mutual information between w and w let t be the set of pairs such that log x 1ris positivewe define the similarity sim between two words w1 and w2 as follows we parsed a 64millionword corpus consisting of the wall street journal san jose mercury and ap newswire from the parsed corpus we extracted 565 million dependency triples in the parsed corpus there are 5469 nouns 2173 verbs and 2632 adjectivesadverbs that occurred at least 100 timeswe computed the pairwise similarity between all the nouns all the verbs and all the adjectivesadverbs using the above similarity measurefor each word we created a thesaurus entry which contains the topn1 words that are most similar to it2 the thesaurus entry for word w has the following format where pos is a part of speech we is a word sesim and se are ordered in descending orderfor example the top10 words in the noun verb and adjective entries for the word quotbriefquot are shown below brief affidavit 013 petition 005 memorandum 005 motion 005 lawsuit 005 deposition 005 slight 005 prospectus 004 document 004 paper 004 brief tell 009 urge 007 ask 007 meet 006 appoint 006 elect 005 name 005 empower 005 summon 005 overrule 004 brief lengthy 013 short 012 recent 009 prolonged 009 long 009 extended 009 daylong 008 scheduled 008 stormy 007 planned 006 two words are a pair of respective nearest neighbors if each is the other most similar wordour program found 543 pairs of rnn nouns 212 pairs of rnn verbs and 382 pairs of rnn adjectivesadverbs in the automatically created thesaurusappendix a lists every 10th of the rnnsthe result looks very strongfew pairs of rnns in appendix a have clearly better alternativeswe also constructed several other thesauri using the same corpus but with the similarity measures in figure 1the measure sirrihindie is the same as the similarity measure proposed in except that it does not use dependency triples with negative mutual informationthe measure simhindi is the same as sim h indle except that all types of dependency relationships are used instead of just subject and object relationshipsthe measures simeosine simdice and simjacard are versions of similarity measures commonly used in information retrieval unlike sim sim h indle and si mhindler they only where s is the set of senses of w in the wordnet super is the set of superclasses of concept c in the wordnet r is the set of words that belong to a same roget category as w make use of the unique dependency triples and ignore their frequency countsin this section we present an evaluation of automatically constructed thesauri with two manually compiled thesauri namely wordnet15 and roget thesauruswe first define two word similarity measures that are based on the structures of wordnet and roget the similarity measure simwn is based on the proposal in the similarity measure sim roget treats all the words in roget as featuresa word w possesses the feature f if f and w belong to a same roget categorythe similarity between two words is then defined as the cosine coefficient of the two feature vectorswith simwn and simroget we transform wordnet and roget into the same format as the automatically constructed thesauri in the previous sectionwe now discuss how to measure the similarity between two thesaurus entriessuppose two thesaurus entries for the same word are as follows for example is the entry for quotbrief quot in our automatically generated thesaurus and and are corresponding entries in wordnet thesaurus and roget thesaurus brief affidavit 013 petition 005 memorandum 005 motion 005 lawsuit 005 deposition 005 slight 005 prospectus 004 document 004 paper 004our evaluation was conducted with 4294 nouns that occurred at least 100 times in the parsed corpus and are found in both wordnet15 and the roget thesaurustable 1 shows the average similarity between corresponding entries in different thesauri and the standard deviation of the average which is the standard deviation of the data items divided by the square root of the number of data itemssince the differences among sim simjacard cosine simdzce and are very small we only included the results for sitncosi in table 1 for the sake of brevityit can be seen that sim hindle and cosine are significantly more similar to wordnet than roget is but are significantly less similar to roget than wordnet isthe differences between hindle and hindle clearly demonstrate that the use of other types of dependencies in addition to subject and object relationships is very beneficialthe performance of sim hindle and cosine are quite closeto determine whether or not the differences are statistically significant we computed their differences in similarities to wordnet and roget thesaurus for each individual entrytable 2 shows the average and standard deviation of the average differencesince the 95 confidence intervals of all the differences in table 2 are on the positive side one can draw the statistical conclusion that simis better than simhzndle which is better than sinacosinewordnet average gra vp sim hindle 0008021 0000428 simcosine 0012798 0000386 hindle cosine 0004777 0000561 roget average cravg simhindle 0002415 0000401 sim cosine 0013349 0000375 hindle cosine 0010933 0000509reliable extraction of similar words from text corpus opens up many possibilities for future workfor example one can go a step further by constructing a tree structure among the most similar words so that different senses of a given word can be identified with different subtreeslet w1 wr be a list of words in descending order of their similarity to a given word w the similarity tree for w is created as follows among lw w1 wz11for example figure 3 shows the similarity tree for the top40 most similar words to dutythe first number behind a word is the similarity of the word to its parentthe second number is the similarity of the word to the root node of the treeinspection of sample outputs shows that this algorithm works wellhowever formal evaluation of its accuracy remains to be future workthere have been many approaches to automatic detection of similar words from text corporaours is similar to in the use of dependency relationship as the word features based on which word similarities are computedevaluation of automatically generated lexical resources is a difficult problemin a small set of sample results are presentedin automatically extracted collocations are judged by a lexicographerin and clusters of similar words are evaluated by how well they are able to recover data items that are removed from the input corpus one at a timein the collocations and their associated scores were evaluated indirectly by their use in parse tree selectionthe merits of different measures for association strength are judged by the differences they make in the precision and the recall of the parser outputsthe main contribution of this paper is a new evaluation methodology for automatically constructed thesauruswhile previous methods rely on indirect tasks or subjective judgments our method allows direct and objective comparison between automatically and manually constructed thesaurithe results show that our automatically created thesaurus is significantly closer to wordnet than roget thesaurus isour experiments also surpasses previous experiments on automatic thesaurus construction in scale and accuracythis research has also been partially supported by nserc research grant 0gp121338 and by the institute for robotics and intelligent systems	P98-2127	automatic retrieval and clustering of similar wordsbootstrapping semantics from text is one of the greatest challenges in natural language learningwe first define a word similarity measure based on the distributional pattern of wordsthe similarity measure allows us to construct a thesaurus using a parsed corpuswe then present a new evaluation methodology for the automatically constructed thesaurusthe evaluation results show that the thesaurns is significantly closer to wordnet than roget thesaurus iswe use dependency relation as word features to compute word similarities from large corpora
robust pronoun resolution with limited knowledge most traditional approaches to anaphora resolution rely heavily on linguistic and domain knowledge one of the disadvantages of developing a knowledgebased system however is that it is a very labourintensive and timeconsuming task this paper presents a robust knowledgepoor approach to resolving pronouns in technical manuals which operates on texts preprocessed by a partofspeech tagger input is checked against agreement and for a number of antecedent indicators candidates are assigned scores by each indicator and the candidate with the highest score is returned as the antecedent evaluation reports a success rate of 897 which is better than the success rates of the approaches selected for comparison and tested on the same data in addition preliminary experiments show that the approach can be successfully adapted for other languages with minimum modifications for the most part anaphora resolution has focused on traditional linguistic methods however to represent and manipulate the various types of linguistic and domain knowledge involved requires considerable human input and computational expensewhile various alternatives have been proposed making use of eg neural networks a situation semantics framework or the principles of reasoning with uncertainty there is still a strong need for the development of robust and effective strategies to meet the demands of practical nlp systems and to enhance further the automatic processing of growing language resourcesseveral proposals have already addressed the anaphora resolution problem by deliberately limiting the extent to which they rely on domain andor linguistic knowledge our work is a continuation of these latest trends in the search for inexpensive fast and reliable procedures for anaphora resolutionit is also an example of how anaphors in a specific genre can be resolved quite successfully without any sophisticated linguistic knowledge or even without parsingfinally our evaluation shows that the basic set of antecedent tracking indicators can work well not only for english but also for other languages with a view to avoiding complex syntactic semantic and discourse analysis we developed a robust knowledgepoor approach to pronoun resolution which does not parse and analyse the input in order to identify antecedents of anaphorsit makes use of only a partofspeech tagger plus simple noun phrase rules and operates on the basis of antecedenttracking preferences the approach works as follows it takes as an input the output of a text processed by a partofspeech tagger identifies the noun phrases which precede the anaphor within a distance of 2 sentences checks them for gender and number agreement with the anaphor and then applies the genrespecific antecedent indicators to the remaining candidates the noun phrase with the highest aggregate score is proposed as antecedent in the rare event of a tie priority is given to the candidate with the higher score for immediate referenceif immediate reference has not been identified then priority is given to the candidate with the best collocation pattern scoreif this does not help the candidate with the higher score for indicating verbs is preferredif still no choice is possible the most recent from the remaining candidates is selected as the antecedentantecedent indicators play a decisive role in tracking down the antecedent from a set of possible candidatescandidates are assigned a score for each indicator the candidate with the highest aggregate score is proposed as the antecedentthe antecedent indicators have been identified empirically and are related to salience to structural matches to referential distance or to preference of termswhilst some of the indicators are more genrespecific and others are less genrespecific the majority appear to be genreindependentin the following we shall outline some the indicators used and shall illustrate them by examplesdefinite noun phrases in previous sentences are more likely antecedents of pronominal anaphors than indefinite ones we regard a noun phrase as definite if the head noun is modified by a definite article or by demonstrative or possessive pronounsthis rule is ignored if there are no definite articles possessive or demonstrative pronouns in the paragraph noun phrases in previous sentences representing the quotgiven informationquot 1 are deemed good candidates for antecedents and score 1 in a coherent text the given or known information or theme usually appears first and thus forms a coreferential link with the preceding textthe new information or rheme provides some information about the themewe use the simple heuristics that the given information is the first noun phrase in a nonimperative sentenceindicating verbs if a verb is a member of the verb_set discuss present illustrate identify summarise examine describe define show check develop review report outline consider investigate explore assess analyse synthesise study survey deal cover we consider the first np following it as the preferred antecedent empirical evidence suggests that because of the salience of the noun phrases which follow them the verbs listed above are particularly good indicatorslexically reiterated items are likely candidates for antecedent lexically reiterated items include repeated synonymous noun phrases which may often be preceded by definite articles or demonstrativesalso a sequence of noun phrases with the same head counts as lexical reiteration section heading preference if a noun phrase occurs in the heading of the section part of which is the current sentence then we consider it as the preferred candidate quotnonprepositionalquot noun phrases a quotpurequot quotnonprepositionalquot noun phrase is given a higher preference than a noun phrase which is part of a prepositional phrase example insert the cassettei into the vcr making sure iti is suitable for the length of recordinghere quotthe vcrquot is penalised for being part of the prepositional phrase quotinto the vcrquotthis preference can be explained in terms of salience from the point of view of the centering theorythe latter proposes the ranking quotsubject direct object indirect objectquot and noun phrases which are parts of prepositional phrases are usually indirect objectsthis preference is given to candidates which have an identical collocation pattern with a pronoun the collocation preference here is restricted to the patterns quotnoun phrase verbquot and quotverb noun phrase quotowing to lack of syntactic information this preference is somewhat weaker than the collocation preference described in example press the keyi down and turn the volume up press iti againimmediate reference in technical manuals the quotimmediate referencequot clue can often be useful in identifying the antecedentthe heuristics used is that in constructions of the form quot v1 np con v2 it v3 itquot where con e andorbeforeafter the noun phrase immediately after v1 is a very likely candidate for antecedent of the pronoun quotitquot immediately following v2 and is therefore given preference this preference can be viewed as a modification of the collocation preferenceit is also quite frequent with imperative constructionsexample to print the paper you can stand the printeri up or lay iti flatto turn on the printer press the power button i and hold iti down for a momentunwrap the paperi form iti and align iti then load iti into the drawerreferential distance in complex sentences noun phrases in the previous clause2 are the best candidate for the antecedent of an anaphor in the subsequent clause followed by noun phrases in the previous sentence then by nouns situated 2 sentences further back and finally nouns 3 sentences further back for anaphors in simple sentences noun phrases in the previous sentence are the best candidate for antecedent followed by noun phrases situated 2 sentences further back and finally nouns 3 sentences further back term preference nps representing terms in the field are more likely to be the antecedent than nps which are not terms as already mentioned each of the antecedent indicators assigns a score with a value these scores have been determined experimentally on an empirical basis and are constantly being updatedtop symptoms like quotlexical reiterationquot assign score quot2quot whereas quotnonprepositionalquot noun phrases are given a negative score of quot1quotwe should point out that the antecedent indicators are preferences and not absolute factorsthere might be cases where one or more of the antecedent indicators do not quotpointquot to the correct antecedentfor instance in the sentence quotinsert the cassette into the vcr i making sure iti is turned onquot the indicator quotnonprepositional noun phrasesquot would penalise the correct antecedentwhen all preferences are taken into account however the right antecedent is still very likely to be tracked down in the above example the quotnonprepositional noun phrasesquot heuristics would be overturned by the quotcollocational preferencequot heuristicsthe algorithm for pronoun resolution can be described informally as follows 3a sentence splitter would already have segmented the text into sentences a pos tagger would already have determined the parts of speech and a simple phrasal grammar would already have detected the noun phrases 4in this project we do not treat cataphora nonanaphoric quotitquot occurring in constructions such as quotit is importantquot quotit is necessaryquot is eliminated by a quotreferential filterquot 5note that this restriction may not always apply in languages other than english on the other hand there are certain collective nouns in english which do not agree in number with their antecedents can be referred to by quotitquot and are exempted from the agreement testfor this purpose we have drawn up a comprehensive list of all such cases to our knowledge no other computational treatment of pronominal anaphora resolution has addressed the problem of quotagreement exceptionsquot antecedentif two candidates have an equal in order to evaluate the effectiveness of the apscore the candidate with the higher score for proach and to explore if how far it is superior over immediate reference is proposed as antecedent the baseline models for anaphora resolution we also if immediate reference does not hold propose tested the sample text on a baseline model which the candidate with higher score for collocational checks agreement in number and gender and where patternif collocational pattern suggests a tie or more than one candidate remains picks as antecedoes not hold select the candidate with higher dent the most recent subject matching the gender score for indicating verbsif this indicator does and number of the anaphor a baseline model not hold again go for the most recent candidate which picks as antecedent the most recent noun 3evaluation phrase that matches the gender and number of the for practical reasons the approach presented does anaphor not incorporate syntactic and semantic information the success rate of the quotbaseline subjectquot was and it is not real 292 whereas the success rate of quotbaseline most istic to expect its performance to be as good as an recent npquot was 625given that our knowledgeapproach which makes use of syntactic and semantic poor approach is basically an enhancement of a knowledge in terms of constraints and preferences baseline model through a set of antecedent indicathe lack of syntactic information for instance tors we see a dramatic improvement in performance means giving up ccommand constraints and subject when these preferences are called upon preference which could be used in center superior to both baseline models when the antecetrackingsyntactic parallelism useful in discrimi dent was neither the most recent subject nor the nating between identical pronouns on the basis of most recent noun phrase matching the anaphor in their syntactic function also has to be forgonelack gender and numberexample of semantic knowledge rules out the use of verb se identify the draweri by the lit paper port led and mantics and semantic parallelismour evaluation add paper to iti however suggests that much less is lost than might the aggregate score for quotthe drawerquot is 7 be fearedin fact our evaluation shows that the re we believe that the good heading 0 collocation 0 referential distance 2 success rate is due to the fact that a number of ante nonprepositional noun phrase 0 immediate refercedent indicators are taken into account and no fac ence 2 7 whereas aggregate score for the most tor is given absolute preferencein particular this recent matching noun phrase out of 223 pronouns in the text 167 were nonanaphoric the evaluation carried out was manual to ensure that no added error was generated another reason for doing it by hand is to ensure a fair comparison with breck baldwin method which not being available to us had to be handsimulated the evaluation indicated 836 success ratethe quotbaseline subjectquot model tested on the same data scored 339 recall and 679 precision whereas quotbaseline most recentquot scored 667note that quotbaseline subjectquot can be assessed both in terms of recall and precision because this quotversionquot is not robust in the event of no subject being available it is not able to propose an antecedent in the second experiment we evaluated the approach from the point of view also of its quotcritical success ratequotthis measure applies only to anaphors quotambiguousquot from the point of view of number and gender and is indicative of the performance of the antecedent indicatorsour evaluation established the critical success rate as 82a case where the system failed was when the anaphor and the antecedent were in the same sentence and where preference was given to a candidate in the preceding sentencethis case and other cases suggest that it might be worthwhile reconsideringrefining the weights for the indicator quotreferential distancequotsimilarly to the first evaluation we found that the robust approach was not very successful on sentences with too complicated syntax a price we have to pay for the quotconveniencequot of developing a knowledgepoor systemthe results from experiment 1 and experiment 2 can be summarised in the following slightly more representative figuresthe lower figure in quotbaseline subjectquot corresponds to quotrecallquot and the higher figure to quotprecisionquotif we regard as quotdiscriminative powerquot of each antecedent indicator the ratio quotnumber of successful antecedent identifications when this indicator was appliedquotquotnumber of applications of this indicatorquot the immediate reference emerges as the most discriminative indicator followed by nonprepositional noun phrase collocation section heading lexical reiteration givenness term preference and referential distance the relatively low figures for the majority of indicators should not be regarded as a surprise firstly we should bear in mind that in most cases a candidate was picked as an antecedent on the basis of applying a number of different indicators and secondly that most anaphors had a relatively high number of candidates for antecedentin terms of frequency of use the most frequently used indicator proved to be referential distance used in 989 of the cases followed by term preference givenness lexical reiteration definiteness section heading immediate reference and collocation as expected the most frequent indicators were not the most discriminative oneswe felt appropriate to extend the evaluation of our approach by comparing it to breck baldwin cogmac approach which features quothigh precision coreference with limited knowledge and linguistics resourcesquotthe reason is that both our approach and breck baldwin approach share common principles and therefore a comparison would be appropriategiven that our approach is robust and returns antecedent for each pronoun in order to make the comparison as fair as possible we used cogniac quotresolve allquot version by simulating it manually on the same training data used in evaluation b abovecogniac successfully resolved the pronouns in 75 of the casesthis result is comparable with the results described in for the training data from the genre of technical manuals it was rule 5 which was most frequently used followed by rule 8 rule 7 rule 1 and rule 3 it would be fair to say that even though the results show superiority of our approach on the training data used they cannot be generalised automatically for other genres or unrestricted texts and for a more accurate picture further extensive tests are necessaryan attractive feature of any nlp approach would be its language quotuniversalityquotwhile we acknowledge that most of the monolingual nlp approaches are not automatically transferable to other languages it would be highly desirable if this could be done with minimal adaptationwe used the robust approach as a basis for developing a genrespecific reference resolution approach in polishas expected some of the preferences had to be modified in order to fit with specific features of polish for the time being we are using the same scores for polishthe evaluation for polish was based technical manuals available on the internet the sample texts contained 180 pronouns among which were 120 instances of exophoric reference the robust approach adapted for polish demonstrated a high success rate of 933 in resolving anaphors similarly to the evaluation for english we compared the approach for polish with a baseline model which discounts candidates on the basis of agreement in number and gender and if there were still competing candidates selects as the antecedent the most recent subject matching the anaphor in gender and number a baseline model which checks agreement in number and gender and if there were still more than one candidate left picks up as the antecedent the most recent noun phrase that agrees with the anaphorour preferencebased approach showed clear superiority over both baseline modelsthe first baseline model was successful in only 237 of the cases whereas the second had a success rate of 684therefore the 933 success rate demonstrates a dramatic increase in precision which is due to the use of antecedent tracking preferenceswe have recently adapted the approach for arabic as well our evaluation based on 63 examples from a technical manual indicates a success rate of 952 we have described a robust knowledgepoor approach to pronoun resolution which operates on texts preprocessed by a partofspeech taggerevaluation shows a success rate of 897 for the genre of technical manuals and at least in this genre the approach appears to be more successful than other similar methodswe have also adapted and evaluated the approach for polish and for arabic	P98-2143	robust pronoun resolution with limited knowledgemost traditional approaches to anaphora resolution rely heavily on linguistic and domain knowledgeone of the disadvantages of developing a knowledgebased system however is that it is a very labourintensive and timeconsuming taskthis paper presents a robust knowledgepoor approach to resolving pronouns in technical manuals which operates on texts preprocessed by a partofspeech taggerinput is checked against agreement and for a number of antecedent indicatorscandidates are assigned scores by each indicator and the candidate with the highest score is returned as the antecedentevaluation reports a success rate of 897 which is better than the success rates of the approaches selected for comparison and tested on the same datain addition preliminary experiments show that the approach can be successfully adapted for other languages with minimum modificationswe first apply a set of constraints to filter grammatically incompatible candidate antecedents and then rank the remaining ones using salience factorswe find that the current evaluation of anaphora resolution algorithms and systems is befeft of any common ground for comparison due to the difference in evaluation data as well as the diversity of preprocessing tools employed by each anaphora resolution system
multilingual authoring using feedback texts there are obvious reasons for trying to automate the production of multilingual documentation especially for routine subjectmatter in restricted domains two approaches have been adopted machine translation of a source text and multilingual natural language generation from a knowledge base for mt information extraction is a major difficulty since the meaning must be derived by analysis of the source text mnlg avoids this difficulty but seems at first sight to require an expensive phase of knowledge engineering in order to encode the meaning we introduce here a new technique which employs mnlg during the phase of knowledge editing a feedback text generated from a possibly incomplete knowledge base describes in natural language the knowledge encoded so far and the options for extending it this method allows anyone speaking one of the supported languages to produce texts in all of them requiring from the author only expertise in the subjectmatter not expertise in knowledge engineering the production of multilingual documentation has an obvious practical importancecompanies seeking global markets for their products must provide instructions or other reference materials in a variety of languageslarge political organizations like the european union are under pressure to provide multilingual versions of official documents especially when communicating with the publicthis need is met mostly by human translation an author produces a source document which is passed to a number of other people for translation into other languageshuman translation has several wellknown disadvantagesit is not only costly but timeconsuming often delaying the release of the product in some markets also the quality is uneven and hard to control for all these reasons the production of multilingual documentation is an obvious candidate for automation at least for some classes of documentnobody expects that automation will be applied in the foreseeable future for literary texts ranging over wide domains however there is a mass of nonliterary material in restricted domains for which automation is already a realistic aim instructions for using equipment are a good examplethe most direct attempt to automize multilingual document production is to replace the human translator by a machinethe source is still a natural language document written by a human author a program takes this source as input and produces an equivalent text in another language as outputmachine translation has proved useful as a way of conveying roughly the information expressed by the source but the output texts are typically poor and overliteralthe basic problem lies in the analysis phase the program cannot extract from the source all the information that it needs in order to produce a good output textthis may happen either because the source is itself poor or because the source uses constructions and concepts that lie outside the program rangesuch problems can be alleviated to some extent by constraining the source document eg through use of a controlled language such as aecma an alternative approach to translation is that of generating the multilingual documents from a nonlinguistic sourcein the case of automatic multilingual natural language generation the source will be a knowledge base expressed in a formal languageby eliminating the analysis phase of mt mnlg can yield highquality output texts free from the literal quality that so often arises from structural imitation of an input textunfortunately this benefit is gained at the cost of a huge increase in the difficulty of obtaining the sourceno longer can the domain expert author the document directly by writing a text in natural languagedefining the source becomes a task akin to building an expert system requiring collaboration between a domain expert and a knowledge engineer owing to this cost mnlg has been applied mainly in contexts where the knowledge base is already available having been created for another purpose for discussion see reiter and mellish is there any way in which a domain expert might author a knowledge base without going through this timeconsuming and costly collaboration with a knowledge engineerassuming that some kind of mediation is needed between domain expert and knowledge formalism the only alternative is to provide easier tools for editing knowledge basessome knowledge management projects have experimented with graphical presentations which allow editing by direct manipulation so that there is no need to learn the syntax of a programming language see for example skuce and lethbridge this approach has also been adopted in two mnlg systems gist which generates social security forms in english italian and german and drafter which generates instructions for software applications in english and frenchthese projects were the first attempts to produce symbolic authoring systems that is systems allowing a domain expert with no training in knowledge engineering to author a knowledge base from which texts in many languages can be generatedalthough helpful graphical tools for managing knowledge bases remain at best a compromise solutiondiagrams may be easier to understand than logical formalisms but they still lack the flexibility and familiarity of natural ianguage text as empirical studies on editing diagrammatic representations have shown for discussion see power et al this observation has led us to explore a new possibility at first sight paradoxical that of a symbolic authoring system in which the current knowledge base is presented through a natural language text generated by the systemthis kills two birds with one stone the source is still a knowledge base not a text so no problem of analysis arises but this source is presented to the author in natural language through what we will call a feedback textas we shall see the feedback text has some special features which allow the author to edit the knowledge base as well as viewing its contentswe have called this editing method wysiwym or what you see is what you meant a natural language text presents a knowledge base that the author has built by purely semantic decisions feedback texts to the authorthe feedback texts will include mousesensitive anchors allowing the author to make semantic decisions eg by selecting options from popup menusthe wysiwym system allows a domain expert speaking any one of the supported languages to produce good output texts in all of thema more detailed description of the architecture is given in scott et al 2 example of a wysiwym system the first application of wysiwym was drafterii a system which generates instuctions for using word processors and diary managersat present three languages are supported english french and italianas an example we will follow a session in which the author encodes instructions for scheduling an appointment with the openwindows calendar managerthe desired content is shown by the following output text which the system will generate when the knowledge base is complete to schedule the appointment before starting open the appointment editor window by choosing the appointment option from the edit menuthen proceed as follows in outline the knowledge base underlying this text is as followsthe whole instruction is represented by a procedure instance with two attributes a goal and a methodthe method instance also has two attributes a precondition and a sequence of steps preconditions and steps are procedures in their turn so they may have methods as well as goalseventually we arrive at subprocedures for which no method is specified it is assumed that the reader of the manual will be able to click on the insert button without being told howsince in drafterh every output text is based on a procedure a newly initialised knowledge base is seeded with a single procedure instance for which the goal and method are undefinedin prolog notation we can represent such a knowledge base by the following assertions procedure goal method here prod l is an identifier for the procedure instance the assertion procedure means that this is an instance of type procedure and the assertion goal means that prod has a goal attribute for which the value is currently undefined when a new knowledge base is created drafterii presents it to the author by generating a feedback text in the currently selected languageassuming that this language is english the instruction to the generator will be generate and the feedback text displayed to the author will be achieve this goal by applying this methodthis text has several special features ing on an anchor the author obtains a popup menu listing the permissible values of the attribute by selecting one of these options the author updates the knowledge basealthough the anchors may be tackled in any order we will assume that the author proceeds from left to rightclicking on this goal yields the popup menu choose click close create save schedule start from which the author selects cheduleeach option in the menu is associated with an updater a prolog term that specifies how the knowledge base should be updated if the option is selectedin this case the updater is insert meaning that an instance of type schedule should become the value of the goal attribute on prodrunning the updater yields an extended knowledge base including a new instance schedl with an undefined attribute actee procedure goal schedule actee method from the updated knowledge base the generator produces a new feedback textschedule this event by applying this methodnote that this text has been completely regeneratedit was not produced from the previous text merely by replacing the anchor this goal by a longer stringcontinuing to specify the goal the author now clicks on this event appointment meeting this time the intended selection is appointment but let us assume that by mistake the author drags the mouse too far and selects meetingthe feedback text schedule the meeting by applying this method immediately shows that an error has been made but how can it be correctedthis problem is solved in wysiwym by allowing the author to select any span of the feedback text that represents an attribute with a specified value and to cut it so that the attribute becomes undefined while its previous value is held in a buffereven large spans representing complex attribute values can be treated in this way so that complex chunks of knowledge can be copied across from one knowledge base to anotherwhen the author selects the phrase the meeting the system displays a popup menu with two options cut copy by selecting cut the author activates the updater cut which updates the knowledge base by removing the instance meetl currently the value of the actee attribute on schedl and holding it in a bufferwith this attribute now undefined the feedback text reverts to schedule this event by applying this method whereupon the author can once again expand this eventthis time however the popup menu that opens on this anchor will include an extra option that of pasting back the material that has just been cutof course this option is only provided if the instance currently held in the buffer is a suitable value for the attribute represented by the anchorpaste appointment meeting the paste option here will be associated with the updater paste which would assign the instance currently in the buffer in this case meet1 as the value of the actee attribute on sched1fortunately the author avoids reinstating this error and selects appointment yielding the following reassuring feedback text schedule the appointment by applying this methodnote incidentally that this text presents a knowledge base that is potentially complete since all obligatory attributes have been specifiedthis can be immediately seen from the absence of any red anchorsintending to add a method the author now clicks on this methodin this case the popup menu shows only one option method running the associated updater yields the following knowledge base procedure goal schedule actee appointment method method precondition steps steps first procedure goal method rest meeting a considerable expansion has taken place here because the system has been configured to automatically instantiate obligatory attributes that have only one permissible type of valuesince the steps attribute on method1 is obligatory and must have a value of type steps the instance steps1 is immediately createdin its turn this instance has the attributes first and rest where first is obligatory and must be filled by a procedurea second procedure instance proc2 is therefore created with its own goal and methodto incorporate all this new material the feedback text is recast in a new pattern the main goal being expressed by an infinitive construction instead of an imperative to schedule the appointment first achieve this preconditionthen follow these stepsnote that at any stage the author can switch to one of the other supported languages egfrenchthis will result in a new call to the generator and hence in a new feedback text expressing the procedure prodinsertion du rendezvous avant de commencer accomplir cette tcicheexecuter les actions suivantesclicking for example on cette action will now yield the usual options for instanciating a goal attribute but expressed in frenchthe associated updaters are identical to those for the corresponding menu in english choix cliquer fermer enregistrement insertion lancement the now be clear so let us advance to a later stage in which the scheduling procedure has been fully encodedto schedule the appointment first open the appointment editor windowthen follow these stepsto open the appointment editor window first achieve this preconditionthen follow these steps basic mechanism should two points about this feedback text are worth generate notingfirst to avoid overcrowding the main paragraph the text planner has deferred the subprocedure for opening the appointment editor window which is presented in a separate paragraphto maintain a connection the action of opening the appointment editor window is mentioned twice secondly no red anchors are left so the knowledge base is potentially completethis means that the author may now generate an output text by switching the modality from feedback to outputthe resulting instruction to the generator will be generate yielding the output text shown at the beginning of the sectionfurther output texts can be obtained by switching to another language egfrench insertion du rendezvous avant de commencer ouvrir la fenetre appointment editor en choisissant loption appointment dans le menu editexecuter les actions suivantes 1 choisir lheure de fin du rendezvous2 inserer la description du rendezvous dans la zone de texte what3 cliquer sur le bouton insertnote that in output modality the generator ignores optional undefined attributes the method for opening the appointment editor window thus reduces to a single action which can be reunited with its goal in the main paragraphwysiwym editing is a new idea that requires practical testingwe have not yet carried out formal usability trials nor investigated the design of feedback texts nor confirmed that adequate response times could be obtained for fullscale applicationshowever if satisfactory largescale implementations prove feasible the method brings many potential benefits a document in natural language is the most flexible existing medium for presenting informationwe cannot be sure that all meanings can be expressed clearly in network diagrams or other specialized presentations we can be sure they can be expressed in a document it seems intuitively obvious that authors will understand feedback texts much better than they understand alternative methods of presenting knowledge bases such as network diagramsour experience has been that people can learn to use the drafterii system in a few minutes authors require no training in a controlled language or any other presentational conventionthis avoids the expense of initial training it also means that presentational conventions need not be relearned when a knowledge base is reexamined after a delay of months or years since the knowledge base is presented through a document in natural language it becomes immediately accessible to anyone peripherally concerned with the project documentation of the knowledge base often a tedious and timeconsuming task becomes automatic the model can be viewed and edited in any natural language that is supported by the generator further languages can be added as neededwhen supported by a multilingual natural language generation system as in drafterii wysiwym editing obviates the need for traditional language localisation of the humancomputer interfacenew linguistic styles can also be added as a result wysiwym editing is ideal for facilitating knowledge sharing and transfer within a multilingual projectspeakers of several different languages could collectively edit the same knowledge base each user viewing and modifying the knowledge in hisher own language navigated by the methods familiar from books and from complex electronic documents obviating any need for special training in navigationthe crucial advantage of wysiwym editing compared with alternative natural language interfaces is that it eliminates all the usual problems associated with parsing and semantic interpretationfeedback texts with menus have been used before in the nlmenu system but only as a means of presenting syntactic optionsnlmenu guides the author by listing the extensions of the current sentence that are covered by its grammar in this way it makes parsing more reliable by enforcing adherence to a sublanguage but parsing and interpretation are still requiredso far wysiwym editing has been implemented in two domains software instructions and patient information leafletswe are currently evaluating the usability of these systems partly to confirm that authors do indeed find them easy to use and partly to investigate issues in the design of feedback texts	P98-2173	multilingual authoring using feedback textsthere are obvious reasons for trying to automate the production of multilingual documentation especially for routine subjectmatter in restricted domains two approaches have been adopted machine translation of a source text and multilingual natural language generation from a knowledge basefor mt information extraction is a major difficulty since the meaning must be derived by analysis of the source text mnlg avoids this difficulty but seems at first sight to require an expensive phase of knowledge engineering in order to encode the meaningwe introduce here a new technique which employs mnlg during the phase of knowledge editinga feedback text generated from a possibly incomplete knowledge base describes in natural language the knowledge encoded so far and the options for extending itthis method allows anyone speaking one of the supported languages to produce texts in all of them requiring from the author only expertise in the subjectmatter not expertise in knowledge engineeringwe propose wysiwym as a method for the authoring of semantic information through direct manipulation of structures rendered in natural language textin this system logical forms are entered interactively and the corresponding linguistic realization of the expressions is generated in several languages
statistical models for unsupervised prepositional phrase attachment several unsupervised statistical models for the prepositional phrase attachment task that approach the accuracy of the best supervised methods for this task our unsupervised approach uses a heuristic based on attachment proximity and trains from raw text that is annotated with only partofspeech tags and morphological base forms as opposed to attachment information it is therefore less resourceintensive and more portable than previous corpusbased algorithm proposed for this task we present results for prepositional phrase attachment in both english and spanish prepositional phrase attachment is the task of deciding for a given preposition in a sentence the attachment site that corresponds to the interpretation of the sentencefor example the task in the following examples is to decide whether the preposition with modifies the preceding noun phrase or the preceding verb phrase in sentence 1 with modifies the noun shirt since with pockets describes the shirthowever in sentence 2 with modifies the verb washed since with soap describes how the shirt is washedwhile this form of attachment ambiguity is usually easy for people to resolve a computer requires detailed knowledge about words in order to successfully resolve such ambiguities and predict the correct interpretationmost of the previous successful approaches to this problem have been statistical or corpusbased and they consider only prepositions whose attachment is ambiguous between a preceding noun phrase and verb phraseprevious work has framed the problem as a classification task in which the goal is to predict n or v corresponding to noun or verb attachment given the head verb v the head noun n the preposition p and optionally the object of the preposition n2for example the tuples corresponding to the example sentences are the correct classifications of tuples 1 and 2 are n and v respectively describes a partially supervised approach in which the fidditch partial parser was used to extract tuples from raw text where p is a preposition whose attachment is ambiguous between the head verb v and the head noun n the extracted tuples are then used to construct a classifier which resolves unseen ambiguities at around 80 accuracylater work such as trains and tests on quintuples of the form extracted from the penn treebank and has gradually improved on this accuracy with other kinds of statistical learning methods yielding up to 845 accuracyrecently have reported 88 accuracy by using a corpusbased model in conjunction with a semantic dictionarywhile previous corpusbased methods are highly accurate for this task they are difficult to port to other languages because they require resources that are expensive to construct or simply nonexistent in other languageswe present an unsupervised algorithm for prepositional phrase attachment in english that requires only an partofspeech tagger and a morphology database and is therefore less resourceintensive and more portable than previous approaches which have all required either treebanks or partial parsersthe exact task of our algorithm will be to construct a classifier c which maps an instance of an ambiguous prepositional phrase to either n or v corresponding to noun attachment or verb attachment respectivelyin the full natural language parsing task there are more than just two potential attachment sites but we limit our task to choosing between a verb v and a noun n so that we may compare with previous supervised attempts on this problemwhile we will be given the candidate attachment sites during testing the training procedure assumes no a priori information about potential attachment siteswe generate training data from raw text by using a partofspeech tagger a simple chunker an extraction heuristic and a morphology databasethe order in which these tools are applied to raw text is shown in table 1the tagger from first annotates sentences of raw text with a sequence of partofspeech tagsthe chunker implemented with two small regular expressions then replaces simple noun phrases and quantifier phrases with their head wordsthe extraction heuristic then finds head word tuples and their likely attachments from the tagged and chunked textthe heuristic relies on the observed fact that in english and in languages with similar word order the attachment site of a preposition is usually located only a few words to the left of the prepositionfinally numbers are replaced by a single token the text is converted to lower case and the morphology database is used to find the base forms of the verbs and nounsthe extracted head word tuples differ from the training data used in previous supervised attempts in an important wayin the supervised case both of the potential sites namely the verb v and the noun n are known before the attachment is resolvedin the unsupervised case discussed here the extraction heuristic only finds what it thinks are unambiguous cases of prepositional phrase attachmenttherefore there is only one possible attachment site for the preposition and either the verb v or the noun n does not exist in the case of nounattached preposition or a verbattached preposition respectivelythis extraction heuristic loosely resembles a step in the bootstrapping procedure used to get training data for the classifier of in that step unambiguous attachments from the fidditch parser output are initially used to resolve some of the ambiguous attachments and the resolved cases are iteratively used to disambiguate the remaining unresolved casesour procedure differs critically from in that we do not iterate we extract unambiguous attachments from unparsed input sentences and we totally ignore the ambiguous casesit is the hypothesis of this approach that the information in just the unambiguous attachment events can resolve the ambiguous attachment events of the test datagiven a tagged and chunked sentence the extraction heuristic returns head word tuples of the form or where v is the verb n is the noun p is the preposition n2 is the object of the prepositionthe main idea of the extraction heuristic is that an attachment site of a preposition is usually within a few words to the left of the prepositionwe extract if table 1 also shows the result of the applying the extraction heuristic to a sample sentencethe heuristic ignores cases where p of since such cases are rarely ambiguous and we opt to model them deterministically as noun attachmentswe will report accuracies on both cases where p of and where p 0 ofalso the heuristic excludes examples with the verb to be from the training set since we found them to be unreliable sources of evidenceapplying the extraction heuristic to 970k unannotated sentences from the 1988 wall st journal data yields approximately 910k unique head word tuples of the form or the extraction heuristic is far from perfect when applied to and compared with the annotated wall st journal data of the penn treebank only 69 of the extracted head word tuples represent correct attachments2 the extracted tuples are meant to be a noisy but abundant substitute for the information that one might get from a treebanktables 2 and 3 list the most frequent extracted head word tuples for unambiguous verb and noun attachments respectivelymany of the frequent nounattached tuples such as num to num3 are incorrectthe prepositional phrase to num is usually attached to a verb such as rise or fall in the wall st journal domain eg profits rose 46 to 52 millionwhile the extracted tuples of the form and represent unambiguous noun and verb attachments in which either the verb or noun is known our eventual goal is to resolve ambiguous attachments in the test data of the form in which both the noun n and verb v are always knownwe therefore must use any information in the unambiguous cases to resolve the ambiguous casesa natural way is to use a classifier that compares the probability of each outcome we do not currently use n2 in the probability model and we omit it from further discussionwe can factor pr as follows the terms pr and pr are independent of the attachment a and need not be computed in c but the estimation of pr and pr is problematic since our training data ie the head words extracted from raw text occur with either n or v but never both n v this leads to make some heuristically motivated approximationslet the random variable 0 range over true false and let it denote the presence or absence of any preposition that is unambiguously attached to the noun or verb in questionthen p is the conditional probability that a particular noun n in free text has an unambiguous prepositional phrase attachmentwe approximate pr as follows the rationale behind this approximation is that the tendency of a v n pair towards a noun attachment is related to the tendency of the noun alone to occur with an unambiguous prepositional phrasethe z term exists only to make the approximation a well formed probability over a e in 171we approximate pr as follows the rationale behind these approximations is that when generating p given a noun attachment only the counts involving the noun are relevant assuming also that the noun has an attached prepositional phrase ie 0 truewe use word statistics from both the tagged corpus and the set of extracted head word tuples to estimate the probability of generating 0 true p and n2counts from the extracted set of tuples assume that 0 true while counts from the corpus itself may correspond to either 0 true or 0 false depending on if the noun if p of otherwise or verb in question is or is not respectively unambiguously attached to a prepositionthe quantities pr and pr denote the conditional probability that n or v will occur with some unambiguously attached preposition and are estimated as follows p and define cn ep cap as the number of noun attached tuplesanalogously define cv e c and cv ep cv the counts c and c are from the extracted head word tuplesusing the above notation we can interpolate as follows where c and c are counts from the tagged corpus and where c and c are counts from the extracted head word tuplesthe terms pr and pr denote the conditional probability that a particular preposition p will occur as an unambiguous attachment to n or v we present two techniques to estimate this probability one based on bigram counts and another based on an interpolation methodthis technique uses the bigram counts of the extracted head word tuples and backs off to the uniform distribution when the denominator is zero where 7 is the set of possible prepositions where all the counts c are from the extracted head word tuplesthis technique is similar to the one in and interpolates between the tendencies of the and bigrams and the tendency of the type of attachment towards a particular preposition p first define cn en c as the number of noun attached tuples with the prepositionapproximately 970k unannotated sentences from the 1988 wall st journal were processed in a manner identical to the example sentence in table 1the result was approximately 910000 head word tuples of the form or note that while the head word tuples represent correct attachments only 69 of the time their quantity is about 45 times greater than the quantity of data used in previous supervised approachesthe extracted data was used as training material for the three classifiers clbasel chnterp and cibigram each classifier is constructed as follows cl base this is the quotbaselinequot classifier that predicts n of p of and v otherwise cinterp this classifier has the form of equation uses the method in section 41 to generate 0 and the method in section 422 to generate p clbigram this classifier has the form of equation uses the method in section 41 to generate 0 and the method in section 421 to generate p table 4 shows accuracies of the classifiers on the test set of which is derived from the manually annotated attachments in the penn treebank wall st journal datathe penn treebank is drawn from the 1989 wall st journal data so there is no possibility of overlap with our training datafurthermore the extraction heuristic was developed and tuned on a quotdevelopment setquot ie a set of annotated examples that did not overlap with either the test set or the training settable 5 shows the two probabilities pr and pr using the same approximations as cbigram for the ambiguous example rise num to num and pr are not neededwhile the tuple is more frequent than the conditional probabilities prefer a v which is the choice that maximizes pr both classifiers cinterp and cbigram clearly outperform the baseline but the classifier does not outperform dbigraml even though it interpolates between the less specific evidence and more specific evidence this may be due to the errors in our extracted training data supervised classifiers that train from clean data typically benefit greatly by combining less specific evidence with more specific evidencedespite the errors in the training data the performance of the unsupervised classifiers begins to approach the best performance of the comparable supervised classifiers furthermore we do not use the second noun n2 whereas the best supervised methods use this informationour result shows that the information in imperfect but abundant data from unambiguous attachments as shown in tables 2 and 3 is sufficient to resolve ambiguous prepositional phrase attachments at accuracies just under the supervised stateoftheart accuracywe claim that our approach is portable to languages with similar word order and we support this claim by demonstrating our approach on the spanish languagewe used the spanish tagger and morphological analyzer developed at the xerox research centre europe4 and we modified the extraction heuristic to account for the new tagset and to account for the spanish equivalents of the words of and to be chunking was not performed on the spanish datawe used 450k sentences of raw text from the linguistic data consortium spanish news text collection to extract a training set and we used a nonoverlapping set of 50k sentences from the collection to create test setsthree native spanish speakers were asked to extract and annotate ambiguous instances of spanish prepositional phrase attachmentsthey annotated two sets one set consisted of all ambiguous prepositional phrase attachments of the form and the other set consisted of cases where p confor testing our classifier we used only those judgments on which all three annotators agreedthe performance of the classifiers clbigram dinterp and cbase when trained and tested on spanish language data are shown in table 6the spanish test set has fewer ambiguous prepositions than the english test set as shown by the accuracy of cibahowever the accuracy improvements of clbygrara over cl ba are statistically significant for both test sets5the unsupervised algorithm for prepositional phrase attachment presented here is the only algorithm in the published literature that can significantly outperform the baseline without using data derived from a treebank or parserthe accuracy of our technique approaches the accuracy of the best supervised methods and does so with only a tiny fraction of the supervisionsince only a small part of the extraction heuristic is specific to english and since partofspeech taggers and morphology databases are widely available in other languages our approach is far more portable than previous approaches for this problemwe successfully demonstrated the portability of our approach by applying it to the prepositional phrase attachment task in the spanish languagewe thank dr lauri kartunnen for lending us the spanish natural language tools and mike collins for helpful discussions on this work	P98-2177	statistical models for unsupervised prepositional phrase attachmentwe present several unsupervised statistical models for the prepositional phrase attachment task that approach the accuracy of the best supervised methods for this taskour unsupervised approach uses a heuristic based on attachment proximity and trains from raw text that is annotated with only partofspeech tags and morphological base forms as opposed to attachment informationit is therefore less resourceintensive and more portable than previous corpusbased algorithm proposed for this taskwe present results for prepositional phrase attachment in both english and spanishwe first assume noun attachment for all ofpps and then apply our disambiguation methods to all remaining pps
mindnet acquiring and structuring semantic information from text as a lexical knowledge base constructed automatically from the definitions and example sentences in two machinereadable dictionaries mindnet embodies several features that distinguish it from prior work with mrds it is however more than this static resource alone mindnet represents a general methodology for acquiring structuring accessing and exploiting semantic information from natural language text this paper provides an overview of the distinguishing characteristics of mindnet the steps involved in its creation and its extension beyond dictionary text in this paper we provide a description of the salient characteristics and functionality of mindnet as it exists today together with comparisons to related workwe conclude with a discussion on extending the mindnet methodology to the processing of other corpora and on future plans for mindnetfor additional details and background on the creation and use of mindnet readers are referred to richardson vanderwende and dolan et al mindnet is produced by a fully automatic process based on the use of a broadcoverage nl parsera fresh version of mindnet is built regularly as part of a normal regression processproblems introduced by daily changes to the underlying system or parsing grammar are quickly identified and fixedalthough there has been much research on the use of automatic methods for extracting information from dictionary definitions handcoded knowledge bases egwordnet continue to be the focus of ongoing researchthe euro wordnet project although continuing in the wordnet tradition includes a focus on semiautomated procedures for acquiring lexical contentoutside the realm of nlp we believe that automatic procedures such as mindnet provide the only credible prospect for acquiring world knowledge on the scale needed to support commonsense reasoningat the same time we acknowledge the potential need for the hand vetting of such information to insure accuracy and consistency in production level systemsthe extraction of the semantic information contained in mindnet exploits the very same broadcoverage parser used in the microsoft word 97 grammar checkerthis parser produces syntactic parse trees and deeper logical forms to which rules are applied that generate corresponding structures of semantic relationsthe parser has not been specially tuned to process dictionary definitionsall enhancements to the parser are geared to handle the immense variety of general text of which dictionary definitions are simply a modest subsetthere have been many other attempts to process dictionary definitions using heuristic pattern matching specially constructed definition parsers and even general coverage syntactic parsers however none of these has succeeded in producing the breadth of semantic relations across entire dictionaries that has been produced for mindnetvanderwende describes in detail the methodology used in the extraction of the semantic relations comprising mindneta truly broadcoverage parser is an essential component of this process and it is the basis for extending it to other sources of information such as encyclopedias and text corporathe different types of labeled semantic relations extracted by parsing for inclusion in mindnet are given in the table below these relation types may be contrasted with simple cooccurrence statistics used to create network structures from dictionaries by researchers including veronis and ide kozima and furugori and wilks et al labeled relations while more difficult to obtain provide greater power for resolving both structural attachment and word sense ambiguitieswhile many researchers have acknowledged the utility of labeled relations they have been at times either unable or unwilling to make the effort to obtain themthis deficiency limits the characterization of word pairs such as riverbank and writepen to simple relatedness whereas the labeled relations of mindnet specify precisely the relations riverpartbank and writemeanspenthe automatic extraction of semantic relations from a definition or example sentence for mindnet produces a hierarchical structure of these relations representing the entire definition or sentence from which they camesuch structures are stored in their entirety in mindnet and provide crucial context for some of the procedures described in later sections of this paperthe semrel structure for a definition of car is given in the figure below car quota vehicle with 3 or usu4 wheels and driven by a motor esp one one for carrying peoplequot early dictionarybased work focused on the extraction of paradigmatic relations in particular hypernym relations almost exclusively these relations as well as other syntagmatic ones have continued to take the form of relational triples the larger contexts from which these relations have been taken have generally not been retainedfor labeled relations only a few researchers have appeared to be interested in entire semantic structures extracted from dictionary definitions though they have not reported extracting a significant number of themafter semrel structures are created they are fully inverted and propagated throughout the entire mindnet database being linked to every word that appears in themsuch an inverted structure produced from a definition for motorist and linked to the entry for car is shown in the figure below researchers who produced spreading activation networks from mrds including veronis and ide and kozima and furugori typically only implemented forward links in those networkswords were not related backward to any of the headwords whose definitions mentioned them and words cooccurring in the same definition were not related directlyin the fully inverted structures stored in mindnet however all words are crosslinked no matter where they appearthe massive network of inverted semrel structures contained in mindnet invalidates the criticism leveled against dictionarybased methods by yarowsky and ide and veronis that lkbs created from mrds provide spotty coverage of a language at bestexperiments described elsewhere demonstrate the comprehensive coverage of the information contained in mindnetsome statistics indicating the size of the current version of mindnet and the processing time required to create it are provided in the table belowthe definitions and example sentences are from the longman dictionary of contemporary english and the american heritage dictionary ri edition inverted semrel structures facilitate the access to direct and indirect relationships between the root word of each structure which is the headword for the mindnet entry containing it and every other word contained in the structuresthese relationships consisting of one or more semantic relations connected together constitute semrel paths between two wordsfor example the semrel path between car and person in figure 2 above is carweighting schemes with similar goals are found in work by bradenharder and bookman many researchers both in the dictionary and corpusbased camps have worked extensively on developing methods to identify similarity between words since similarity determination is crucial to many word sense disambiguation and parametersmoothinginference procedureshowever some researchers have failed to distinguish between substitutional similarity and general relatednessthe similarity procedure of mindnet focuses on measuring substitutional similarity but a function is also provided for producing clusters of generally related wordstwo general strategies have been described in the literature for identifying substitutional similarityone is based on identifying direct paradigmatic relations between the words such as hypernym or synonymfor example paradigmatic relations in wordnet have been used by many to determine similarity including li et al and agirre and rigau the other strategy is based on identifying syntagmatic relations with other words that similar words have in commonsyntagmatic strategies for determining similarity have often been based on statistical analyses of large corpora that yield clusters of words occurring in similar bigram and trigram contexts as well as in similar predicateargument structure contexts there have been a number of attempts to combine paradigmatic and syntagmatic similarity strategies however none of these has completely integrated both syntagmatic and paradigmatic information into a single repository as is the case with mindnetthe mindnet similarity procedure is based on the topranked semrel paths between wordsfor example some of the top semrel paths in mindnet between pen and pencil are shown below penfmeansdrawmeans4pencil penfmeanswritemeanspencil penhyp4instrumenthyppencil penhypwritemeanspencil penmeanswrite many patterns being symmetrical but others notseveral experiments were performed in which word pairs from a thesaurus and an antithesaurus were used in a training phase to identify semrel path patterns that indicate similaritythese path patterns were then used in a testing phase to determine the substitutional similarity or dissimilarity of unseen word pairs the results summarized in the table below demonstrate the strength of this integrated approach which uniquely exploits both the paradigmatic and the syntagmatic relations in mindnettraining over 100000 word pairs from a thesaurus and antithesaurus produced 285000 semrel paths containing approx13500 unique path patternstesting over 100000 word pairs from a thesaurus and antithesaurus were evaluated using the path patternssimilar correct dissimilar correct 84 82 human benchmark random sample of 200 similar and dissimilar word pairs were evaluated by 5 humans and by mindnet similar correct dissimilar correct this powerful similarity procedure may also be used to extend the coverage of the relations in mindnetequivalent to the use of similarity determination in corpusbased approaches to infer absent ngrams or triples an inference procedure has been developed which allows semantic relations not presently in mindnet to be inferred from those that areit also exploits the topranked paths between the words in the relation to be inferredfor example if the relation watchmeansgelescope were not in mindnet it could be inferred by first finding the semrel paths between watch and telescope examining those paths to see if another word appears in a means relation with telescope and then checking the similarity between that word and watchas it turns out the word observe satisfies these conditions in the path watchhypobservemeanstelescope and therefore it may be inferred that one can watch by means of a telescopethe seamless integration of the inference and similarity procedures both utilizing the weighted extended paths derived from inverted semrel structures in mindnet is a unique strength of this approachan additional level of processing during the creation of mindnet seeks to provide sense identifiers on the words of semrel structurestypically word sense disambiguation occurs during the parsing of definitions and example sentences following the construction of logical forms detailed information from the parse both morphological and syntactic sharply reduces the range of senses that can be plausibly assigned to each wordother aspects of dictionary structure are also exploited including domain information associated with particular senses in processing normal input text outside of the context of mindnet creation wsd relies crucially on information from mindnet about how word senses are linked to one anotherto help mitigate this bootstrapping problem during the initial construction of mindnet we have experimented with a twopass approach to wsdduring a first pass a version of mindnet that does not include wsd is constructedthe result is a semantic network that nonetheless contains a great deal of quotambientquot information about sense assignmentsfor instance processing the definition spin 101 to produce thread yields a semrel structure in which the sense node spin101 is linked by a deep_subject relation to the undisambiguated form spideron the subsequent pass this information can be exploited by wsd in assigning sense 101 to the word spin in unrelated definitions wolf spider 100 any of various spidersthatdo not spin websthis kind of bootstrapping reflects the broader nature of our approach as discussed in the next section a fully and accurately disambiguated mindnet allows us to bootstrap senses onto words encountered in free text outside the dictionary domainthe creation of mindnet was never intended to be an end unto itselfinstead our emphasis has been on building a broadcoverage nlp understanding systemwe consider the methodology for creating mindnet to consist of a set of general tools for acquiring structuring accessing and exploiting semantic information from nl textour techniques for building mindnet are largely rulebasedhowever we arrive at these representations though the overall structure of mindnet can be regarded as crucially dependent on statisticswe have much more in common with traditional corpusbased approaches than a first glance might suggestan advantage we have over these approaches however is the rich structure imposed by the parse logical form and word sense disambiguation components of our systemthe statistics we use in the context of mindnet allow richer metrics because the data themselves are richerour first foray into the realm of processing free text with our methods has already been accomplished table 2 showed that some 58000 example sentences from ldoce and ahd3 were processed in the creation of our current mindnetto put our hypothesis to a much more rigorous test we have recently embarked on the assimilation of the entire text of the microsoft encarta 98 encyclopediawhile this has presented several new challenges in terms of volume alone we have nevertheless successfully completed a first pass and have produced and added semrel structures from the encarta 98 text to mindnetstatistics on that pass are given below besides our venture into additional english data we fully intend to apply the same methodologies to text in other languages as wellwe are currently developing nlp systems for 3 european and 3 asian languages french german and spanish chinese japanese and koreanthe syntactic parsers for some of these languages are already quite advanced and have been demonstrated publiclyas the systems for these languages mature we will create corresponding mindnets beginning as we did in english with the processing of machinereadable reference materials and then adding information gleaned from corpora	P98-2180	mindnet acquiring and structuring semantic information from textas a lexical knowledge base constructed automatically from the definitions and example sentences in two machinereadable dictionaries mindnet embodies several features that distinguish it from prior work with mrdsit is however more than this static resource alonemindnet represents a general methodology for acquiring structuring accessing and exploiting semantic information from natural language textmindnet is both an extraction methodology and a lexical ontology different from a word net since it was created automatically from a dictionary and its structure is based on such resources
nounphrase cooccurrence statistics for semiautomatic semantic lexicon construction semantic lexicons semiautomatically could be a great time saver relative to creating them by hand in this paper we present an algorithm for extracting potential entries for a category from an online corpus based upon a small set of exemplars our algorithm finds more correct terms and fewer incorrect ones than previous work in this area additionally the entries that are generated potentially provide broader coverage of the category than would occur to an individual coding them by hand our algorithm finds many terms not included within wordnet and could be viewed as an quotenhancerquot of existing broadcoverage resources semantic lexicons play an important role in many natural language processing taskseffective lexicons must often include many domainspecific terms so that available broad coverage resources such as wordnet are inadequatefor example both escort and chinook are types of vehicles but neither are cited as so in wordnetmanually building domainspecific lexicons can be a costly timeconsuming affairutilizing existing resources such as online corpora to aid in this task could improve performance both by decreasing the time to construct the lexicon and by improving its qualityextracting semantic information from word cooccurrence statistics has been effective particularly for sense disambiguation in riloff and shepherd noun cooccurrence statistics were used to indicate nominal category membership for the purpose of aiding in the construction of semantic lexiconsgenerically their algorithm can be outlined as follows our algorithm uses roughly this same generic structure but achieves notably superior results by changing the specifics of what counts as cooccurrence which figures of merit to use for new seed word selection and final ranking the method of initial seed word selection and how to manage compound nounsin sections 25 we will cover each of these topics in turnwe will also present some experimental results from two corpora and discuss criteria for judging the quality of the outputthe first question that must be answered in investigating this task is why one would expect it to work at allwhy would one expect that members of the same semantic category would cooccur in discoursein the word sense disambiguation task no such claim is made words can serve their disambiguating purpose regardless of partofspeech or semantic characteristicsin motivating their investigations riloff and shepherd cited several very specific noun constructions in which cooccurrence between nouns of the same semantic class would be expected including conjunctions lists appositives and noun compounds our algorithm focuses exclusively on these constructionsbecause the relationship between nouns in a compound is quite different than that between nouns in the other constructions the algorithm consists of two separate components one to deal with conjunctions lists and appositives and the other to deal with noun compoundsall compound nouns in the former constructions are represented by the head of the compoundwe made the simplifying assumptions that a compound noun is a string of consecutive nouns and that the head of the compound is the rightmost nounto identify conjunctions lists and appositives we first parsed the corpus using an efficient statistical parser trained on the penn wall street journal treebank we defined cooccurrence in these constructions using the standard definitions of dominance and precedencethe relation is stipulated to be transitive so that all head nouns in a list cooccur with each other two head nouns cooccur in this algorithm if they meet the following four conditions in contrast rs counted the closest noun to the left and the closest noun to the right of a head noun as cooccuring with itconsider the following sentence from the muc4 corpus quota cargo aircraft may drop bombs and a truck may be equipped with artillery for warquot in their algorithm both cargo and bombs would be counted as cooccuring with aircraftin our algorithm cooccurrence is only counted within a noun phrase between head nouns that are separated by a comma or conjunctionif the sentence had read quota cargo aircraft fighter plane or combat helicopter quot then aircraft plane and helicopter would all have counted as cooccuring with each other in our algorithmrs used the same figure of merit both for selecting new seed words and for ranking words in the final outputtheir figure of merit was simply the ratio of the times the noun coocurs with a noun in the seed list to the total frequency of the noun in the corpusthis statistic favors low frequency nouns and thus necessitates the inclusion of a minimum occurrence cutoffthey stipulated that no word occuring fewer than six times in the corpus would be considered by the algorithmthis cutoff has two effects it reduces the noise associated with the multitude of low frequency words and it removes from consideration a fairly large number of certainly valid category membersideally one would like to reduce the noise without reducing the number of valid nounsour statistics allow for the inclusion of rare occcurancesnote that this is particularly important given our algorithm since we have restricted the relevant occurrences to a specific type of structure even relatively common nouns may not occur in the corpus more than a handful of times in such a contextthe two figures of merit that we employ one to select and one to produce a final rank use the following two counts for each noun to select new seed words we take the ratio of count 1 to count 2 for the noun in questionthis is similar to the figure of merit used in rs and also tends to promote low frequency nounsfor the final ranking we chose the log likelihood statistic outlined in dunning which is based upon the cooccurrence counts of all nouns this statistic essentially measures how surprising the given pattern of cooccurrence would be if the distributions were completely randomfor instance suppose that two words occur forty times each and they cooccur twenty times in a millionword corpusthis would be more surprising for two completely random distributions than if they had each occurred twice and had always cooccurreda simple probability does not capture this factthe rationale for using two different statistics for this task is that each is well suited for its particular role and not particularly well suited to the otherwe have already mentioned that the simple ratio is ill suited to dealing with infrequent occurrencesit is thus a poor candidate for ranking the final output if that list includes words of as few as one occurrence in the corpusthe log likelihood statistic we found is poorly suited to selecting new seed words in an iterative algorithm of this sort because it promotes high frequency nouns which can then overly influence selections in future iterations if they are selected as seed wordswe termed this phenomenon infection and found that it can be so strong as to kill the further progress of a categoryfor example if we are processing the category vehicle and the word artillery is selected as a seed word a whole set of weapons that cooccur with artillery can now be selected in future iterationsif one of those weapons occurs frequently enough the scores for the words that it cooccurs with may exceed those of any vehicles and this effect may be strong enough that no vehicles are selected in any future iterationin addition because it promotes high frequency terms such a statistic tends to have the same effect as a minimum occurrence cutoff ie few if any low frequency words get addeda simple probability is a much more conservative statistic insofar as it selects far fewer words with the potential for infection it limits the extent of any infection that does occur and it includes rare wordsour motto in using this statistic for selection is quotfirst do no harmquotthe simple ratio used to select new seed words will tend not to select higher frequency words in the categorythe solution to this problem is to make the initial seed word selection from among the most frequent head nouns in the corpusthis is a sensible approach in any case since it provides the broadest coverage of category occurrences from which to select additional likely category membersin a task that can suffer from sparse data this is quite importantwe printed a list of the most common nouns in the corpus and selected category members by scanning through this listanother option would be to use head nouns identified in wordnet which as a set should include the most common members of the category in questionin general however the strength of an algorithm of this sort is in identifying infrequent or specialized termstable 1 shows the seed words that were used for some of the categories testedthe relationship between the nouns in a compound noun is very different from that in the other constructions we are consideringthe nonhead nouns in a compound noun may or may not be legitimate members of the categoryfor instance either pickup truck or pickup is a legitimate vehicle whereas cargo plane is legitimate but cargo is notfor this reason cooccurrence within noun compounds is not considered in the iterative portions of our algorithminstead all noun compounds with a head that is included in our final ranked list are evaluated for inclusion in a second listthe method for evaluating whether or not to include a noun compound in the second list is intended to exclude constructions such as government plane and include constructions such as fighter planesimply put the former does not correspond to a type of vehicle in the same way that the latter doeswe made the simplifying assumption that the higher the probability of the head given the nonhead noun the better the construction for our purposesfor instance if the noun government is found in a noun compound how likely is the head of that compound to be planehow does this compare to the noun fighterfor this purpose we take two counts for each noun in the compound for each nonhead noun in the compound we evaluate whether or not to omit it in the outputif all of them are omitted or if the resulting compound has already been output the entry is skippedeach noun is evaluated as follows first the head of that noun is determinedto get a sense of what is meant here consider the following compound nuclearpowered aircraft carrierin evaluating the word nuclearpowered it is unclear if this word is attached to aircraft or to carrierwhile we know that the head of the entire compound is carrier in order to properly evaluate the word in question we must determine which of the words following it is its headthis is done in the spirit of the dependency model of lauer by selecting the noun to its right in the compound with the highest probability of occuring with the word in question when occurring in a noun compoundonce the head of the word is determined the ratio of count 1 to count 2 is compared to an empirically set cutoffif it falls below that cutoff it is omittedif it does not fall below the cutoff then it is kept the input to the algorithm is a parsed corpus and a set of initial seed words for the desired categorynouns are matched with their plurals in the corpus and a single representation is settled upon for both eg carcooccurrence bigrams are collected for head nouns according to the notion of cooccurrence outlined abovethe algorithm then proceeds as followswe ran our algorithm against both the muc4 corpus and the wall street journal corpus for a variety of categories beginning with the categories of vehicle and weapon both included in the five categories that rks investigated in their paperother categories that we investigated were crimes people commercial sites states and machinesthis last category was run because of the sparse data for the category weapon in the wall street journalit represents roughly the same kind of category as weapon namely technological artifactsit in turn produced sparse results with the muc4 corpustables 3 and 4 show the top results on both the head noun and the compound noun lists generated for the categories we testedrs evaluated terms for the degree to which they are related to the categoryin contrast we counted valid only those entries that are clear members of the categoryrelated words did not counta valid instance was novel unique and a proper class within the category as an illustration of this last condition neither galileo probe nor gray plane is a valid entry the former because it denotes an individual and the latter because it is a class of planes based upon an incidental feature in the interests of generating as many valid entries as possible we allowed for the inclusion in noun compounds of words tagged as adjectives or cardinality wordsin certain occasions this is necessary to avoid losing key parts of the compoundmost common adjectives are dropped in our compound noun analysis since they occur with a wide variety of headswe determined three ways to evaluate the output of the algorithm for usefulnessthe first is the ratio of valid entries to total entries producedrs reported a ratio of 17 valid to total entries for both the vehicle and weapon categories on the same corpus our algorithm yielded a ratio of 329 valid to total entries for the category vehicle and 36 for the category weaponthis can be seen in the slope of the graphs in figure 1tables 2 and 5 give the relevant data for the categories that we investigatedin general the ratio of valid to total entries fell between 2 and 4 even in the cases that the output was relatively smalla second way to evaluate the algorithm is by the total number of valid entries producedas can be seen from the numbers reported in table 2 our algorithm generated from 24 to nearly 3 times as many valid terms for the two contrasting categories from the muc corpus than the algorithm of rseven more valid terms were generated for appropriate categories using the wall street journalanother way to evaluate the algorithm is with the number of valid entries produced that are not in wordnettable 2 presents these numbers for the categories vehicle and weaponwhereas the rs algorithm produced just 11 terms not already present in wordnet for the two categories combined our algorithm produced 106 or over 3 for every 5 valid terms producedit is for this reason that we are billing our algorithm as something that could enhance existing broadcoverage resources with domainspecific lexical informationwe have outlined an algorithm in this paper that as it stands could significantly speed up the task of building a semantic lexiconwe have also examined in detail the reasons why it works and have shown it to work well for multiple corpora and multiple categoriesthe algorithm generates many words not included in broad coverage resources such as wordnet and could be thought of as a wordnet quotenhancerquot for domainspecific applicationsmore generally the relative success of the algorithm demonstrates the potential benefit of narrowing corpus input to specific kinds of constructions despite the danger of compounding sparse data problemsto this end parsing is invaluablethanks to mark johnson for insightful discussion and to julie sedivy for helpful comments	P98-2182	nounphrase cooccurrence statistics for semiautomatic semantic lexicon constructiongenerating semantic lexicons semiautomatically could be a great time saver relative to creating them by handin this paper we present an algorithm for extracting potential entries for a category from an online corpus based upon a small set of exemplarsour algorithm finds more correct terms and fewer incorrect ones than previous work in this areaadditionally the entries that are generated potentially provide broader coverage of the category than would occur to an individual coding them by handour algorithm finds many terms not included within wordnet and could be viewed as an enhancer of existing broadcoverage resourceswe use cooccurrence statistics in local context to discover sibling relationsour experiments were performed using the muc4 and wall street journal corpuses to select seed words we rank all of the head nouns in the training corpus by frequency and manually select the first 10 nouns that unambiguously belong to each categorywe find that 3 of every 5 words learned by our system are not present in wordnet
never look back an alternative to centering i propose a model for determining the hearer attentional state which depends solely on a list of salient discourse entities the ordering among the elements of the slist covers also the of the center the centering model the ranking criteria for the slist based on the distinction between entities and incorporate preferences for interand intrasentential anaphora the model is the basis for an algorithm which operates incrementally word by word i propose a model for determining the hearer attentional state in understanding discoursemy proposal is inspired by the centering model and draws on the conclusions of strube hahn approach for the ranking of the forwardlooking center list for germantheir approach has been proven as the point of departure for a new model which is valid for english as wellthe use of the centering transitions in brennan et al algorithm prevents it from being applied incrementally in my approach i propose to replace the functions of the backwardlooking center and the centering transitions by the order among the elements of the list of salient discourse entities the slist ranking criteria define a preference for hearerold over hearernew discourse entities generalizing strube hahn approachbecause of these ranking criteria i can account for the difference in salience between definite nps and indefinite nps the slist is not a local data structure associated with individual utterancesthe slist rather describes the attentional state of the hearer at any given point in processing a discoursethe slist is generated incrementally word by word and used immediatelytherefore the slist integrates in the simplest manner preferences for inter and intrasentential anaphora making further specifications for processing complex sentences unnecessarysection 2 describes the centering model as the relevant background for my proposalin section 3 i introduce my model its only data structure the slist and the accompanying algorithmin section 4 i compare the results of my algorithm with the results of the centering algorithm with and without specifications for complex sentences the centering model describes the relation between the focus of attention the choices of referring expressions and the perceived coherence of discoursethe model has been motivated with evidence from preferences for the antecedents of pronouns and has been applied to pronoun resolution inter alia whose interpretation differs from the original modelthe centering model itself consists of two constructs the backwardlooking center and the list of forwardlooking centers and a few rules and constraintseach utterance ui is assigned a list of forwardlooking centers cf and a unique backwardlooking center cba ranking imposed on the elements of the cf reflects the assumption that the most highly ranked element of c f is most likely to be the cbthe most highly ranked element of c f that is realized in u2f1 is the cbtherefore the ranking on the cf plays a crucial role in the modelgrosz et al and brennan et al use grammatical relations to rank the cf but state that other factors might also play a rolefor their centering algorithm brennan et al extend the notion of centering transition relations which hold across adjacent utterances to differentiate types of shift preferred to retain is preferred to smoothshift is preferred to roughshiftthe bfpalgorithm consists of three basic steps to illustrate this algorithm we consider example which has two different final utterances and utterance contains one pronoun utterance two pronounswe look at the interpretation of and after step 2 the algorithm has produced two readings for each variant which are rated by the corresponding transitions in step 3in the pronoun quotshequot is resolved to quotherquot because the continue transition is ranked higher than smoothshift in the second readingin the pronoun quotshequot is resolved to quotfriedmanquot because smoothshift is preferred over roughshiftthe realization and the structure of my model departs significantly from the centering model in contrast to the centering model my model does not need a construct which looks back it does not need transitions and transition ranking criteriainstead of using the cb to account for local coherence in my model this is achieved by comparing the first element of the slist with the preceding statestrube hahn rank the cf according to the information status of discourse entitiesi here generalize these ranking criteria by redefining them in prince termsi distinguish between three different sets of expressions hearerold discourse entities mediated discourse entities and hearernew discourse entities these sets consist of the elements of prince familiarity scale old consists of evoked and unused discourse entities while new consists of brandnew discourse entitiesmed consists of inferrables containing inferrables and anchored brandnew discourse entitiesthese discourse entities are discoursenew but mediated by some hearerold discourse entity i do not assume any difference between the elements of each set with respect to their information statuseg evoked and unused discourse entities have the same information status because both belong to oldfor an operationalization of prince terms i stipulate that evoked discourse entitites are coreferring expressions unused discourse entities are proper names and titlesin texts brandnew proper names are usually accompanied by a relative clause or an appositive which relates them to the hearer knowledgethe corresponding discourse entity is evoked only after this elaborationwhenever these linguistic devices are missing proper names are treated as unusedl i restrict inferrables to the particular subset defined by hahn et al anchored brandnew discourse entities require that the anchor is either evoked or unusedi assume the following conventions for the ranking constraints on the elements of the slistthe 3tuple denotes a discourse entity x which is evoked in utterance uttx at the text position posxwith respect to any two discourse entities and uttx and utty specifying the current utterance ui or the preceding utterance u2_1 i set up the following ordering constraints on elements in the slist 2for any state of the processorhearer the ordering of discourse entities in the slist that can be derived from the ordering constraints to is denoted by the precedence relation if x e old and y e med then x yif x e old and y e new then x yif x e med and y e new then x y if x y e old or x y e med or x y e new then if utt utty then x y if utt utty and pos and indicate that the utterance containing x follows the utterance containing y or that x and y are elements of the same utterance between discourse entities in ui and discourse entities in u2_1 i am able to deal with intrasentential anaphorathere is no need for further specifications for complex sentencesa finer grained ordering is achieved by ranking discourse entities within each of the sets according to their text positionanaphora resolution is performed with a simple lookup in the slist3the elements of the slist are tested in the given order until one test succeedsjust after an anaphoric expression is resolved the slist is updatedthe algorithm processes a text from left to right 2if the analysis of utterance u5 is finished remove all discourse entities from the slist which are not realized in youthe analysis for example is given in table 36i show only these steps which are of interest for the computation of the slist and the pronoun resolutionthe preferences for pronouns are given by the slist immediately above themthe pronoun quotshequot in is resolved to the first element of the slistwhen the pronoun quotherquot in is encountered friedman is the first element of the slist since friedman is unused and in the current utterancebecause of binding restrictions quotherquot cannot be resolved to friedman but to the second element brennanin both and the pronoun quotshequot is resolved to friedman3the slist consists of referring expressions which are specified for text position agreement sortal information and information statuscoordinated nps are collected in a setthe slist does not contain predicative nps pleonastic quotitquot and any elements of direct speech enclosed in double quotesthe difference between my algorithm and the bfpalgorithm becomes clearer when the unused discourse entity quotfriedmanquot is replaced by a brandnew discourse entity eg quota professional driverquot7 in the bfpalgorithm the ranking of the cflist depends on grammatical roleshence driver is ranked higher than brennan in the cft2cin the pronoun quotshequot is resolved to brennan because of the preference for continue over retainin quotshequot is resolved to driver because smoothshift is preferred over roughshiftin my algorithm at the end of the evoked phrase quotherquot is ranked higher than the brandnew phrase quota professional driverquot in both and the pronoun quotshequot is resolved to brennanexample 8 illustrates how the preferences for intra and intersentential anaphora interact with the information status of discourse entitites sentence starts a new discourse segmentthe phrase quota judgequot is brandnewquotmr curtisquot is mentioned several times before in the text hence 71 owe this variant andrew kehlerthis example can misdirect readers because the phrase quota professional driverquot is assigned the quotdefaultquot gender masculineanyway this example like the original example seems not to be felicitous english and has only illustrative character81n the new york tunesdecember 7 1997 pa48 the discourse entity curtis is evoked and ranked higher than the discourse entity judgein the next step the ellipsis refers to judge which is evoked thenthe nouns quotrequestquot and quotprosecutorsquot are brandnew9the pronoun quothequot and the possessive pronoun quothisquot are resolved to curtisquotconditionquot is brandnew but anchored by the possessive pronounfor and i show only the steps immediately before the pronouns are resolvedin both quotmr curtisquot and quotthe judgequot are evokedhowever quotmr curtisquot is the leftmost evoked phrase in this sentence and therefore the most preferred antecedent for the pronoun quothimquotfor my experiments i restricted the length of the slist to five elementstherefore quotprosecutorsquot in is not contained in the slistthe discourse entity smirga is introduced in it becomes evoked after the appositivehence sm1rga is the most preferred antecedent for the pronoun quothequotin the first experiment i compare my algorithm with the bfpalgorithm which was in a second experiment extended by the constraints for complex sentences as described by kameyama methodi use the following guidelines for the handsimulated analysis i do not assume any world knowledge as part of the anaphora resolution processonly agreement criteria binding and sortal constraints are appliedi do not account for false positives and error chainsfollowing walker a segment is defined as a paragraph unless its first sentence has a pronoun in subject position or a pronoun where none of the preceding sentenceinternal noun phrases matches its syntactic featuresat the beginning of a segment anaphora resolution is preferentially performed within the same utterancemy algorithm starts with an empty slist at the beginning of a segmentthe basic unit for which the centering data structures are generated is the utterance youfor the bfpalgorithm i define you as a simple sentence a complex sentence or each full clause of a compound sentencekameyama intrasentential centering operates at the clause levelwhile tensed clauses are defined as utterances on their own untensed clauses are processed with the main clause so that the cflist of the main clause contains the elements of the untensed embedded clausekameyama distinguishes for tensed clauses further between sequential and hierarchical centeringexcept for reported speech nonreport complements and relative clauses all other types of tensed clauses build a chain of utterances on the same levelaccording to the preference for intersentential candidates in the centering model i define the following anaphora resolution strategy for the bepalgorithm test elements of uj_1 test elements of ui lefttoright test elements of cf cf in my algorithm steps and fall together is performed using previous states of the systemresultsthe test set consisted of the beginnings of three short stories by hemingway and three articles from the new york times the results of my experiments are given in table 6the first row gives the number of personal and possessive pronounsthe remainder of the table shows the results for the bfpalgorithm for the bfpalgorithm extended by kameyama intrasentential specifications and for my algorithmthe overall error rate of each approach is given in the rows marked with wrongthe rows marked with wrong give the numbers of errors directly produced by the algorithms strategy the rows marked with wrong the number of analyses with ambiguities generated by the bfpalgorithm the rows marked with wrong give the number of errors caused by specifications for intrasentential anaphorasince my algorithm integrates the specifications for intrasentential anaphora i count these errors as strategic errorsthe rows marked with wrong give the numbers of errors contained in error chainsthe rows marked with wrong give the numbers of the remaining errors interpretationthe results of my experiments showed not only that my algorithm performed better than the centering approaches but also revealed insight in the interaction between inter and intrasentential preferences for anaphoric antecedentskameyama specifications reduce the complexity in that the cflists in general are shorter after splitting up a sentence into clausestherefore the bfpalgorithm combined with her specifications has almost no strategic errors while the number of ambiguities remains constantbut this benefit is achieved at the expense of more errors caused by the intrasentential specificationsthese errors occur in cases like example in which kameyama intrasentential strategy makes the correct antecedent less salient indicating that a clausebased approach is too finegrained and that the hierarchical syntactical structure as assumed by kameyama does not have a great impact on anaphora resolutioni noted too that the bfpalgorithm can generate ambiguous readings for ui when the pronoun in ui does not cospecify the cbin cases where the c1 contains more than one possible antecedent for the pronoun several ambiguous readings with the same transitions are generatedan examplem there is no cb because no element of the preceding utterance is realized in the pronoun quotthemquot in cospecifies quotdeerquot but the bfpalgorithm generates two readings both of which are marked by a retain transition a jim pulled the burlap sacks off the deer b and liz looked at themin general the strength of the centering model is that it is possible to use the cb as the most preferred antecedent for a pronoun in youin my model this effect is achieved by the preference for hearerold discourse entitieswhenever this preference is misleading both approaches give wrong resultssince the cb is defined strictly local while hearerold discourse entities are defined global my model produces less errorsin my model the preference is available immediately while the bfpalgorithm can use its preference not before the second utterance has been processedthe more global definition of hearerold discourse entities leads also to shorter error chains however the test set is too small to draw final conclusions but at least for the texts analyzed the preference for hearerold discourse entities is more appropriate than the preference given by the bfp algorithmkameyama version of centering also omits the centering transitionsbut she uses the cb and a ranking over simplified transitions preventing the incremental application of her modelthe focus model accounts for evoked discourse entities explicitly because it uses the discourse focus which is determined by a successful anaphora resolutionincremental processing is not a topic of these paperseven models which use salience measures for determining the antecedents of pronoun use the concept of evoked discourse entitieshajieova et al assign the highest value to an evoked discourse entityalso lappin leass who give the subject of the current sentence the highest weight have an implicit notion of evokednessthe salience weight degrades from one sentence to another by a factor of two which implies that a repeatedly mentioned discourse entity gets a higher weight than a brandnew subjectin this paper i proposed a model for determining the hearer attentional state which is based on the distinction between hearerold and hearernew discourse entitiesi showed that my model though it omits the backwardlooking center and the centering transitions does not lose any of the predictive power of the centering model with respect to anaphora resolutionin contrast to the centering model my model includes a treatment for intrasentential anaphora and is sufficiently well specified to be applied to real textsits incremental character seems to be an answer to the question kehler recently raisedfurthermore it neither has the problem of inconsistency kehler mentioned with respect to the bfpalgorithm nor does it generate unnecessary ambiguitiesfuture work will address whether the text position which is the weakest grammatical concept is sufficient for the order of the elements of the slist at the second layer of my ranking constraintsi will also try to extend my model for the analysis of definite noun phrases for which it is necessary to integrate it into a more global model of discourse processingacknowledgments this work has been funded by a postdoctoral grant from dfg and is supported by a postdoctoral fellowship award from ircsi would like to thank nobo komagata rashmi prasad and matthew stone who commented on earlier drafts of this paperi am grateful for valuable comments by barbara grosz udo hahn aravind joshi lauri karttunen andrew kehler ellen prince and bonnie webber	P98-2204	never look back an alternative to centeringi propose a model for determining the hearer attentional state which depends solely on a list of salient discourse entities the ordering among the elements of the slist covers also the function of the backwardlooking center in the centering modelthe ranking criteria for the slist are based on the distinction between hearerold and hearernew discourse entities and incorporate preferences for inter and intrasentential anaphorathe model is the basis for an algorithm which operates incrementally word by wordwe argue that the information status of an antecedent is more important than the grammatical role in which it occurswe evaluate on handannotated datawe restrict our algorithm to the current and last sentence
measures of distributional similarity distributional similarity measures for the purpose of improving probability estimation for unseen cooccurrences our contributions are threefold an empirical comparison of a broad range of measures a classification an inherent problem for statistical methods in natural language processing is that of sparse data the inaccurate representation in any training corpus of the probability of low frequency eventsin particular reasonable events that happen to not occur in the training set may mistakenly be assigned a probability of zerothese unseen events generally make up a substantial portion of novel data for example essen and steinbiss report that 12 of the testset bigrams in a 7525 split of one million words did not occur in the training partitionwe consider here the question of how to estimate the conditional cooccurrence probability p of an unseen word pair drawn from some finite set n x v two stateoftheart technologies are katz backoff method and jelinek and mercer interpolation methodboth use p to estimate p when is unseen essentially ignoring the identity of n an alternative approach is distanceweighted averaging which arrives at an estimate for unseen cooccurrences by combining estimates for where s is a set of candidate similar words and sim is a function of the similarity between n and m we focus on distributional rather than semantic similarity because the goal of distanceweighted averaging is to smooth probability distributions although the words quotchancequot and quotprobabilityquot are synonyms the former may not be a good model for predicting what cooccurrences the latter is likely to participate inthere are many plausible measures of distributional similarityin previous work we compared the performance of three different functions the jensenshannon divergence the l1 norm and the confusion probabilityour experiments on a frequencycontrolled pseudoword disambiguation task showed that using any of the three in a distanceweighted averaging scheme yielded large improvements over katz backoff smoothing method in predicting unseen coocurrencesfurthermore by using a restricted version of model that stripped incomparable parameters we were able to empirically demonstrate that the confusion probability is fundamentally worse at selecting useful similar wordsd lin also found that the choice of similarity function can affect the quality of automaticallyconstructed thesauri to a statistically significant degree and the ability to determine common morphological roots by as much as 49 in precision 3the term quotsimilaritybasedquot which we have used previously has been applied to describe other models as well these empirical results indicate that investigating different similarity measures can lead to improved natural language processingon the other hand while there have been many similarity measures proposed and analyzed in the information retrieval literature there has been some doubt expressed in that community that the choice of similarity metric has any practical impact several authors have pointed out that the difference in retrieval performance achieved by different measures of association is insignificant providing that these are appropriately normalised but no contradiction arises because as van rijsbergen continues quotone would expect this since most measures incorporate the same informationquotin the languagemodeling domain there is currently no agreedupon best similarity metric because there is no agreement on what the quotsame informationquot the key data that a similarity function should incorporate isthe overall goal of the work described here was to discover these key characteristicsto this end we first compared a number of common similarity measures evaluating them in a parameterfree way on a decision taskwhen grouped by average performance they fell into several coherent classes which corresponded to the extent to which the functions focused on the intersection of the supports of the distributionsusing this insight we developed an informationtheoretic metric the skew divergence which incorporates the supportintersection data in an asymmetric fashionthis function yielded the best performance overall an average error rate reduction of 4 with respect to the jensenshannon divergence the best predictor of unseen events in our earlier experiments our contributions are thus threefold an empirical comparison of a broad range of similarity metrics using an evaluation methodology that factors out inessential degrees of freedom a proposal building on this comparison of a characteristic for classifying similarity functions and the introduction of a new similarity metric incorporating this characteristic that is superior at evaluating potential proxy distributionsin this section we describe the seven distributional similarity functions we initally evaluated2 for concreteness we choose n and v to be the set of nouns and the set of transitive verbs respectively a cooccurrence pair results when n appears as the head noun of the direct object of v we use p to denote probabilities assigned by a base language model let n and m be two nouns whose distributional similarity is to be determined for notational simplicity we write q for p and are for p their respective conditional verb cooccurrence probabilitiesfigure 1 lists several familiar functionsthe cosine metric and jaccard coefficient are commonly used in information retrieval as measures of association note that jaccard coefficient differs from all the other measures we consider in that it is essentially combinatorial being based only on the sizes of the supports of q r and q r rather than the actual values of the distributionspreviously we found the jensenshannon divergence to be a useful measure of the distance between distributions the function d is the kl divergence which measures the average inefficiency in using one distribution to code for another the function avgq denotes the average distribution avgqr r2 observe that its use ensures that the jensenshannon divergence is always definedin contrast d is undefined if q is not absolutely continuous with respect to are 2strictly speaking some of these functions are dissimilarity measures but each such function f can be recast as a similarity function via the simple transformation c f where c is an appropriate constantwhether we mean f or c f should be clear from context1 the confusion probability has been used by several authors to smooth word cooccurrence probabilities it measures the degree to which word m can be substituted into the contexts in which n appearsif the base language model probabilities obey certain bayesian consistency conditions as is the case for relative frequencies then we may write the confusion probability as follows note that it incorporates unigram probabilities as well as the two distributions q and r finally kendall t which appears in work on clustering similar adjectives is a nonparametric measure of the association between random variables in our context it looks for correlation between the behavior of q and r on pairs of verbsthree versions exist we use the simplest ta here sign q r 2 where sign is 1 for positive arguments 1 for negative arguments and 0 at 0the intuition behind kendall t is as followsassume all verbs have distinct conditional probabilitiesif sorting the verbs by the likelihoods assigned by q yields exactly the same ordering as that which results from ranking them according to r then t 1 if it yields exactly the opposite ordering then t 1we treat a value of 1 as indicating extreme dissimilarity3 it is worth noting at this point that there are several wellknown measures from the nlp literature that we have omitted from our experimentsarguably the most widely used is the mutual information it does not apply in the present setting because it does not measure the similarity between two arbitrary probability distributions and p but rather the similarity between a joint distribution p and the corresponding product distribution pp hammingtype metrics are intended for data with symbolic features since they count feature label mismatches whereas we are dealing feature values that are probabilitiesvariations of the value difference metric have been employed for supervised disambiguation but it is not reasonable in language modeling to expect training data tagged with correct probabilitiesthe dice coefficient is monotonic in jaccard coefficient so its inclusion in our experiments would be redundantfinally we did not use the kl divergence because it requires a smoothed base language modelwe evaluated the similarity functions introduced in the previous section on a binary decision task using the same experimental framework as in our previous preliminary comparison that is the data consisted of the verbobject cooccurrence pairs in the 1988 associated press newswire involving the 1000 most frequent nouns extracted via church and yarowsky processing tools587833 of the pairs served as a training set from which to calculate base probabilitiesfrom the other 20 we prepared test sets as follows after discarding pairs occurring in the training data we split the remaining pairs into five partitions and replaced each nounverb pair with a nounverbverb triple such that p pthe task for the language model under evaluation was to reconstruct which of and was the original cooccurrencenote that by construction was always the correct answer and furthermore methods relying solely on unigram frequencies would perform no better than chancetestset performance was measured by the error rate defined as 12 where t is the number of test triple tokens in the set and a tie results when both alternatives are deemed equally likely by the language model in questionto perform the evaluation we incorporated each similarity function into a decision rule as followsfor a given similarity measure f and neighborhood size k let sfk denote the k most similar words to n according to f we define the evidence according to f for the cooccurrence as then the decision rule was to choose the alternative with the greatest evidencethe reason we used a restricted version of the distanceweighted averaging model was that we sought to discover fundamental differences in behaviorbecause we have a binary decision task efk simply counts the number of k nearest neighbors to n that make the right decisionif we have two functions f and g such that efk egk then the k most similar words according to f are on the whole better predictors than the k most similar words according to g hence f induces an inherently better similarity ranking for distanceweighted averagingthe difficulty with using the full model for comparison purposes is that fundamental differences can be obscured by issues of weightingfor example suppose the probability estimate e are performed poorlywe would not be able to tell whether the because was an inherent deficiency in the l1 norm or just a poor choice of weight function perhaps 2 would have yielded better estimatesfigure 2 shows how the average error rate varies with k for the seven similarity metrics introduced aboveas previously mentioned a steeper slope indicates a better similarity rankingall the curves have a generally upward trend but always lie far below backoff they meet at k 1000 because sf jdx1 is always the set of all nounswe see that the functions fall into four groups the l2 norm kendall t the confusion probability and the cosine metric and the l1 norm jensenshannon divergence and jaccard coefficientwe can account for the similar performance of various metrics by analyzing how they incorporate information from the intersection of the supports of q and are consider the following supports we can rewrite the similarity functions from section 2 in terms of these sets making use of the identities e vevqwqr q evevqr q evevrwqr are evevqr are 1table 1 lists these alternative forms in order of performancewe see that for the noncombinatorial functions the groups correspond to the degree to which the measures rely on the verbs in vqrthe jensenshannon divergence and the l1 norm can be computed simply by knowing the values of q and r on vqrfor the cosine and the confusion probability the distribution values on vqr are key but other information is also incorporatedthe statistic ta takes into account all verbs including those that occur neither with the superior performance of jac seems to underscore the importance of the set vqrjaccard coefficient ignores the values of q and r on vqr but we see that simply knowing the size of vqr relative to the supports of q and r leads to good rankings4 the skew divergence based on the results just described it appears that it is desirable to have a similarity function that focuses on the verbs that cooccur with both of the nouns being comparedhowever we can make a further observation with the exception of the confusion probability all the functions we compared are symmetric that is f f but the substitutability of one word for another need not symmetricfor instance quotfruitquot may be the best possible approximation to quotapplequot but the distribution of quotapplequot may not be a suitable proxy for the distribution of quotfruitquot 4 in accordance with this insight we developed a novel asymmetric generalization of the kl divergence the askew divergence scr d r for 0 a 1it can easily be shown that sc depends only on the verbs in vqrnote that at a 1 the skew divergence is exactly the kl divergence and s12 is twice one of the summands of js 40n a related note an anonymous reviewer cited the 30 following example from the psychology literature we can say smith lecture is like a sleeping pill but quotnot the other way roundquot average error rate error rates definition of commonality is left to the user we view the empirical approach taken in this paper as complementary to linthat is we are working in the context of a particular application and while we have no mathematical certainty of the importance of the quotcommon supportquot information we did not assume it a priori rather we let the performance data guide our thinkingfinally we observe that the skew metric seems quite promisingwe conjecture that appropriate values for a may inversely correspond to the degree of sparseness in the data and intend in the future to test this conjecture on largerscale prediction taskswe also plan to evaluate skewed versions of the jensenshannon divergence proposed by rao and j lin thanks to claire cardie jon kleinberg fernando pereira and stuart shieber for helpful discussions the anonymous reviewers for their insightful comments fernando pereira for access to computational resources at att and stuart shieber for the opportunity to pursue this work at harvard university under nsf grant noiri9712068	P99-1004	measures of distributional similaritywe study distributional similarity measures for the purpose of improving probability estimation for unseen cooccurrenceswe use verbobject relations in both active and passive voice constructionswe find that our asymmetric skew divergence a generalisation of kullbackleibler divergence persorms best for improving probability estimates for unseen word cooccurrences
finding parts in very large corpora we present a method for extracting parts of objects from wholes given a very large corpus our method finds part words with 55 accuracy for the top 50 words as ranked by the system the part list could be scanned by an enduser and added to an existing ontology or used as a part of a rough semantic lexicon we present a method of extracting parts of objects from wholes to be more precise given a single word denoting some entity that has recognizable parts the system finds and rankorders other words that may denote parts of the entity in questionthus the relation found is strictly speaking between words a relation miller 1 calls quotmeronymyquot in this paper we use the more colloquial quotpartofquot terminologywe produce words with 55 accuracy for the top 50 words ranked by the system given a very large corpuslacking an objective definition of the partof relation we use the majority judgment of five human subjects to decide which proposed parts are correctthe program output could be scanned by an enduser and added to an existing ontology or used as a part of a rough semantic lexiconto the best of our knowledge there is no published work on automatically finding parts from unlabeled corporacasting our nets wider the work most similar to what we present here is that by hearst 2 on acquisition of hyponyms in that paper hearst finds lexical correlates to the hyponym relations by looking in text for cases where known hyponyms appear in proximity as in quotboats cars and other vehiclesquot tests the proposed patterns for validity and uses them to extract relations from a corpusin this paper we apply much the same methodology to the partof relationindeed in 2 hearst states that she tried to apply this strategy to the partof relation but failedwe comment later on the differences in our approach that we believe were most important to our comparative successlooking more widely still there is an evergrowing literature on the use of statisticalcorpusbased techniques in the automatic acquisition of lexicalsemantic knowledge we take it as axiomatic that such knowledge is tremendously useful in a wide variety of tasks from lowerlevel tasks like nounphrase reference and parsing to userlevel tasks such as web searches question answering and digestingcertainly the large number of projects that use wordnet 1 would support this contentionand although wordnet is handbuilt there is general agreement that corpusbased methods have an advantage in the relative completeness of their coverage particularly when used as supplements to the more laborintensive methodswebster dictionary defines quotpartquot as quotone of the often indefinite or unequal subdivisions into which something is or is regarded as divided and which together constitute the wholequot the vagueness of this definition translates into a lack of guidance on exactly what constitutes a part which in turn translates into some doubts about evaluating the results of any procedure that claims to find themmore specifically note that the definition does not claim that parts must be physical objectsthus say quotnovelquot might have quotplotquot as a partin this study we handle this problem by asking informants which words in a list are parts of some target word and then declaring majority opinion to be correctwe give more details on this aspect of the study laterhere we simply note that while our subjects often disagreed there was fair consensus that what might count as a part depends on the nature of the word a physical object yields physical parts an institution yields its members and a concept yields its characteristics and processesin other words quotfloorquot is part of quotbuildingquot and quotplotquot is part of quotbookquot our first goal is to find lexical patterns that tend to indicate partwhole relationsfollowing hearst 2 we find possible patterns by taking two words that are in a partwhole relation and finding sentences in our corpus from ldc that have these words within close proximitythe first few such sentences are the basement of the building the basement in question is in a fourstory apartment building the basement of the apartment buildingfrom the building basement the basement of a building the basements of buildings from these examples we construct the five patterns shown in table 1we assume here that parts and wholes are represented by individual lexical items as opposed to complete noun phrases or as a sequence of quotimportantquot noun modifiers together with the headthis occasionally causes problems eg quotconditionerquot was marked by our informants as not part of quotcarquot whereas quotair conditionerquot probably would have made it into a part listnevertheless in most cases head nouns have worked quite well on their ownwe evaluated these patterns by observing how they performed in an experiment on a single exampletable 2 shows the 20 highest ranked part words for each of the patterns ae table 2 shows patterns a and b clearly outperform patterns c d and e although parts occur in all five patterns the lists for a and b are predominately partsorientedthe relatively poor performance of patterns c and e was anticipated as many things occur quotinquot cars other than their partspattern d is not so obviously bad as it differs from the plural case of pattern b only in the lack of the determiner quotthequot or quotaquothowever this difference proves critical in that pattern d tends to pick up quotcountingquot nouns such as quottruckloadquot on the basis of this experiment we decided to proceed using only patterns a and b from table 1we use the ldc north american news corpus which is a compilation of the wire output of several us newspapersthe total corpus is about 100000000 wordswe ran our program on the whole data set which takes roughly four hours on our networkthe bulk of that time is spent tagging the corpusas is typical in this sort of work we assume that our evidence is independently and identically distributed we have found this assumption reasonable but its breakdown has led to a few errorsin particular a drawback of the nanc is the occurrence of repeated articles since the corpus consists of all of the articles that come over the wire some days include multiple updated versions of the same story containing identical paragraphs or sentenceswe wrote programs to weed out such cases but ultimately found them of little usefirst quotupdatequot articles still have substantial variation so there is a continuum between these and articles that are simply on the same topicsecond our data is so sparse that any such repeats are very unlikely to manifest themselves as repeated examples of parttype patternsnevertheless since two or three occurrences of a word can make it rank highly our results have a few anomalies that stem from failure of the lid assumption our seeds are one word and its pluralwe do not claim that all single words would fare as well as our seeds as we picked highly probable words for our corpus that we thought would have parts that might also be mentioned thereinwith enough text one could probably get reasonable results with any noun that met these criteriathe program has three phasesthe first identifies and records all occurrences of patterns a and b in our corpusthe second filters out all words ending with quotingquot quotnessquot or quotityquot since these suffixes typically occur in words that denote a quality rather than a physical objectfinally we order the possible parts by the likelihood that they are true parts according to some appropriate metricwe took some care in the selection of this metricat an intuitive level the metric should be something like p states that w appears in the patterns ab as a whole while p states that p appears as a partmetrics of the form p have the desirable property that they are invariant over p with radically different base frequencies and for this reason have been widely used in corpusbased lexical semantic research 369however in making this intuitive idea someone more precise we found two closely related versions we call metrics based on the first of these quotloosely conditionedquot and those based on the second quotstrongly conditionedquotwhile invariance with respect to frequency is generally a good property such invariant metrics can lead to bad results when used with sparse datain particular if a part word p has occurred only once in the data in the ab patterns then perforce p 1 for the entity w with which it is pairedthus this metric must be tempered to take into account the quantity of data that supports its conclusionto put this another way we want to pick pairs that have two properties p is high and i to p is largewe need a metric that combines these two desiderata in a natural waywe tried two such metricsthe first is dunning 10 loglikelihood metric which measures how quotsurprisedquot one would be to observe the data counts wp wp i i to p i and i w19 i if one assumes that pintuitively this will be high when the observed p p and when the counts supporting this calculation are largethe second metric is proposed by johnson he suggests asking the question how far apart can we be sure the distributions pand p are if we require a particular significance level say 05 or 01we call this new test the quotsignificantdifferencequot test or sigdiffjohnson observes that compared to sigdiff loglikelihood tends to overestimate the importance of data frequency at the expense of the distance between p and ptable 3 shows the 20 highest ranked words for each statistical method using the seed word quotcarquot the first group contains the words found for the method we perceive as the most accurate sigdiff and strong conditioningthe other groups show the differences between them and the first groupthe category means that this method adds the word to its list means the oppositefor example quotbackquot is on the sigdiffloose list but not the sigdiffstrong listin general sigdiff worked better than surprise and strong conditioning worked better than loose conditioningin both cases the less favored methods tend to promote words that are less specific furthermore the combination of sigdiff and strong conditioning worked better than either by itselfthus all results in this paper unless explicitly noted otherwise were gathered using sigdiff and strong conditioning combinedwe tested five subjects for their concept of a quotpartquot we asked them to rate sets of 100 words of which 50 were in our final results settables 6 11 show the top 50 words for each of our six seed words along with the number of subjects who marked the word as a part of the seed conceptthe score of individual words vary greatly but there was relative consensus on most wordswe put an asterisk next to words that the majority subjects marked as correctlacking a formal definition of part we can only define those words as correct and the rest as wrongwhile the scoring is admittedly not perfect it provides an adequate reference resulttable 4 summarizes these resultsthere we show the number of correct part words in the top 10 20 30 40 and 50 parts for each seed overall about 55 of the top 50 words for each seed are parts and about 70 of the top 20 for each seedthe reader should also note that we tried one ambiguous word quotplantquot to see what would happenour program finds parts corresponding to both senses though given the nature of our text the industrial use is more commonour subjects marked both kinds of parts as correct but even so this produced the weakest part list of the six words we triedas a baseline we also tried using as our quotpatternquot the head nouns that immediately surround our target wordwe then applied the same quotstrong conditioning sigdiffquot statistical test to rank the candidatesthis performed quite poorlyof the top 50 candidates for each target only 8 were parts as opposed to the 55 for our programwe also compared out parts list to those of wordnettable 5 shows the parts of quotcarquot in wordnet that are not in our top 20 and the words in our top 20 that are not in wordnet there are definite tradeoffs although we would argue that our top20 set is both more specific and more comprehensivetwo notable words our top 20 lack are quotenginequot and quotdoorquot both of which occur before 100more generally all wordnet parts occur somewhere before 500 with the exception of quottailfinquot which never occurs with carit would seem that our program would be a good tool for expanding wordnet as a person can to the entire statistical nlp group at brown and scan and mark the list of part words in a few minutes particularly to mark johnson brian roark gideon mann and anamaria popescu who provided invaluable help on the projectthe program presented here can find parts of objects given a word denoting the whole object and a large corpus of unmarked textthe program is about 55 accurate for the top 50 proposed parts for each of six examples upon which we tested itthere does not seem to be a single because for the 45 of the cases that are mistakeswe present here a few problems that have caught our attentionidiomatic phrases like quota jalopy of a carquot or quotthe son of a gunquot provide problems that are not easily weeded outdepending on the data these phrases can be as prevalent as the legitimate partsin some cases problems arose because of tagger mistakesfor example quotreenactmentquot would be found as part of a quotcarquot using pattern b in the phrase quotthe reenactment of the car crashquot if quotcrashquot is tagged as a verbthe program had some tendency to find qualities of objectsfor example quotdriveabilityquot is strongly correlated with carwe try to weed out most of the qualities by removing words with the suffixes quotnessquot quotingquot and quotityquot the most persistent problem is sparse data which is the source of most of the noisemore data would almost certainly allow us to produce better lists both because the statistics we are currently collecting would be more accurate but also because larger numbers would allow us to find other reliable indicatorsfor example idiomatic phrases might be recognized as suchso we see quotjalopy of a carquot but not of course quotthe car jalopyquotwords that appear in only one of the two patterns are suspect but to use this rule we need sufficient counts on the good words to be sure we have a representative sampleat 100 million words the nanc is not exactly small but we were able to process it in about four hours with the machines at our disposal so still larger corpora would not be out of the questionfinally as noted above hearst 2 tried to find parts in corpora but did not achieve good resultsshe does not say what procedures were used but assuming that the work closely paralleled her work on hyponyms we suspect that our relative success was due to our very large corpus and the use of more refined statistical measures for ranking the outputthis research was funded in part by nsf grant iri9319516 and onr grant n00149610549thanks	P99-1008	finding parts in very large corporawe present a method for extracting parts of objects from wholes given a very large corpus our method finds part words with 55 accuracy for the top 50 words as ranked by the systemthe part list could be scanned by an enduser and added to an existing ontology or used as a part of a rough semantic lexiconto filter out attributes that are regarded as qualities rather than parts we remove words ending with the suffixes ness ing and ity
inducing a semantically annotated lexicon via thembased clustering we present a technique for automatic induction of slot annotations for subcategorization frames based on induction of hidden classes in the them framework of statistical estimation the models are empirically evalutated by a general decision test induction of slot labeling for subcategorization frames is accomplished by a further application of them and applied experimentally on frame observations derived from parsing large corpora we outline an interpretation of the learned representations as theoreticallinguistic decompositional lexical entries an important challenge in computational linguistics concerns the construction of largescale computational lexicons for the numerous natural languages where very large samples of language use are now availableresnik initiated research into the automatic acquisition of semantic selectional restrictionsribas presented an approach which takes into account the syntactic position of the elements whose semantic relation is to be acquiredhowever those and most of the following approaches require as a prerequisite a fixed taxonomy of semantic relationsthis is a problem because entailment hierarchies are presently available for few languages and we regard it as an open question whether and to what degree existing designs for lexical hierarchies are appropriate for representing lexical meaningboth of these considerations suggest the relevance of inductive and experimental approaches to the construction of lexicons with semantic informationthis paper presents a method for automatic induction of semantically annotated subcategorization frames from unannotated corporawe use a statistical subcatinduction system which estimates probability distributions and corpus frequencies for pairs of a head and a subcat frame the statistical parser can also collect frequencies for the nominal fillers of slots in a subcat framethe induction of labels for slots in a frame is based upon estimation of a probability distribution over tuples consisting of a class label a selecting head a grammatical relation and a filler headthe class label is treated as hidden data in the emframework for statistical estimationin our clustering approach classes are derived directly from distributional dataa sample of pairs of verbs and nouns gathered by parsing an unannotated corpus and extracting the fillers of grammatical relationssemantic classes corresponding to such pairs are viewed as hidden variables or unobserved data in the context of maximum likelihood estimation from incomplete data via the them algorithmthis approach allows us to work in a mathematically welldefined framework of statistical inference ie standard monotonicity and convergence results for the them algorithm extend to our methodthe two main tasks of thembased clustering are i the induction of a smooth probability model on the data and ii the automatic discovery of classstructure in the databoth of these aspects are respected in our application of lexicon inductionthe basic ideas of our thembased clustering approach were presented in rooth our approach constrasts with the merely heuristic and empirical justification of similaritybased approaches to clustering for which so far no clear probabilistic interpretation has been giventhe probability model we use can be found earlier in pereira et al however in contrast to this approach our statistical inference method for clustering is formalized clearly as an themalgorithmapproaches to probabilistic clustering similar to ours were presented recently in saul and pereira and hofmann and puzicha there also themalgorithms for similar probability models have been derived but applied only to simpler tasks not involving a combination of embased clustering models as in our lexicon induction experimentfor further applications of our clustering model see rooth et al we seek to derive a joint distribution of verbnoun pairs from a large sample of pairs of verbs v e v and nouns n e n the key idea is to view v and n as conditioned on a hidden class c e c where the classes are given no prior interpretationthe semantically smoothed probability of a pair is defined to be the joint distribution p is defined by p pppnote that by construction conditioning of v and n on each other is solely made through the classes c in the framework of the them algorithm we can formalize clustering as an estimation problem for a latent class model as followswe are given a sample space of observed incomplete data corresponding to pairs from v x n a sample space x of unobserved complete data corresponding to triples from cx v x n a set x x e x i x c e ci of complete data related to the observation y a completedata specification po corresponding to the joint probability p over cx v x n with parametervector 0 an incomplete data specification p0 which is related to the completedata specification as the marginal probability p0 expo the them algorithm is directed at finding a value a of 0 that maximizes the incompletedata loglikelihood function l as a function of 0 for a given sample ie a arg max l where l ln fly po 0 as prescribed by the them algorithm the parameters of l are estimated indirectly by proceeding iteratively in terms of completedata estimation for the auxiliary function q which is the conditional expectation of the completedata loglikelihood lnpo given the observed data y and the current fit of the parameter values 0this auxiliary function is iteratively maximized as a function of 0 where each iteration is defined by note that our application is an instance of the themalgorithm for contextfree models from which the following particularily simple reestimation formulae can be derivedlet x for fixed c and ythen probabilistic contextfree grammar of gave for the british national corpus intuitively the conditional expectation of the number of times a particular v n or c choice is made during the derivation is prorated by the conditionally expected total number of times a choice of the same kind is madeas shown by baum et al these expectations can be calculated efficiently using dynamic programming techniquesevery such maximization step increases the loglikelihood function l and a sequence of reestimates eventually converges to a maximum of l in the following we will present some examples of induced clustersinput to the clustering algorithm was a training corpus of 1280715 tokens of verbnoun pairs participating in the grammatical relations of intransitive and transitive verbs and their subject and objectfillersthe data were gathered from the maximalprobability parses the headlexicalized fig2 shows an induced semantic class out of a model with 35 classesat the top are listed the 20 most probable nouns in the p distribution and their probabilities and at left are the 30 most probable verbs in the p distribution5 is the class indexthose verbnoun pairs which were seen in the training data appear with a dot in the class matrixverbs with suffix as s indicate the subject slot of an active intransitivesimilarily aso s denotes the subject slot of an active transitive and aso o denotes the object slot of an active transitivethus v in the above discussion actually consists of a combination of a verb with a subcat frame slot as s aso s or aso oinduced classes often have a basis in lexical semantics class 5 can be interpreted as clustering agents denoted by proper names quotmanquot and quotwomanquot together with verbs denoting communicative actionfig1 shows a cluster involving verbs of scalar change and things which can move along scalesfig5 can be interpreted as involving different dispositions and modes of their execution60 100 160 200 260 300 nix el cleeeewe evaluated our clustering models on a pseudodisambiguation task similar to that performed in pereira et al but differing in detailthe task is to judge which of two verbs v and v is more likely to take a given noun n as its argument where the pair has been cut out of the original corpus and the pair is constructed by pairing 71 with a randomly chosen verb v such that the combination is completely unseenthus this test evaluates how well the models generalize over unseen verbsthe data for this test were built as followswe constructed an evaluation corpus of triples by randomly cutting a test corpus of 3000 pairs out of the original corpus of 1280712 tokens leaving a training corpus of 1178698 tokenseach noun n in the test corpus was combined with a verb v which was randomly chosen according to its frequency such that the pair did appear neither in the training nor in the test corpushowever the elements v v and n were required to be part of the training corpusfurthermore we restricted the verbs and nouns in the evalutation corpus to the ones which occured at least 30 times and at most 3000 times with some verbfunctor v in the training corpusthe resulting 1337 evaluation triples were used to evaluate a sequence of clustering models trained from the training corpusthe clustering models we evaluated were parametrized in starting values of the training algorithm in the number of classes of the model and in the number of iteration steps resulting in a sequence of 3 x 10 x 6 modelsstarting from a lower bound of 50 random choice accuracy was calculated as the number of times the model decided for p p out of all choices madefig3 shows the evaluation results for models trained with 50 iterations averaged over starting values and plotted against class cardinalitydifferent starting values had an effect of 2 on the performance of the testwe obtained a value of about 80 accuracy for models between 25 and 100 classesmodels with more than 100 classes show a small but stable overfitting effecta second experiment addressed the smoothing power of the model by counting the number of pairs in the set v xn of all possible combinations of verbs and nouns which received a positive joint probability by the modelthe v x nspace for the above clustering models included about 425 million combinations we approximated the smoothing size of a model by randomly sampling 1000 pairs from v x n and returning the percentage of positively assigned pairs in the random samplefig4 plots the smoothing results for the above models against the number of classesstarting values had an influence of 1 on performancegiven the proportion of the number of types in the training corpus to the v x nspace without clustering we have a smoothing power of 014 whereas for example a model with 50 classes and 50 iterations has a smoothing power of about 93 corresponding to the maximum likelihood paradigm the number of training iterations had a decreasing effect on the smoothing performance whereas the accuracy of the pseudodisambiguation was increasing in the number of iterationswe found a number of 50 iterations to be a good compromise in this tradeoffthe goal of the following experiment was to derive a lexicon of several hundred intransitive and transitive verbs with subcat slots labeled with latent classesto induce latent classes for the subject slot of a fixed intransitive verb the following statistical inference step was performedgiven a latent class model plc for verbnoun pairs and a sample n1 nm of subjects for a fixed intransitive verb we calculate the probability of an arbitrary subject n e n by the estimation of the parametervector 0 can be formalized in the them framework by viewing p or p as a function of 0 for fixed plcthe reestimation formulae resulting from the incomplete data estimation for these probability functions have the following form is the frequency of n in the sample of subjects of the fixed verb a similar them induction process can be applied also to pairs of nouns thus enabling induction of latent semantic annotations for transitive verb framesgiven a lc model na for verbnoun pairs and a sample 1 of noun arguments for a fixed transitive verb we calculate the probability of its noun argument pairs by in an them framework by viewing p or p as a function of 0 for fixed plcthe reestimation formulae resulting from this incomplete data estimation problem have the following simple form is the frequency of in the sample of noun argument pairs of the fixed verb note that the class distributions p and p for intransitive and transitive models can be computed also for verbs unseen in the lc model blush 5 0982975 snarl 5 0962094 constance 3 mandeyille 2 christina 3 jinkwa 2 willie 299737 man 199859 ronni 2 scott 199761 claudia 2 omalley 199755 gabriel 2 shamlou 1 maggie 2 angalo 1 bathsheba 2 corbett 1 sarah 2 southgate 1 girl 19977 ace 1 experiments used a model with 35 classesfrom maximal probability parses for the british national corpus derived with a statistical parser we extracted frequency tables for intransitve verbsubject pairs and transitive verbsubjectobject triplesthe 500 most frequent verbs were selected for slot labelingfig6 shows two verbs v for which the most probable class label is 5 a class which we earlier described as communicative action together with the estimated frequencies of f pe for those ten nouns n for which this estimated frequency is highestfig7 shows corresponding data for an intransitive scalar motion sense of increasefig8 shows the intransitive verbs which take 17 as the most probable labelintuitively the verbs are semantically coherentwhen compared to levin 48 toplevel verb classes we found an agreement of our classification with her class of quotverbs of changes of statequot except for the last three verbs in the list in fig8 which is sorted by probability of the class labelsimilar results for german intransitive scalar motion verbs are shown in fig9the data for these experiments were extracted from the maximalprobability parses of a 41 million word corpus of german subordinate clauses yielding 418290 tokens of pairs of verbs or adjectives and nounsthe lexicalized probabilistic grammar for german used is described in beil et al we compared the german example of scalar motion verbs to the linguistic classification of verbs given by schuhmacher and found an agreement of our classification with the class of quoteinfache anderungsverbenquot except for the verbs anwachsen and stagnieren which were not classified there at allfig10 shows the most probable pair of classes for increase as a transitive verb together with estimated frequencies for the head filler pairnote that the object label 17 is the class found with intransitive scalar motion verbs this correspondence is exploited in the next sectionin some linguistic accounts multiplace verbs are decomposed into representations involving one predicate or relation per argumentfor instance the transitive causativeinchoative verb increase is composed of an actorcausative verb combining with a oneplace predicate in the structure on the left in fig11linguistically such representations are motivated by argument alternations case linking and deep word order language acquistion scope ambiguity by the desire to represent aspects of lexical meaning and by the fact that in some languages the postulated decomposed representations are overt with each primitive predicate corresponding to a morphemefor references and recent discussion of this kind of theory see hale and keyser and kural we will sketch an understanding of the lexical representations induced by latentclass labeling in terms of the linguistic theories mentioned above aiming at an interpretation which combines computational learnability linguistic motivation and denotationalsemantic adequacythe basic idea is that latent classes are computational models of the atomic relation symbols occurring in lexicalsemantic representationsas a first implementation consider replacing the relation symbols in the first tree in fig11 with relation symbols derived from the latent class labelingin the second tree in fig 11 r17 and r8 are relation symbols with indices derived from the labeling procedure of sect4such representations can be semantically interpreted in standard ways for instance by interpreting relation symbols as denoting relations between events and individualssuch representations are semantically inadequate for reasons given in philosophical critiques of decomposed linguistic representations see fodor for recent discussiona lexicon estimated in the above way has as many primitive relations as there are latent classeswe guess there should be a few hundred classes in an approximately complete lexicon fodor arguments which are based on the very limited degree of genuine interdefinability of lexical items and on putnam arguments for contextual determination of lexical meaning indicate that the number of basic concepts has the order of magnitude of the lexicon itselfmore concretely a lexicon constructed along the above principles would identify verbs which are labelled with the same latent classes for instance it might identify the representations of grab and touchfor these reasons a semantically adequate lexicon must include additional relational constantswe meet this requirement in a simple way by including as a conjunct a unique constant derived from the openclass root as in the third tree in fig11we introduce indexing of the open class root in order that homophony of open class roots not result in common conjuncts in semantic representationsfor instance we do not want the two senses of decline exemplified in decline the proposal and decline five percent to have an common entailment represented by a common conjunctthis indexing method works as long as the labeling process produces different latent class labels for the different sensesthe last tree in fig11 is the learned representation for the scalar motion sense of the intransitive verb increasein our approach learning the argument alternation relating the transitive increase to the intransitive increase amounts to learning representations with a common component ri7 a increaserin this case this is achievedwe have proposed a procedure which maps observations of subcategorization frames with their complement fillers to structured lexical entrieswe believe the method is scientifically interesting practically useful and flexible because 1the algorithms and implementation are efficient enough to map a corpus of a hundred million words to a lexicon2the model and induction algorithm have foundations in the theory of parameterized families of probability distributions and statistical estimationas exemplified in the paper learning disambiguation and evaluation can be given simple motivated formulations3the derived lexical representations are linguistically interpretablethis suggests the possibility of largescale modeling and observational experiments bearing on questions arising in linguistic theories of the lexicon4because a simple probabilistic model is used the induced lexical entries could be incorporated in lexicalized syntaxbased probabilistic language models in particular in headlexicalized modelsthis provides for potential application in many areas5the method is applicable to any natural language where text samples of sufficient size computational morphology and a robust parser capable of extracting subcategorization frames with their fillers are available	P99-1014	inducing a semantically annotated lexicon via thembased clusteringwe present a technique for automatic induction of slot annotations for subcategorization frames based on induction of hidden classes in the them framework of statistical estimationthe models are empirically evalutated by a general decision testinduction of slot labeling for subcategorization frames is accomplished by a further application of them and applied experimentally on frame observations derived from parsing large corporawe outline an interpretation of the learned representations as theoreticallinguistic decompositional lexical entrieswe test 3000 random verbnoun pairs requiring the erbs and nouns to appear between 30 and 3000 times in trainingwe use soft clustering to form classes for generalization and do not take recourse to any handcrafter resources in our approach to selectional preference induction
automatic construction of a hypernymlabeled noun hierarchy from text previous work has shown that automatic methods can be used in building semantic lexicons this work goes a step further by automatically creating not just clusters of related words but a hierarchy of nouns and their hypernyms akin to the handbuilt hierarchy in wordnet the purpose of this work is to build something like the hypernymlabeled noun hierarchy of wordnet automatically from text using no other lexical resourceswordnet has been an important research tool but it is insufficient for domainspecific text such as that encountered in the mucs our work develops a labeled hierarchy based on a text corpusin this project nouns are clustered into a hierarchy using data on conjunctions and appositives appearing in the wall street journalthe internal nodes of the resulting tree are then labeled with hypernyms for the nouns clustered underneath them also based on data extracted from the wall street journalthe resulting hierarchy is evaluated by human judges and future research directions are discussedthe first stage in constructing our hierarchy is to build an unlabeled hierarchy of nouns using bottomup clustering methods nouns are clustered based on conjunction and appositive data collected from the wall street journal corpussome of the data comes from the parsed files 221 of the wall street journal penn treebank corpus and additional parsed text was obtained by parsing the 1987 wall street journal text using the parser described in charniak et al from this parsed text we identified all conjunctions of noun phrases and all appositives the idea here is that nouns in conjunctions or appositives tend to be semantically related as discussed in riloff and shepherd and roark and charniak taking the head words of each np and stemming them results in data for about 50000 distinct nounsa vector is created for each noun containing counts for how many times each other noun appears in a conjunction or appositive with itwe can then measure the similarity of the vectors for two nouns by computing the cosine of the angle between these vectors as 111iwr to compare the similarity of two groups of nouns we define similarity as the average of the cosines between each pair of nouns made up of one noun from each of the two groupsev cos sizesize where v ranges over all vectors for nouns in group a w ranges over the vectors for group b and size represents the number of nouns which are descendants of node xwe want to create a tree of all of the nouns in this data using standard bottomup clustering techniques as follows put each noun into its own nodecompute the similarity between each pair of nodes using the cosine methodfind the two most similar nouns and combine them by giving them a common parent we can then compute the new node similarity to each other node by computing a weighted average of the similarities between each of its children and the other nodein other words assuming nodes a and b have been combined under a new parent c the similarity between c and any other node i can be computed as once again we combine the two most similar nodes under a common parentrepeat until all nouns have been placed under a common ancestornouns which have a cosine of 0 with every other noun are not included in the final treein practice we cannot follow exactly that algorithm because maintaining a list of the cosines between every pair of nodes requires a tremendous amount of memorywith 50000 nouns we would initially require a 50000 x 50000 array of values with our current hardware the largest array we the way we handled this limitation is to process the nouns in batchesinitially 5000 nouns are read inwe cluster these until we have 2500 nodesthen 2500 more nouns are read in to bring the total to 5000 again and once again we cluster until 2500 nodes remainthis process is repeated until all nouns have been processedsince the lowestfrequency nouns are clustered based on very little information and have a greater tendency to be clustered badly we chose to filter some of these outby reducing the number of nouns to be read a much nicer structure is obtainedwe now only consider nouns with a vector of length at least 2there are approximately 20000 nouns as the leaves in our final binary tree structureour next step is to try to label each of the internal nodes with a hypernym describing its descendant nounsfollowing wordnet a word a is said to be a hypernym of a word b if native speakers of english accept the sentence quotb is a aquot to determine possible hypernyms for a particular noun we use the same parsed text described in the previous sectionas suggested in hearst we can find some hypernym data in the text by looking for conjunctions involving the word quototherquot as in quotx y and other zsquot from this phrase we can extract that z is likely a hypernym for both x and ythis data is extracted from the parsed text and for each noun we construct a vector of hypernyms with a value of 1 if a word has been seen as a hypernym for this noun and 0 otherwisethese vectors are associated with the leaves of the binary tree constructed in the previous sectionfor each internal node of the tree we construct a vector of hypernyms by adding together the vectors of its childrenwe then assign a hypernym to this node by simply choosing the hypernym with the largest value in this vector that is the hypernym which appeared with the largest number of the node descendant nounswe also list the second and thirdbest hypernyms to account for cases where a single word does not describe the cluster adequately or cases where there are a few good hypernyms which tend to alternate such as quotcountryquot and quotnationquotif a hypernym has occurred with only one of the descendant nouns it is not listed as one of the best hypernyms since we have insufficient evidence that the word could describe this class of nounsnot every node has sufficient data to be assigned a hypernymthe labeled tree constructed in the previous section tends to be extremely redundantrecall that the tree is binaryin many cases a group of nouns really do not have an inherent tree structure for example a cluster of countriesalthough it is possible that a reasonable tree structure could be created with subtrees of say european countries asian countries etc recall that we are using singleword hypernymsa large binary tree of countries would ideally have quotcountryquot as the best hypernym at every levelwe would like to combine these subtrees into a single parent labeled quotcountryquot or quotnationquot with each country appearing as a leaf directly beneath this parentanother type of redundancy can occur when an internal node is unlabeled meaning a hypernym could not be found to describe its descendant nounssince the tree root is labeled somewhere above this node there is necessarily a node labeled with a hypernym which applies to its descendant nouns including those which are a descendant of this nodewe want to move this node children directly under the nearest labeled ancestorwe compress the tree using the following very simple algorithm in depthfirst order examine the children of each internal nodeif the child is itself an internal node and it either has no best hypernym or the same three best hypernyms as its parent delete this child and make its children into children of the parent insteadthere are 20014 leaves and 654 internal nodes in the final tree the toplevel node in our learned tree is labeled quotproductanalystofficialquot since these hypernyms are learned from the wall street journal they are domainspecific labels rather than the more general quotthingpersonquothowever if the hierarchy were to be used for text from the financial domain these labels may be preferredthe next level of the hierarchy the children of the root is as shown in table 1these numbers do not add up to 20014 because 1288 nouns are attached directly to the root meaning that they could not be clustered to any greater level of detailthese tend to be nouns for which little data was available generally proper nouns to evaluate the hierarchy 10 internal nodes dominating at least 20 nouns were selected at randomfor each of these nodes we randomly selected 20 of the nouns from the cluster under that nodethree human judges were asked to evaluate for each noun and each of the three hypernyms listed as quotbestquot for that cluster whether they were actually in a hyponymhypernym relationthe judges were students working in natural language processing or computational linguistics at our institution who were not directly involved in the research for this project5 quotnoisequot nouns randomly selected from elsewhere in the tree were also added to each cluster without the judges knowledge to verify that the judges were not overly generoussome nouns especially proper nouns were not recognized by the judgesfor any noun that was not evaluated by at least two judges we evaluated the nounhypernym pair by examining the appearances of that noun in the source text and verifying that the hypernym was correct for the predominant sense of the nountable 2 presents the results of this evaluationthe table lists only results for the actual candidate hyponym nouns not the noise wordsthe quothypernym 1quot column indicates whether the quotbestquot hypernym was considered correct while the quotany hypernymquot column indicates whether any of the listed hypernyms were acceptedwithin those columns quotmajorityquot lists the opinion of the majority of judges and quotanyquot indicates the hypernyms that were accepted by even one of the judgesthe quothypernym 1anyquot column can be used to compare results to riloff and shepherd for five handselected categories each with a single hypernym and the 20 nouns their algorithm scored as the best members of each category at least one judge marked on average about 31 of the nouns as correctusing randomlyselected categories and randomlyselected category members we achieved 39by the strictest criteria our algorithm produces correct hyponyms for a randomlyselected hypernym 33 of the timeroark and charniak report that for a handselected category their algorithm generally produces 20 to 40 correct entriesfurthermore if we loosen our criteria to consider also the second and thirdbest hypernyms 60 of the nouns evaluated were assigned to at least one correct hypernym according to at least one judgethe quotbankfirmstationquot cluster consists largely of investment firms which were marked as incorrect for quotbankquot resulting in the poor performance on the hypernym 1 measures for this clusterthe last cluster in the list labeled quotcompanyquot is actually a very good cluster of cities that because of sparse data was assigned a poor hypernymsome of the suggestions in the following section might correct this problemof the 50 noise words a few of them were actually rated as correct as well as shown in table 3this is largely because the noise words were selected truly at random so that a noise word for the quotcompanyquot cluster may not have been in that particular cluster but may still have appeared under a quotcompanyquot hypernym elsewhere in the hierarchyfuture work should benefit greatly by using data on the hypernyms of hypernymsin our current tree the best hypernym for the entire tree is quotproductquot however many times nodes deeper in the tree are given this label alsofor example we have a cluster including many forms of currency but because there is little data for these particular words the only hypernym found was quotproductquothowever the parent of this node has the best hypernym of quotcurrencyquotif we knew that quotproductquot was a hypernym of quotcurrencyquot we could detect that the parent node label is more specific and simply absorb the child node into the parentfurthermore we may be able to use data on the hypernyms of hypernyms to give better labels to some nodes that are currently labeled simply with the best hypernyms of their subtrees such as a node labeled quotproductanalystquot which has two subtrees one labeled quotproductquot and containing words for things the other labeled quotanalystquot and containing names of peoplewe would like to instead label this node something like quotentityquotit is not yet clear whether corpus data will provide sufficient data for hypernyms at such a high level of the tree but depending on the intended application for the hierarchy this level of generality might not be requiredas noted in the previous section one major spurious result is a cluster of 51 nouns mainly people which is given the hypernym quotconductorquotthe reason for this is that few of the nouns appear with hypernyms and two of them appear in the same phrase listing conductors thus giving quotconductorquot a count of two sufficient to be listed as the only hypernym for the clusterit might be useful to have some stricter criterion for hypernyms say that they occur with a certain percentage of the nouns below them in the treeadditional hypernym data would also be helpful in this case and should be easily obtainable by looking for other patterns in the text as suggested by hearst because the tree is built in a binary fashion when eg three clusters should all be distinct children of a common parent two of them must merge first giving an artificial intermediate level in the treefor example in the current tree a cluster with best hypernym quotagencyquot and one with best hypernym quotexchangequot have a parent with two best hypernyms quotagencyexchangequot rather than both of these nodes simply being attached to the next level up with best hypernym quotgroupquotit might be possible to correct for this situation by comparing the hypernyms for the two clusters and if there is little overlap deleting their parent node and attaching them to their grandparent insteadit would be useful to try to identify terms made up of multiple words rather than just using the head nouns of the noun phrasesnot only would this provide a more useful hierarchy or at least perhaps one that is more useful for certain applications but it would also help to prevent some errorshearst gives an example of a potential hyponymhypernym pair quotbroken boneinjuryquot using our algorithm we would learn that quotinjuryquot is a hypernym of quotbonequotideally this would not appear in our hierarchy since a more common hypernym would be chosen instead but it is possible that in some cases a bad hypernym would be found based on multiple word phrasesa discussion of the difficulties in deciding how much of a noun phrase to use can be found in hearstideally a useful hierarchy should allow for multiple senses of a word and this is an area which can be explored in future workhowever domainspecific text tends to greatly constrain which senses of a word will appear and if the learned hierarchy is intended for use with the same type of text from which it was learned it is possible that this would be of limited benefitwe used parsed text for these experiments because we believed we would get better results and the parsed data was readily availablehowever it would be interesting to see if parsing is necessary or if we can get equivalent or nearlyequivalent results doing some simpler text processing as suggested in ahlswede and evens both hearst and riloff and shepherd use unparsed textpereira et al used clustering to build an unlabeled hierarchy of nounstheir hierarchy is constructed topdown rather than bottomup with nouns being allowed membership in multiple clusterstheir clustering is based on verbobject relations rather than on the nounnoun relations that we usefuture work on our project will include an attempt to incorporate verbobject data as well in the clustering processthe tree they construct is also binary with some internal nodes which seem to be quotartificialquot but for evaluation purposes they disregard the tree structure and consider only the leaf nodesunfortunately it is difficult to compare their results to ours since their evaluation is based on the verbobject relationsriloff and shepherd suggested using conjunction and appositive data to cluster nouns however they approximated this data by just looking at the nearest np on each side of a particular nproark and charniak built on that work by actually using conjunction and appositive data for noun clustering as we do hereboth of these projects have the goal of building a single cluster of eg vehicles and both use seed words to initialize a cluster with nouns belonging to ithearst introduced the idea of learning hypernymhyponym relationships from text and gives several examples of patterns that can be used to detect these relationships including those used here along with an algorithm for identifying new patternsthis work shares with ours the feature that it does not need large amounts of data to learn a hypernym unlike in much statistical work a single occurrence is sufficientthe hyponymhypernym pairs found by hearst algorithm include some that hearst describes as quotcontext and pointofview dependentquot such as quotwashingtonnationalistquot and quotaircrafttargetquot our work is somewhat less sensitive to this kind of problem since only the most common hypernym of an entire cluster of nouns is reported so much of the noise is filteredwe have shown that hypernym hierarchies of nouns can be constructed automatically from text with similar performance to semantic lexicons built automatically for handselected hypernymswith the addition of some improvements we have identified we believe that these automatic methods can be used to construct truly useful hierarchiessince the hierarchy is learned from sample text it could be trained on domainspecific text to create a hierarchy that is more applicable to a particular domain than a generalpurpose resource such as wordnetthanks to eugene charniak for helpful discussions and for the data used in this projectthanks also to brian roark heidi jfox and keith hall for acting as judges in the project evaluationthis research is supported in part by nsf grant iri9319516 and by onr grant n00149610549	P99-1016	automatic construction of a hypernymlabeled noun hierarchy from textprevious work has shown that automatic methods can be used in building semantic lexiconsthis work goes a step further by automatically creating not just clusters of related words but a hierarchy of nouns and their hypernyms akin to the handbuilt hierarchy in wordnetwe let three judges evaluate ten internal nodes in the hyponym hierarchy that had at least twenty descendants
development and use of a goldstandard data set for subjectivity classifications this paper presents a case study of analyzing and improving intercoder reliability in discourse using statistical techniques corrected tags are formulated and successfully used to guide a revision of the coding manual and develop an automatic classifier this paper presents a case study of analyzing and improving intercoder reliability in discourse tagging using the statistical techniques presented in our approach is data driven we refine our understanding and presentation of the classification scheme guided by the results of the intercoder analysiswe also present the results of a probabilistic classifier developed on the resulting annotationsmuch research in discourse processing has focused on taskoriented and instructional dialogsthe task addressed here comes to the fore in other genres especially news reportingthe task is to distinguish sentences used to objectively present factual information from sentences used to present opinions and evaluationsthere are many applications for which this distinction promises to be important including text categorization and summarizationthis research takes a large step toward developing a reliably annotated gold standard to support experimenting with such applicationsthis research is also a case study of analyzing and improving manual tagging that is applicable to any tagging taskwe perform a statistical analysis that provides information that complements the information provided by cohen kappa in particular we analyze patterns of agreement to identify systematic disagreements that result from relative bias among judges because they can potentially be corrected automaticallythe corrected tags serve two purposes in this workthey are used to guide the revision of the coding manual resulting in improved kappa scores and they serve as a gold standard for developing a probabilistic classifierusing biascorrected tags as goldstandard tags is one way to define a single best tag when there are multiple judges who disagreethe coding manual and data from our experiments are available at hap wwwcsnmsuedur wiebeprojectsin the remainder of this paper we describe the classification being performed the statistical tools used to analyze the data and produce the biascorrected tags the case study of improving intercoder agreement and the results of the classifier for automatic subjectivity tagging we address evidentiality in text which concerns issues such as what is the source of information and whether information is being presented as fact or opinionthese questions are particularly important in news reporting in which segments presenting opinions and verbal reactions are mixed with segments presenting objective fact the definitions of the categories in our coding manual are intentionbased quotif the primary intention of a sentence is objective presentation of material that is factual to the reporter the sentence is objectiveotherwise the sentence is subjectivequot we focus on sentences about private states such as belief knowledge emotions etc and sentences about speech events such as speaking and writingsuch sentences may be either subjective or objectivefrom the coding manual quotsubjective speechevent sentences are used to communicate the speaker evaluations opinions emotions and speculationsthe primary intention of objective speechevent sentences on the other hand is to objectively communicate material that is factual to the reporterthe speaker in these cases is being used as a reliable source of informationquot following are examples of subjective and objective sentences in sentence 4 there is no uncertainty or evaluation expressed toward the speaking eventthus from one point of view one might have considered this sentence to be objectivehowever the object of the sentence is not presented as material that is factual to the reporter so the sentence is classified as subjectivelinguistic categorizations usually do not cover all instances perfectlyfor example sentences may fall on the borderline between two categoriesto allow for uncertainty in the annotation process the specific tags used in this work include certainty ratings ranging from 0 for least certain to 3 for most certainas discussed below in section 32 the certainty ratings allow us to investigate whether a model positing additional categories provides a better description of the judges annotations than a binary model doessubjective and objective categories are potentially important for many text processing applications such as information extraction and information retrieval where the evidential status of information is importantin generation and machine translation it is desirable to generate text that is appropriately subjective or objective in summarization subjectivity judgments could be included in document profiles to augment automatically produced document summaries and to help the user make relevance judgments when using a search enginein addition they would be useful in text categorizationin related work we found that article types such as announcement and opinion piece are significantly correlated with the subjective and objective classificationour subjective category is related to but differs from the statementopinion category of the switchboarddamsl discourse annotation project as well as the gives opinion category of bale model of smallgroup interactionall involve expressions of opinion but while our category specifications focus on evidentiality in text theirs focus on how conversational participants interact with one another in dialogtable 1 presents data for two judgesthe rows correspond to the tags assigned by judge 1 and the columns correspond to the tags assigned by judge 2let nj denote the number of sentences that judge 1 classifies as i and judge 2 classifies as j and let be the probability that a randomly selected sentence is categorized as i by judge 1 and j by judge 2then the maximum likelihood estimate of fiii is 11771 where n eii nii 504table 1 shows a fourcategory data configuration in which certainty ratings 0 and 1 are combined and ratings 2 and 3 are combinednote that the analyses described in this section cannot be performed on the twocategory data configuration due to insufficient degrees of freedom evidence of confusion among the classifications in table 1 can be found in the marginal totals ni and njwe see that judge 1 has a relative preference or bias for objective while judge 2 has a bias for subjectiverelative bias is one aspect of agreement among judgesa second is whether the judges disagreements are systematic that is correlatedone pattern of systematic disagreement is symmetric disagreementwhen disagreement is symmetric the differences between the actual counts and the counts expected if the judges decisions were not correlated are symmetric that is snii for i j where 5i is the difference from independenceour goal is to correct correlated disagreements automaticallywe are particularly interested in systematic disagreements resulting from relative biaswe test for evidence of such correlations by fitting probability models to the dataspecifically we study bias using the model for marginal homogeneity and symmetric disagreement using the model for quasisymmetrywhen there is such evidence we propose using the latent class model to correct the disagreements this model posits an unobserved variable to explain the correlations among the judges observationsthe remainder of this section describes these models in more detailall models can be evaluated using the freeware package coco which was developed by badsberg and is available at http webmathaucdkr jhbcocoa probability model enforces constraints on the counts in the datathe degree to which the counts in the data conform to the constraints is called the fit of the modelin this work model fit is reported in terms of the likelihood ratio statistic g2 and its significance the higher the g2 value the poorer the fitwe will consider model fit to be acceptable if its reference significance level is greater than 001 bias of one judge relative to another is evidenced as a discrepancy between the marginal totals for the two judges bias is measured by testing the fit of the model for marginal homogeneity 25i for all ithe larger the g2 value the greater the biasthe fit of the model can be evaluated as described on pages 293294 of bishop et al judges who show a relative bias do not always agree but their judgments may still be correlatedas an extreme example judge 1 may assign the subjective tag whenever judge 2 assigns the objective tagin this example there is a kind of symmetry in the judges responses but their agreement would be lowpatterns of symmetric disagreement can be identified using the model for quasisymmetrythis model constrains the offdiagonal counts ie the counts that correspond to disagreementit states that these counts are the product of a table for independence and a symmetric table nii ai x ai x aii such that aij iiin this formula ai x a3 is the model for independence and ai3 is the symmetric interaction termintuitively aii represents the difference between the actual counts and those predicted by independencethis model can be evaluated using coco as described on pages 289290 of bishop et al we use the latent class model to correct symmetric disagreements that appear to result from biasthe latent class model was first introduced by lazarsfeld and was later made computationally efficient by goodman goodman procedure is a specialization of the them algorithm which is implemented in the freeware program coco since its development the latent class model has been widely applied and is the underlying model in various unsupervised machine learning algorithms including autoclass the form of the latent class model is that of naive bayes the observed variables are all conditionally independent of one another given the value of the latent variablethe latent variable represents the true state of the object and is the source of the correlations among the observed variablesas applied here the observed variables are the classifications assigned by the judgeslet b d j and m be these variables and let l be the latent variablethen the latent class model is the parameters of the model are p p p pp once estimates of these parameters are obtained each clause can be assigned the most probable latent category given the tags assigned by the judgesthe them algorithm takes as input the number of latent categories hypothesized ie the number of values of l and produces estimates of the parametersfor a description of this process see goodman dawid skene or pedersen bruce three versions of the latent class model are considered in this study one with two latent categories one with three latent categories and one with fourwe apply these models to three data configurations one with two categories one with four categories and one with eight categories all combinations of model and data configuration are evaluated except the fourcategory latent class model with the twocategory data configuration due to insufficient degrees of freedomin all cases the models fit the data well as measured by g2the model chosen as final is the one for which the agreement among the latent categories assigned to the three data configurations is highest that is the model that is most consistent across the three data configurationsour annotation project consists of the following steps2 biascorrected tag in many cases but arguing for his or her own tag in some casesbased on the judges feedback 22 of the 504 biascorrected tags are changed and a second draft of the coding manual is written5a second corpus is annotated by the same four judges according to the new coding manualeach spends about five hours6the results of the second tagging experiment are analyzed using the methods described in section 3 and biascorrected tags are produced for the second data settwo disjoint corpora are used in steps 2 and 5 both consisting of complete articles taken from the wall street journal treebank corpus in both corpora judges assign tags to each noncompound sentence and to each conjunct of each compound sentence 504 in the first corpus and 500 in the secondthe segmentation of compound sentences was performed manually before the judges received the datajudges j and b the first two authors of this paper are nlp researchersjudge m is an undergraduate computer science student and judge d has no background in computer science or linguisticsjudge j with help from m developed the original coding instructions and judge j directed the process in step 4the analysis performed in step 3 reveals strong evidence of relative bias among the judgeseach pairwise comparison of judges also shows a strong pattern of symmetric disagreementthe twocategory latent class model produces the most consistent clusters across the data configurationsit therefore is used to define the biascorrected tagsin step 4 judge b was excluded from the interactive discussion for logistical reasonsdiscussion is apparently important because although b kappa values for the first study are on par with the others b kappa values for agreement with the other judges change very little from the first to the second study in contrast agreement among the other judges noticeably improvesbecause judge b poor performance in the second tagging experiment is linked to a difference in procedure judge b tags are excluded from our subsequent analysis of the data gathered during the second tagging experimenttable 2 shows the changes from study 1 to study 2 in the kappa values for pairwise agreement among the judgesthe best results are clearly for the two who are not authors of this paper the kappa value for the agreement between d and m considering all certainty ratings reaches 76 which allows tentative conclusions on krippendorf scale if we exclude the sentences with certainty rating 0 the kappa values for pairwise agreement between m and d and between j and m are both over 8 which allows definite conclusions on krippendorf scalefinally if we only consider sentences with certainty 2 or 3 the pairwise agreements among m d and j all have high kappa values 087 and overwe are aware of only one previous project reporting intercoder agreement results for similar categories the switchboarddamsl project mentioned abovewhile their kappa results are very good for other tags the opinionstatement tagging was not very successful quotthe distinction was very hard to make by labelers and accounted for a large proportion of our interlabeler errorquot in step 6 as in step 3 there is strong evidence of relative bias among judges d j and m each pairwise comparison of judges also shows a strong pattern of symmetric disagreementthe results of this analysis are presented in table 33 also as in step 3 the twocategory latent class model produces the most consistent clusters across the data configurationsthus it is used to define the biascorrected tags for the second data set as wellrecently there have been many successful applications of machine learning to discourse processing such as in this section we report the results of machine learning experiments in which we develop probablistic classifiers to automatically perform the subjective and objective classificationin the method we use for developing classifiers a search is performed to find a probability model that captures important interdependencies among featuresbecause features can be dropped and added during search the method also performs feature selectionin these experiments the system considers naive bayes full independence full interdependence and models generated from those using forward and backward searchthe model selected is the one with the highest accuracy on a heldout portion of the training data setson each fold one set is used for testing and the other nine are used for trainingfeature selection model selection and parameter estimation are performed anew on each foldthe following are the potential features considered on each folda binary feature is included for each of the following the presence in the sentence of a pronoun an adjective a cardinal number a modal other than will and an adverb other than notwe also include a binary feature representing whether or not the sentence begins a new paragraphfinally a feature is included representing cooccurrence of word tokens and punctuation marks with the subjective and objective classification4 there are many other features to investigate in future work such as features based on tags assigned to previous utterances and features based on semantic classes such as positive and negative polarity adjectives and reporting verbs the data consists of the concatenation of the two corpora annotated with biascorrected tags as described abovethe baseline accuracy ie the frequency of the more frequent class is only 51the results of the experiments are very promisingthe average accuracy across all folds is 7217 more than 20 percentage points higher than the baseline accuracyinterestingly the system performs better on the sentences for which the judges are certainin a post hoc analysis we consider the sentences from the second data set for which judges m j and d rate their certainty as 2 or 3there are 299500 such sentencesfor each fold we calculate the system accuracy on the subset of the test set consisting of such sentencesthe average accuracy of the subsets across folds is 815taking human performance as an upper bound the system has room for improvementthe average pairwise percentage agreement between d j and m and the biascorrected tags in the entire data set is 895 while the system percentage agreement with the biascorrected tags is 7217this paper demonstrates a procedure for automatically formulating a single best tag when there are multiple judges who disagreethe procedure is applicable to any tagging task in which the judges exhibit symmetric disagreement resulting from biaswe successfully use biascorrected tags for two purposes to guide a revision of the coding manual and to develop an automatic classifierthe revision of the coding manual results in as much as a 16 point improvement in pairwise kappa values and raises the average agreement among the judges to a kappa value of over 087 for the sentences that can be tagged with certaintyusing only simple features the classifier achieves an average accuracy 21 percentage points higher than the baseline in 10fold cross validation experimentsin addition the average accuracy of the classifier is 815 on the sentences the judges tagged with certaintythe strong performance of the classifier and its consistency with the judges demonstrate the value of this approach to developing goldstandard tagsthis research was supported in part by the office of naval research under grant number n000149510776we are grateful to matthew t bell and richard a wiebe for participating in the annotation study and to the anonymous reviewers for their comments and suggestions	P99-1032	development and use of a goldstandard data set for subjectivity classificationsthis paper presents a case study of analyzing and improving intercoder reliability in discourse tagging using statistical techniquesbiascorrected tags are formulated and successfully used to guide a revision of the coding manual and develop an automatic classifierwe use a sentencelevel naive bayes classifier using as features the presence or absence of particular syntactic classes punctuation and sentence positionwe define a subjective sentences as sentences expressing evaluations opinions emotions and speculations
automatic identification of noncompositional phrases noncompositional expressions present a special challenge to nlp applications we present a method for automatic identification of noncompositional expressions using their statistical properties in a text corpus our method is based on the hypothesis that when a phrase is noncomposition its mutual information differs significantly from the mutual informations of phrases obtained by substituting one of the word in the phrase with a similar word noncompositional expressions present a special challenge to nlp applicationsin machine translation wordforword translation of noncompositional expressions can result in very misleading translationsin information retrieval expansion of words in a noncompositional expression can lead to dramatic decrease in precision without any gain in recallless obviously noncompositional expressions need to be treated differently than other phrases in many statistical or corpusbased nlp methodsfor example an underlying assumption in some word sense disambiguation systems eg is that if two words occurred in the same context they are probably similarsuppose we want to determine the intended meaning of quotproductquot in quothot productquotwe can find other words that are also modified by quothotquot and then choose the meaning of quotproductquot that is most similar to meanings of these wordshowever this method fails when noncompositional expressions are involvedfor instance using the same algorithm to determine the meaning of quotlinequot in quothot linequot the words quotproductquot quotmerchandisequot quotcarquot etc would lead the algorithm to choose the quotline of productquot sense of quotlinequotwe present a method for automatic identification of noncompositional expressions using their statistical properties in a text corpusthe intuitive idea behind the method is that the metaphorical usage of a noncompositional expression causes it to have a different distributional characteristic than expressions that are similar to its literal meaningthe input to our algorithm is a collocation database and a thesauruswe briefly describe the process of obtaining this inputmore details about the construction of the collocation database and the thesaurus can be found in we parsed a 125million word newspaper corpus with minipar1 a descendent of principar and extracted dependency relationships from the parsed corpusa dependency relationship is a triple where head and modifier are words in the input sentence and type is the type of the dependency relationfor example is an example dependency tree and the set of dependency triples extracted from are shown in there are about 80 million dependency relationships in the parsed corpusthe frequency counts of dependency relationships are filtered with the loglikelihood ratio we call a dependency relationship a collocation if its loglikelihood ratio is greater than a threshold the number of unique collocations in the resulting database2 is about 11 millionusing the similarity measure proposed in we constructed a corpusbased thesaurus3 consisting of 11839 nouns 3639 verbs and 5658 adjectiveadverbs which occurred in the corpus at least 100 timeswe define the probability space to consist of all possible collocation tripleswe use is ft 14i to denote the frequency count of all the collocations that match the pattern where h and m are either words or the wild card and r is either a dependency type or the wild cardfor example to compute the mutual information in a collocation we treat a collocation as the conjunction of three events the mutual information of a collocation is the logarithm of the ratio between the probability of the collocation and the probability of events a b and c cooccur if we assume b and c are conditionally independent given a d type modifierlxl type l ogin this section we use several examples to demonstrate the basic idea behind our algorithmconsider the expression quotspill gutquotusing the automatically constructed thesaurus we find the following top10 most similar words to the verb quotspillquot and the noun quotgutquot spill leak 0153 pour 0127 spew 0125 dump 0118 pump 0098 seep 0096 burn 0095 explode 0094 burst 0092 spray 0091 gut intestine 0091 instinct 0089 foresight 0085 creativity 0082 heart 0079 imagination 0076 stamina 0074 soul 0073 liking 0073 charisma 0071 the collocation quotspill gutquot occurred 13 times in the 125millionword corpusthe mutual information of this collocation is 624searching the collocation database we find that it does not contain any collocation in the form nor where simvsptil is a verb similar to quotspillquot and sirnngut is a noun similar to quotgutquotthis means that the phrases such as quotleak gutquot quotpour gutquot or quotspill intestinequot quotspill instinctquot either did not appear in the corpus at all or did not occur frequent enough to pass the loglikelihood ratio testthe second example is quotred tapequotthe top10 most similar words to quotredquot and quottapequot in our thesaurus are red yellow 0164 purple 0149 pink 0146 green 0136 blue 0125 white 0122 color 0118 orange 0111 brown 0101 shade 0094 tape videotape 0196 cassette 0177 videocassette 0168 video 0151 disk 0129 recording 0117 disc 0113 footage 0111 recorder 0106 audio 0106 the following table shows the frequency and mutual information of quotred tapequot and word combinations in which one of quotredquot or quottapequot is substituted by a similar word even though many other similar combinations exist in the collocation database they have very different frequency counts and mutual information values than quotred tapequotfinally consider a compositional phrase quoteconomic impactquotthe top10 most similar words are economic financial 0305 political 0243 social 0219 fiscal 0209 cultural 0202 budgetary 02 technological 0196 organizational 019 ecological 0189 monetary 0189 impact effect 0227 implication 0163 consequence 0156 significance 0146 repercussion 0141 fallout 0141 potential 0137 ramification 0129 risk 0126 influence 0125 the frequency counts and mutual information values of quoteconomic impactquot and phrases obtained by replacing one of quoteconomicquot and quotimpactquot with a similar word are in table 4not only many combinations are found in the corpus many of them have very similar mutual information values to that of nomial distribution can be accurately approximated by a normal distribution since all the potential noncompositional expressions that we are considering have reasonably large frequency counts we assume their distributions are normallet head type modifierj k and 1 1 n the maximum likelihood estimation of the true probability p of the collocation is k n even though we do not know what p is since p is normally distributed there is n chance that it fails within the interval where zn is a constant related to the confidence level n and the last step in the above derivation is due to the fact that is very smalltable 3 shows the zn values for a sample set of confidence intervalsquoteconomic impactquotin fact the difference of mutual information values appear to be more important to the phrasal similarity than the similarity of individual wordsfor example the phrases quoteconomic falloutquot and quoteconomic repercussionquot are intuitively more similar to quoteconomic impactquot than quoteconomic implicationquot or quoteconomic significancequot even though quotimplicationquot and quotsignificancequot have higher similarity values to quotimpactquot than quotfalloutquot and quotrepercussionquot dothese examples suggest that one possible way to separate compositional phrases and noncompositional ones is to check the existence and mutual information values of phrases obtained by substituting one of the words with a similar worda phrase is probably noncompositional if such substitutions are not found in the collocation database or their mutual information values are significantly different from that of the phrasein order to implement the idea of separating noncompositional phrases from compositional ones with mutual information we must use a criterion to determine whether or not the mutual information values of two collocations are significantly differentalthough one could simply use a predetermined threshold for this purpose the threshold value will be totally arbitraryfurthermore such a threshold does not take into account the fact that with different frequency counts we have different levels confidence in the mutual information valueswe propose a more principled approachthe frequency count of a collocation is a random variable with binomial distributionwhen the frequency count is reasonably large a bin 50 80 90 95 98 99 zn 067 128 164 196 233 258 we further assume that the estimations of p p and p in are accuratethe confidence interval for the true probability gives rise to a confidence interval for the true mutual information the upper and lower bounds of this interval are obtained by substituting 11 with kzs vk and lz1114 in since our confidence of p falling between1vin is n we can have n confidence that the true mutual information is within the upper and lower boundwe use the following condition to determine whether or not a collocation is compositional a collocation a is noncompositional if there does not exist another collocation 3 such that 13 is obtained by substituting the head or the modifier in a with a similar word and there is an overlap between the 95 confidence interval of the mutual information values of a and 0for example the following table shows the frequency count mutual information and the lower and upper bounds of the 95 confidence interval of the true mutual information verbobject freq mutual lower upper count info bound bound make difference 1489 2928 2876 2978 make change 1779 2194 2146 2239 since the intervals are disjoint the two collocations are considered to have significantly different mutual information valuesthere is not yet a wellestablished methodology for evaluating automatically acquired lexical knowledgeone possibility is to compare the automatically identified relationships with relationships listed in a manually compiled dictionaryfor example compared automatically created thesaurus with the wordnet 1990 and roget thesaurushowever since the lexicon used in our parser is based on the wordnet the phrasal words in wordnet are treated as a single wordfor example quottake advantage ofquot is treated as a transitive verb by the parseras a result the extracted noncompositional phrases do not usually overlap with phrasal entries in the wordnettherefore we conducted the evaluation by manually examining sample resultsthis method was also used to evaluate automatically identified hyponyms word similarity and translations of collocations our evaluation sample consists of 5 most frequent open class words in the our parsed corpus have company make do take and 5 words whose frequencies are ranked from 2000 to 2004 path lock resort column gulfwe examined three types of dependency relationships objectverb nounnoun and adjectivenouna total of 216 collocations were extracted shown in appendix awe compared the collocations in appendix a with the entries for the above 10 words in the ntc english idioms dictionary which contains approximately 6000 definitions of idiomsfor our evaluation purposes we selected the idioms in ntceid that satisfy both of the following two conditions a the head word of the idiom is one of the above 10 words b there is a verbobject nounnoun or adjectivenoun relationship in the idiom and the modifier in the phrase is not a variablefor example quottake a stab at somethingquot is included in the evaluation whereas quottake something at face valuequot is notthere are 249 such idioms in ntceid 34 of which are also found in appendix a if we treat the 249 entries in ntceid as the gold standard the precision and recall of the phrases in appendix a are shown in table 4to compare the performance with manually compiled dictionaries we also compute the precision and recall of the entries in the longman dictionary of english idioms that satisfy the two conditions in it can be seen that the overlap between manually compiled dictionaries are quite low reflecting the fact that different lexicographers may have quite different opinion about which phrases are noncompositionalthe collocations in appendix a are classified into three categoriesthe ones marked with sign are found in ntceidthe ones marked withx are parsing errors the unmarked collocations satisfy the condition but are not found in ntceidmany of the unmarked collocation are clearly idioms such as quottake fifth amendmentquot and quottake tollquot suggesting that even the most comprehensive dictionaries may have many gaps in their coveragethe method proposed in this paper can be used to improve the coverage manually created lexical resourcesmost of the parser errors are due to the incompleteness of the lexicon used by the parserfor example quotoptquot is not listed in the lexicon as a verbthe lexical analyzer guessed it as a noun causing the erroneous collocation quot do optquotthe collocation quottrig lockquot should be quottrigger lockquotthe lexical analyzer in the parser analyzed quottriggerquot as the er form of the adjective quottrigquot duplications in the corpus can amplify the effect of a single mistakefor example the following disclaimer occurred 212 times in the corpusquotannualized average rate of return after expenses for the past 30 days not a forecast of future returnsquot the parser analyzed quota forecast of future returnsquot as s ni a forecast of future vp returnsas a result satisfied the condition duplications can also skew the mutual information of correct dependency relationshipsfor example the verbobject relationship between quottakequot and quotbridequot passed the mutual information filter because there are 4 copies of the article containing this phraseif we were able to throw away the duplicates and record only one count of quottakebridequot it would have not pass the mutual information filter the fact that systematic parser errors tend to pass the mutual information filter is both a curse and a blessingon the negative side there is no obvious way to separate the parser errors from true noncompositional expressionson the positive side the output of the mutual information filter has much higher concentration of parser errors than the database that contains millions of collocationsby manually sifting through the output one can construct a list of frequent parser errors which can then be incorporated into the parser so that it can avoid making these mistakes in the futuremanually going through the output is not unreasonable because each noncompositional expression has to be individually dealt with in a lexicon anywayto find out the benefit of using the dependency relationships identified by a parser instead of simple cooccurrence relationships between words we also created a database of the cooccurrence relationship between partofspeech tagged wordswe aggregated all word pairs that occurred within a 4word window of each otherthe same algorithm and similarity measure for the dependency database are used to construct a thesaurus using the cooccurrence databaseappendix b shows all the word pairs that satisfies the condition and that involve one of the 10 words have company make do take path lock resort column gulfit is clear that appendix b contains far fewer true noncompositional phrases than appendix athere have been numerous previous research on extracting collocations from corpus eg and they do not however make a distinction between compositional and noncompositional collocationsmutual information has often been used to separate systematic associations from accidental onesit was also used to compute the distributional similarity between words a method to determine the compositionality of verbobject pairs is proposed in the basic idea in there is that quotif an object appears only with one verb in a large corpus we expect that it has an idiomatic naturequot for each object noun o computes the distributed frequency df and rank the noncompositionality of o according to this valueusing the notation introduced in section 3 df is computed as follows where vi v2 vn are verbs in the corpus that took o as the object and where a and b are constantsthe first column in table 5 lists the top 40 verbobject pairs in the quotmiquot column show the result of our mutual information filterthe sign means that the verbobject pair is also consider to be noncompositional according to mutual information filter the sign means that the verbobject pair is present in our dependency database but it does not satisfy condition for each marked pairs the quotsimilar collocationquot column provides a similar collocation with a similar mutual information value the of marked pairs are not found in our collocation database for various reasonsfor example quotfinish seventhquot is not found because quotseventhquot is normalized as quotnumquot quothave a goquot is not found because quota goquot is not an entry in our lexicon and quottake advantagequot is not found because quottake advantage ofquot is treated as a single lexical item by our parserthe v marks in the quotntcquot column in table 5 indicate that the corresponding verbobject pairs is an idiom in it can be seen that none of the verbobject pairs in table 5 that are filtered out by condition is listed as an idiom in ntceidwe have presented a method to identify noncompositional phrasesthe method is based on the assumption that noncompositional phrases have a significantly different mutual information value than the phrases that are similar to their literal meaningsour experiment shows that this hypothesis is generally truehowever many collocations resulted from systematic parser errors also tend to posses this propertythe author wishes to thank acl reviewers for their helpful comments and suggestionsthis research was partly supported by natural sciences and engineering research council of canada grant ogp121338y chouelca1988looking for needles in a haystack or locating interesting collocational expressions in large textual databasesin proceedings of the riao conference on useroriented contentbased text and image handling cambridge ma march 2124ido dagan and alon itai1994word sense disambiguation using a second language monolingual corpuscomputational linguistics 20563596ted dunning1993accurate methods for the statistics of surprise and coincidencecomputational linguistics 196174 marchmarti a hearst1998automated discovery of wordnet relationsin c fellbaum editor wordnet an electronic lexical database pages 131151mit press	P99-1041	automatic identification of noncompositional phrasesnoncompositional expressions present a special challenge to nlp applicationswe present a method for automatic identification of noncompositional expressions using their statistical properties in a text corpusour method is based on the hypothesis that when a phrase is noncomposition its mutual information differs significantly from the mutual informations of phrases obtained by substituting one of the word in the phrase with a similar wordwe use lsa to distinguish between compositional and non compositional verbparticle constructions and noun noun compoundswe define a decision criterion for non compositional phrases based on the change in the mutual information of a phrase when substituting one word for a similar one based on an automatically constructed thesaurus
deep read a reading comprehension system paper describes initial work on read an automated reading comprehension system that accepts arbitrary text input and answers questions about it we have acquired a corpus of 60 and 60 test stories of to grade material each story is followed by shortanswer questions we used these to construct and evaluate a baseline system that uses pattern matching techniques augmented with additional automated linguistic processing this simple system retrieves the sentence containing the answer 3040 of the time this paper describes our initial work exploring reading comprehension tests as a research problem and an evaluation method for language understanding systemssuch tests can take the form of standardized multiplechoice diagnostic reading skill tests as well as fillintheblank and shortanswer teststypically such tests ask the student to read a story or article and to demonstrate herhis understanding of that article by answering questions about itfor an example see figure 1reading comprehension tests are interesting because they constitute quotfoundquot test material these tests are created in order to evaluate children reading skills and therefore test materials scoring algorithms and human performance measures already existfurthermore human performance measures provide a more intuitive way of assessing the capabilities of a given system than current measures of precision recall fmeasure operating curves etcin addition reading comprehension tests are written to test a range of skill levelswith proper choice of test material it should be possible to challenge systems to successively higher levels of performancefor these reasons reading comprehension tests offer an interesting alternative to the kinds of specialpurpose carefully constructed evaluations that have driven much recent research in language understandingmoreover the current stateoftheart in computerbased language understanding makes this project a good choice it is beyond current systems capabilities but tractableour it was 150 years ago this year that our nation biggest library burned to the groundcopies of all the written books of the time were kept in the library of congressbut they were destroyed by fire in 1814 during a war with the britishthat fire did not stop book loversthe next year they began to rebuild the libraryby giving it 6457 of his books thomas jefferson helped get it startedthe first libraries in the united states could be used by members onlybut the library of congress was built for all the peoplefrom the start it was our national librarytoday the library of congress is one of the largest libraries in the worldpeople can find a copy of just about every book and magazine printedlibraries have been with us since people first learned to writeone of the oldest to be found dates back to about 800 years bcthe books were written on tablets made from claythe people who took care of the books were called quotmen of the written tabletsquot simple bagofwords approach picked an appropriate sentence 3040 of the time with only a few months work much of it devoted to infrastructurewe believe that by adding additional linguistic and world knowledge sources to the system it can quickly achieve primaryschoollevel performance and within a few years quotgraduatequot to realworld applicationsreading comprehension tests can serve as a testbed providing an impetus for research in a number of areas bottlenecks for lexical and world knowledgein addition research into collaboration might lead to insights about intelligent tutoringfinally reading comprehension evaluates systems abilities to answer ad hoc domainindependent questions this ability supports fact retrieval as opposed to document retrieval which could augment future search engines see kupiec for an example of such workthere has been previous work on story understanding that focuses on inferential processing common sense reasoning and world knowledge required for indepth understanding of storiesthese efforts concern themselves with specific aspects of knowledge representation inference techniques or question types see lehnert or schubert in contrast our research is concerned with building systems that can answer ad hoc questions about arbitrary documents from varied domainswe report here on our initial pilot study to determine the feasibility of this taskwe purchased a small corpus of development and test materials consisting of remedial reading materials for grades 36 these materials are simulated news stories followed by shortanswer quot5wquot questions who what when where and why questionswe developed a simple modular baseline system that uses pattern matching techniques and limited linguistic processing to select the sentence from the text that best answers the querywe used our development corpus to explore several alternative evaluation techniques and then evaluated on the test set which was kept blindwe had three goals in choosing evaluation metrics for our systemfirst the evaluation should be automaticsecond it should maintain comparability with human benchmarksthird it should require little or no effort to prepare new answer keyswe used three metrics pr humsent and autsent which satisfy these constraints to varying degreespr was the precision and recall on stemmed content words2 comparing the system response at the word level to the answer key provided by the test publisherhumsent and autsent compared the sentence chosen by the system to a list of acceptable answer sentences scoring one point for a response on the list and zero points otherwisein all cases the score for a set of questions was the average of the scores for each questionfor pr the answer key from the publisher was used unmodifiedthe answer key for humsent was compiled by a human annotator i these materials consisted of levels 25 of quotthe 5 wquot written by linda miller which can be purchased from remedia publications 10135 e via linda d124 scottsdale az 85258repeated words in the answer key match or fail togetherall words are stemmed and stop words are removedat present the stopword list consists of forms of be have and do personal and possessive pronouns the conjunctions and or the prepositions to in at of the articles a and the and the relative and demonstrative pronouns this that and which who examined the texts and chose the sentence that best answered the question even where the sentence also contained additional informationfor autsent an automated routine replaced the human annotator examining the texts and choosing the sentences this time based on which one had the highest recall compared against the published answer keyfor pr we note that in figure 2 there are two content words in the answer key and sentence 1 matches both of them for 22 100 recallthere are seven content words in sentence 1 so it scores 27 29 precisionsentence 2 scores 1250 recall and 1617 precisionthe human preparing the list of acceptable sentences for humsent has a problemsentence 2 responds to the question but requires pronoun coreference to give the full answer sentence 1 contains the words of the answer but the sentence as a whole does not really answer the questionin this and other difficult cases we have chosen to list no answers for the human metric in which case the system receives zero points for the questionthis occurs 11 of the time in our test corpusthe question is still counted meaning that the system receives a penalty in these casesthus the highest score a system could achieve for humsent is 89given that our current system can only respond with sentences from the text this penalty is appropriatethe automated routine for preparing the answer key in autsent selects as the answer key the sentence with the highest recall thus only sentence 1 would be counted as a correct answerwe have implemented all three metricshumsent and autsent are comparable with human benchmarks since they provide a binary score as would a teacher for a student answerin contrast the precision and recall scores of pr lack such a straightforward comparabilityhowever word recall from pare closely mimics the scores of humsent and autsentthe correlation coefficient for answdrecall to humsent in our test set is 98 and from humsent to autsent is also 98with respect to ease of answer key preparation pr and autsent are clearly superior since they use the publisherprovided answer keyhumsent requires human annotation for each questionwe found this annotation to be of moderate difficultyfinally we note that precision as well as recall will be useful to evaluate systems that can return clauses or phrases possibly constructed rather than whole sentence extracts as answerssince most national standardized tests feature a large multiplechoice component many available benchmarks are multiplechoice examsalso although our shortanswer metrics do not impose a penalty for incorrect answers multiplechoice exams such as the scholastic aptitude tests doin realworld applications it might be important that the system be able to assign a confidence level to its answerspenalizing incorrect answers would help guide development in that regardwhile we were initially concerned that adapting the system to multiplechoice questions would endanger the goal of realworld applicability we have experimented with minor changes to handle the multiple choice formatinitial experiments indicate that we can use essentially the same system architecture for both shortanswer and multiple choice teststhe process of taking shortanswer reading comprehension tests can be broken down into the following subtasks a crucial component of all three of these subtasks is the representation of information in textbecause our goal in designing our system was to explore the difficulty of various reading comprehension exams and to measure baseline performance we tried to keep this initial implementation as simple as possibleour system represents the information content of a sentence as the set of words in the sentencethe word sets are considered to have no structure or order and contain unique elementsfor example the representation for is the set in la by giving it 6457 of his books thomas jefferson helped get it started lb 16457 books by get giving helped his it jefferson of started thomas extraction of information content from text both in documents and questions then consists of tokenizing words and determining sentence boundary punctuationfor english written text both of these tasks are relatively easy although not trivialsee palmer and hearst the search subtask consists of finding the best match between the word set representing the question and the sets representing sentences in the documentour system measures the match by size of the intersection of the two word setsfor example the question in would receive an intersection score of 1 because of the mutual set element booksbecause match size does not produce a complete ordering on the sentences of the document we additionally prefer sentences that first match on longer words and second occur earlier in the documentin this section we describe extensions to the extraction approach described abovein the next section we will discuss the performance benefits of these extensionsthe most straightforward extension is to remove function or stop words such as the of a etc from the word sets reasoning that they offer little semantic information and only muddle the signal from the more contentful wordssimilarly one can use stemming to remove inflectional affixes from the words such normalization might increase the signal from contentful wordsfor example the intersection between and would include give if inflection were removed from gave and givingwe used a stemmer described by abney a different type of extension is suggested by the fact that who questions are likely to be answered with words that denote people or organizationssimilarly when and where questions are answered with words denoting temporal and locational words respectivelyby using name taggers to identify person location and temporal information we can add semantic class symbols to the question word sets marking the type of the question and then add corresponding class symbols to the word sets whose sentences contain phrases denoting the proper type of entityfor example due to the name thomas jefferson the word set in would be extended by person as would the word set because it is a who questionthis would increase the matching score by onethe system makes use of the alembic automated named entity system for finding named entitiesin a similar vein we also created a simple common noun classification module using wordnet it works by looking up all nouns of the text and adding person or location classes if any of a noun senses is subsumed by the appropriate wordnet classwe also created a filtering module that ranks sentences higher if they contain the appropriate class identifier even though they may have fewer matching words eg if the bag representation of a sentence does not contain person it is ranked lower as an answer to a who question than sentences which do contain personfinally the system contains an extension which substitutes the referent of personal pronouns for the pronoun in the bag representationfor example if the system were to choose the sentence he gave books to the library the answer returned and scored would be thomas jefferson gave books to the library if he were resolved to thomas jeffersonthe current system uses a very simplistic pronoun resolution system whichour modular architecture and automated scoring metrics have allowed us to explore the effect of various linguistic sources of information on overall system performancewe report here on three sets of findings the value added from the various linguistic modules the questionspecific results and an assessment of the difficulty of the reading comprehension taskwe were able to measure the effect of various linguistic techniques both singly and in combination with each other as shown in figure 3 and table 1the individual modules are indicated as follows name is the alembic named tagger described abovenamehum is handtagged named entitystem is abney automatic stemming algorithmflit is the filtering modulepro is automatic name and personal pronoun coreferenceprolitun is handtagged full reference resolutionsem is the wordnetbased common noun semantic classificationwe computed significance using the nonparametric significance test described by noreen the following performance improvements of the answdrecall metric were statistically significant results at a confidence level of 95 base vs namestem namestem vs filtnamehumstem and filtnamehumstem vs filtprohumnamehumstemthe other adjacent performance differences in figure 3 are suggestive but not statistically significantremoving stop words seemed to hurt overall performance slightlyit is not shown herestemming on the other hand produced a small but fairly consistent improvementwe compared these results to perfect stemming which made little difference leading us to conclude that our automated stemming module worked well enoughname identification provided consistent gainsthe alembic name tagger was developed for newswire text and used here with no modificationswe created handtagged named entity data which allowed us to measure the performance of alembic the accuracy was 765 see chinchor and sundheim for a description of the standard muc scoring metricthis also allowed us to simulate perfect tagging and we were able to determine how much we might gain by improving the name tagging by tuning it to this domainas the results indicate there would be little gain from improved name tagging however some modules that seemed to have little effect with automatic name tagging provided small gains with perfect name tagging specifically wordnet common noun semantics and automatic pronoun resolutionwhen used in combination with the filtering module these also seemed to helpsimilarly the handtagged reference resolution data allowed us to evaluate automatic coreference resolutionthe latter was a combination of name coreference as determined by alembic and a heuristic resolution of personal pronouns to the most recent prior named personusing the muc coreference scoring algorithm this had a precision of 77 and a recall of 183 the use of full handtagged reference resolution caused a substantial increase of the answdrecall metricthis was because the system substitutes the antecedent for all referring expressions improving the wordbased measurethis did not however provide an increase in the sentencebased measuresfinally we plan to do similar human labeling experiments for semantic class identification to determine the potential effect of this knowledge sourceour results reveal that different questiontypes behave very differently as shown in figure 4why questions are by far the hardest because they require understanding of rhetorical structure and because answers tend to be whole clauses rather than phrases embedded in a context that matches the query closelyon the other hand who and when queries benefit from reliable person name and time extractionwho questions seem to benefit most dramatically from perfect name tagging combined with filtering and pronoun resolutionwhat questions show relatively little benefit from the various linguistic techniques probably because there are many types of what question most of which are not answered by a person time or placefinally where question results are quite variable perhaps because location expressions often do not include specific place names3 the low recall is attributable to the fact that the heuristic asigned antecedents only for names and pronouns and completely ignored definite noun phrases and plural pronousthese results indicate that the sample tests are an appropriate and challenging taskthe simple techniques described above provide a system that finds the correct answer sentence almost 40 of the timethis is much better than chance which would yield an average score of about 45 for the sentence metrics given an average document length of 20 sentencessimple linguistic techniques enhance the baseline system score from the low 30 range to almost 40 in all three metricshowever capturing the remaining 60 will clearly require more sophisticated syntactic semantic and world knowledge sourcesour pilot study has shown that reading comprehension is an appropriate task providing a reasonable starting level it is tractable but not trivialour next steps include standardized multiplechoice reading comprehension testthis will require some minor changes in strategyfor example in preliminary experiments our system chose the answer that had the highest sentence matching score when composed with the questionthis gave us a score of 45 on a small multiplechoice test setsuch tests require us to deal with a wider variety of question types eg what is this story aboutthis will also provide an opportunity to look at rejection measures since many tests penalize for random guessing moving from whole sentence retrieval towards answer phrase retrievalthis will allow us to improve answer word precision which provides a good measure of how much extraneous material we are still returning adding new linguistic knowledge sourceswe need to perform further hand annotation experiments to determine the effectiveness of semantic class identification and lexical semantics encoding more semantic information in our representation for both question and document sentencesthis information could be derived from syntactic analysis including noun chunks verb chunks and clause groupings cooperation with educational testing and content providerswe hope to work together with one or more major publishersthis will provide the research community with a richer collection of training and test material while also providing educational testing groups with novel ways of checking and benchmarking their testswe have argued that taking reading comprehension exams is a useful task for developing and evaluating natural language understanding systemsreading comprehension uses found material and provides humancomparable evaluations which can be computed automatically with a minimum of human annotationcrucially the reading comprehension task is neither too easy nor too hard as the performance of our pilot system demonstratesfinally reading comprehension is a task that is sufficiently close to information extraction applications such as ad hoc question answering fact verification situation tracking and document summarization that improvements on the reading comprehension evaluations will result in improved systems for these applicationswe gratefully acknowledge the contribution of lisa ferro who prepared much of the handtagged data used in these experiments	P99-1042	deep read a reading comprehension systemthis paper describes initial work on deep read an automated reading comprehension system that accepts arbitrary text input and answers questions about itwe have acquired a corpus of 60 development and 60 test stories of 3rd to 6th grade material each story is followed by shortanswer questions we used these to construct and evaluate a baseline system that uses pattern matching techniques augmented with additional automated linguistic processing this simple system retrieves the sentence containing the answer 3040 of the timewe use a statistical bagofwords approach matching the question with the lexically most similar sentence in the story
corpusbased identification of nonanaphoric noun phrases coreference resolution involves finding antecedents for anaphoric discourse entities such as definite noun phrases but many definite noun phrases are not anaphoric because their meaning can be understood from general world knowledge we have developed a corpusbased algorithm for automatically identifying definite noun phrases that are nonanaphoric which has the potential to improve the efficiency and accuracy of coreference resolution systems our algorithm generates lists of nonanaphoric noun phrases and noun phrase patterns from a training corpus and uses them to recognize nonanaphoric noun phrases in new texts using 1600 muc4 terrorism news articles as the training corpus our approach achieved 78 recall and 87 precision at identifying such noun phrases in 50 test documents most automated approaches to coreference resolution attempt to locate an antecedent for every potentially coreferent discourse entity in a textthe problem with this approach is that a large number of de may not have antecedentswhile some discourse entities such as pronouns are almost always referential definite descriptionsquot may not beearlier work found that nearly 50 of definite descriptions had no prior referents and we found that number to be even higher 63 in our corpussome nonanaphoric definite descriptions can be identified by looking for syntactic clues like attached prepositional phrases or restrictive relative clausesbut other definite descriptions are nonanaphoric because readers understand their meaning due to common knowledgefor example readers of this in this work we define a definite description to be a noun phrase beginning with the paper will probably understand the real world referents of quotthe fbiquot quotthe white housequot and quotthe golden gate bridgequot these are instances of definite descriptions that a coreference resolver does not need to resolve because they each fully specify a cognitive representation of the entity in the reader mindone way to address this problem is to create a list of all nonanaphoric nps that could be used as a filter prior to coreference resolution but hand coding such a list is a daunting and intractable taskwe propose a corpusbased mechanism to identify nonanaphoric nps automaticallywe will refer to nonanaphoric definite noun phrases as existential nps our algorithm uses statistical methods to generate lists of existential noun phrases and noun phrase patterns from a training corpusthese lists are then used to recognize existential nps in new textscomputational coreference resolvers fall into two categories systems that make no attempt to identify nonanaphoric discourse entities prior to coreference resolution and those that apply a filter to discourse entities identifying a subset of them that are anaphoricthose that do not practice filtering include decision tree models that consider all possible combinations of potential anaphora and referentsexhaustively examining all possible combinations is expensive and we believe unnecessaryof those systems that apply filtering prior to coreference resolution the nature of the filtering variessome systems recognize when an anaphor and a candidate antecedent are incompatiblein sri probabilistic model a pair of extracted templates may be removed from consideration because an outside knowledge base indicates contradictory featuresother systems look for particular constructions using certain trigger wordsfor example pleonastic2 pronouns are identified by looking for modal adjectives or cognitive verbs in a set of patterned constructions a more recent system recognizes a large percentage of nonanaphoric definite noun phrases during the coreference resolution process through the use of syntactic cues and casesensitive rulesthese methods were successful in many instances but they could not identify them allthe existential nps that were missed were existential to the reader not because they were modified by particular syntactic constructions but because they were part of the reader general world knowledgedefinite noun phrases that do not need to be resolved because they are understood through world knowledge can represent a significant portion of the existential noun phrases in a textin our research we found that existential nps account for 63 of all definite nps and 24 of them could not be identified by syntactic or lexical meansthis paper details our method for identifying existential nps that are understood through general world knowledgeour system requires no hand coded information and can recognize a larger portion of existential nps than vieira and poesio systemto better understand what makes an np anaphoric or nonanaphoric we found it useful to classify definite nps into a taxonomywe first classified definite nps into two broad categories referential nps which have prior referents in the texts and existential nps which do notin figure 1 examples of referential nps are quotthe mass kidnappingquot quotthe terroristsquot and quotthe individualsquot while examples of existential nps are quotthe arce battalion commandquot and quotthe farabundo marti national liberation frontquot we should clarify an important pointwhen we say that a definite np is existential we say this because it completely specifies a cognitive representation of the entity in the reader mindthat is suppose quotthe fbiquot appears in both sentence 1 and sentence 7 of a textalthough there may be a cohesive relationship between the noun phrases because they both completely specify independently we consider them to be nonanaphoricdefinite noun phrases we further classified existential nps into two categories independent and associative which are distinguished by their need for contextindependent existentials can be understood in isolationassociative existentials are inherently associated with an event action object or other context3in a text about a basketball game for example we might find quotthe scorequot quotthe hoopquot and quotthe bleachersquot although they may that our independent existentials roughly equate to her new class our associative existentials to her inferable class and our referentials to her evoked class not have direct antecedents in the text we understand what they mean because they are all associated with basketball gamesin isolation a reader would not necessarily understand the meaning of quotthe scorequot because context is needed to disambiguate the intended word sense and provide a complete specificationbecause associative nps represent less than 10 of the existential nps in our corpus our efforts were directed at automatically identifying independent existentialsunderstanding how to identify independent existential nps requires that we have an understanding of why these nps are existentialwe classified independent existentials into two groups semantic and syntacticsemantically independent nps are existential because they are understood by readers who share a collective understanding of current events and world knowledgefor example we understand the meaning of quotthe fbiquot without needing any other informationsyntactically independent nps on the other hand gain this quality because they are modified structurallyfor example in quotthe man who shot liberty valencequot quotthe manquot is existential because the relative clause uniquely identifies its referentour goal is to build a system that can identify independent existential noun phrases automaticallyin the previous section we observed that quotexistentialismquot can be granted to a definite noun phrase either through syntax or semanticsin this section we introduce four methods for recognizing both classes of existentialswe began by building a set of syntactic heuristics that look for the structural cues of restrictive premodification and restrictive postmodificationrestrictive premodification is often found in noun phrases in which a proper noun is used as a modifier for a head noun for example quotthe yous presidentquot quotthe presidentquot itself is ambiguous but quotthe yous presidentquot is notrestrictive postmodification is often represented by restrictive relative clauses prepositional phrases and appositivesfor example quotthe president of the united statesquot and quotthe president who governs the yousquot are existential due to a prepositional phrase and a relative clause respectivelywe also developed syntactic heuristics to recognize referential npsmost nps of the form quotthe quot have an antecedent so we classified them as referentialalso if the head noun of the np appeared earlier in the text we classified the np as referentialthis method then consists of two groups of syntactic heuristicsthe first group which we refer to as the rulein heuristics contains seven heuristics that identify restrictive premodification or postmodification thus targeting existential npsthe second group referred to as the ruleout heuristics contains two heuristics that identify referential npsmost referential nps have antecedents that precede them in the textthis observation is the basis of our first method for identifying semantically independent npsif a definite np occurs in the first sentence4 of a text we assume the np is existentialusing a training corpus we create a list of presumably existential nps by collecting the first sentence of every text and extracting all definite nps that were not classified by the syntactic heuristicswe call this list the 51 extractionswhile examining the si extractions we found many similar nps for example quotthe salvadoran governmentquot quotthe guatemalan governmentquot and quotthe yous governmentquot the similarities indicate that some head nouns when premodified represent existential entitiesby using the si extractions as input to a pattern generation algorithm we built a set of existential head patterns that identify such constructionsthese patterns are of the form quotthe 5 quot such as quotthe governmentquot or quotthe salvadoran governmentquot figure 3 shows the algorithm for creating ehpsit also became clear that some existentials never appear in indefinite constructionsquotthe fbiquot quotthe contraryquot quotthe national guardquot are definite nps which are rarely if ever seen in indefinite constructionsthe chances that a reader will encounter quotan fbiquot are slim to nonethese nps appeared to be perfect candidates for a corpusbased approachto locate quotdefiniteonlyquot nps we made two passes over the corpusthe first pass produced a list of every definite np and its frequencythe second pass counted indefinite uses of all nps cataloged during the first passknowing how often an np was used in definite and indefinite constructions allowed us to sort the nps first by the probability of being used as a definite and second by definiteuse frequencyfor example quotthe contraryquot appeared high on this list because its head noun occurred 15 times in the training corpus and every time it was in a definite constructionfrom this we created a definiteonly list by selecting those nps which occurred at least 5 times and only in definite constructionsexamples from the three methods can be found in the appendixour methods for identifying existential nps are all heuristicbased and therefore can be incorrect in certain situationswe identified two types of common errorsto address these problems we developed a vaccineit was clear that we had a number of infections in our si list including quotthe basequot quotthe for every definite np in a text individualsquot quotthe attackquot and quotthe banksquot we noticed however that many of these incorrect nps also appeared near the bottom of our definiteindefinite list indicating that they were often seen in indefinite constructionswe used the definite probability measure as a way of detecting errors in the si and ehp listsif the definite probability of an np was above an upper threshold the np was allowed to be classified as existentialif the definite probability of an np fell below a lower threshold it was not allowed to be classified by the si or ehp methodthose nps that fell between the two thresholds were considered occasionally existentialoccasionally existential nps were handled by observing where the nps first occurred in the textfor example if the first use of quotthe guerrillasquot was in the first few sentences of a text it was usually an existential useif the first use was later it was usually a referential use because a prior definition appeared in earlier sentenceswe applied an early allowance threshold of three sentences occasionally existential nps occuring under this threshold were classified as existential and those that occurred above were left unclassifiedfigure 4 details the vaccine algorithmwe trained and tested our methods on the latin american newswire articles from muc4 the training set contained 1600 texts and the test set contained 50 textsall texts were first parsed by sundance our heuristicbased partial parser developed at the university of utahwe generated the si extractions by processing the first sentence of all training textsthis produced 849 definite npsusing these nps as input to the existential head pattern algorithm we generated 297 ehpsthe do list was built by using only those nps which appeared at least 5 times in the corpus and 100 of the time as definiteswe generated the do list in two iterations once for head nouns alone and once for full nps resulting in a list of 65 head nouns and 321 full nps6once the methods had been trained we classified each definite np in the test set as referential or existential using the algorithm in figure 5figure 6 graphically represents the main elements of the algorithmnote that we applied vaccines to the si and ehp lists but not to the do list because gaining entry to the do list is much more difficult an np must occur at least 5 times in the training corpus and every time it must occur in a definite constructionto evaluate the performance of our algorithm we handtagged each definite np in the 50 test texts as a syntactically independent existential a semantically independent existential an associative existential or a referential npfigure 8 shows the distribution of definite np types in the test textsof the 1001 definite nps tested 63 were independent existentials so removing these nps from the coreference resolution process could have substantial savingswe measured the accuracy of our classifications using recall and precision metricsresults are shown in figure 7as a baseline measurement we considered the accuracy of classifying every definite np as existentialgiven the distribution of definite np types in our test set this would result in recall of 100 and precision of 72note that we are more interested in high measures of precision than recall because we view this method to be the precursor to a coreference resolution algorithmincorrectly removing an anaphoric np means that the coreference resolver would never have a chance to resolve it on the other hand nonanaphoric nps that slip through can still be ruled as nonanaphoric by the coreference resolverwe first evaluated our system using only the syntactic heuristics which produced only 43 recall but 92 precisionalthough the syntactic heuristics are a reliable way to identify existential definite nps they miss 57 of the true existentialswe expected the si ehp and do methods to increase coveragefirst we evaluated each method independently the results appear in rows 24 of figure 7each method increased recall to between 6169 but decreased precision to 8487all of these methods produced a substantial gain in recall at some cost in precisionnext we tried combining the methods to make sure that they were not identifying exactly the same set of existential npswhen we combined the si and ehp heuristics recall increased to 80 with precision dropping only slightly to 82when we combined all three methods recall increased to 82 without any corresponding loss of precisionthese experiments show that these heuristics substantially increase recall and are identifying different sets of existential npsfinally we tested our vaccine algorithm to see if it could increase precision without sacrificing much recallwe experimented with two variations va used an upper definite probability threshold of 70 and vi used an upper definite probability threshold of 50both variations used a lower definite probability threshold of 25the results are shown in rows 78 of figure 7both vaccine variations increased precision by several percentage points with only a slight drop in recallin previous work the system developed by vieria poesio achieved 74 recall and 85 precision for identifying quotlarger situation and unfamiliar usequot npsthis set of nps does not correspond exactly to our definition of existential nps because we consider associative nps to be existential and they do noteven so our results are slightly better than their previous resultsa more equitable comparison is to measure our system performance on only the independent existential noun phrasesusing this measure our algorithm achieved 818 recall with 856 precision using va and achieved 829 recall with 835 precision using vbwe have developed several methods for automatically identifying existential noun phrases using a training corpusit accomplishes this task with recall and precision measurements that exceed those of the earlier vieira rz poesio system while not exploiting full parse trees appositive constructions handcoded lists or case sensitive text7in addition because the system is fully automated and corpusbased it is suitable for applications that require portability across domainsgiven the large percentage of nonanaphoric discourse entities handled by most coreference resolvers we believe that using a system like ours to filter existential nps has the potential to reduce processing time and complexity and improve the accuracy of coreference resolution	P99-1048	corpusbased identification of nonanaphoric noun phrasescoreference resolution involves finding antecedents for anaphoric discourse entities such as definite noun phrasesbut many definite noun phrases are not anaphoric because their meaning can be understood from general world knowledge we have developed a corpusbased algorithm for automatically identifying definite noun phrases that are nonanaphoric which has the potential to improve the efficiency and accuracy of coreference resolution systemsour algorithm generates lists of nonanaphoric noun phrases and noun phrase patterns from a training corpus and uses them to recognize nonanaphoric noun phrases in new textsusing 1600 muc4 terrorism news articles as the training corpus our approach achieved 78 recall and 87 precision at identifying such noun phrases in 50 test documentswe develop a system for identifying discoursenew dds that incorporates in addition to syntaxbased heuristics aimed at recogznizing predicative and established dds additional techniques for mining from corpora unfamiliar dds including proper names larger situation and semantically functionalwe develop an unsupervised learning algorithm that automatically recognizes definite nps that are existential without syntactic modification because their meaning is universally understood
efficient parsing for bilexical contextfree grammars and head automaton grammars stochastic parsers use grammars where each word type idiosyncratically prefers particular complements with parhead words we present parsing algorithms for two bilexical formalisms improvthe prior upper bounds of for a comspecial case that was known to allow we present an algorithm with an improved grammar constant lexicalized grammar formalisms are of both theoretical and practical interest to the computational linguistics communitysuch formalisms specify syntactic facts about each word of the languagein particular the type of arguments that the word can or must takeearly mechanisms of this sort included categorial grammar and subcategorization frames other lexicalized formalisms include besides the possible arguments of a word a naturallanguage grammar does well to specify possible head words for those argumentsquotconvenequot requires an np object but some nps are more semantically or lexically appropriate here than others and the appropriateness depends largely on the np head we use the general term bilexical for a grammar that records such factsa bilexical grammar makes many stipulations about the compatibility of particular pairs of words in particular rolesthe acceptability of quotnora convened the the authors were supported respectively under arpa grant n6600194c6043 quothuman language technologyquot and ministero delluniversita e della ricerca scientifica e tecnologica project quotmethodologies and tools of high performance systems for multimedia applicationsquot partyquot then depends on the grammar writer assessment of whether parties can be convenedseveral recent realworld parsers have improved stateoftheart parsing accuracy by relying on probabilistic or weighted versions of bilexical grammars the rationale is that soft selectional restrictions play a crucial role in disambiguation1 the chart parsing algorithms used by most of the above authors run in time 0 because bilexical grammars are enormous in practiceheavy probabilistic pruning is therefore needed to get acceptable runtimesbut in this paper we show that the complexity is not so bad after all grammars where an 0 algorithm was previously known the grammar constant can be reduced without harming the 0 propertyour algorithmic technique throughout is to propose new kinds of subderivations that are not constituentswe use dynamic programming to assemble such subderivations into a full parsethe reader is assumed to be familiar with contextfree grammarsour notation follows a contextfree grammar is a tuple g where vn and vt are finite disjoint sets of nonterminal and terminal symbols respectively and s e vn is the start symbolset p is a finite set of productions having the form a a where a e vn a e if every production in p has the form a 4 because or a a for a because e vn a e vt then the grammar is said to be in chomsky normal form 2 every language that can be generated by a cfg can also be generated by a cfg in cnfin this paper we adopt the following conventions a b c d denote symbols in vt w x y denote strings in vat and a 0 denote strings in the input to the parser will be a cfg g together with a string of terminal symbols to be parsed w d1d2 dnalso hi jk denote positive integers which are assumed to be ja quotderivesquot relation written is associated with a cfg as usualwe also use the reflexive and transitive closure of written and define l accordinglywe write a 8 a0 for a derivation in which only 3 is rewrittenwe introduce next a grammar formalism that captures lexical dependencies among pairs of words in vtthis formalism closely resembles stochastic grammatical formalisms that are used in several existing natural language processing systems we will specify a nonstochastic version noting that probabilities or other weights may be attached to the rewrite rules exactly as in stochastic cfg suppose g is a cfg in cnf3 we say that g is bilexical if there exists a set of quotdelexicalized nonterminalsquot vd such that vn aa a e vd a e vt and every production in p has one of the following forms thus every nonterminal is lexicalized at some terminal aa constituent of nonterminal type aa is said to have terminal symbol a as its lexical head quotinheritedquot from the constituent head child in the parse tree notice that the start symbol is necessarily a lexicalized nonterminal thence appears in every string of l it is usually convenient to define g so that the language of interest is actually l x x e l such a grammar can encode lexically specific preferencesfor example p might contain the productions in order to allow the derivation vpsolve solve two puzzles but meanwhile omit the similar productions since puzzles are not edible a goat is not solvable quotsleepquot is intransitive and quotgoatquot cannot take plural determinersthe cost of this expressiveness is a very large grammarstandard contextfree parsing algorithms are inefficient in such a casethe cky algorithm is time 0 where in the worst case i p1 ivni3 for a bilexical grammar the worst case is ipi vd i 3 i vt12 which is large for a large vocabulary vt we may improve the analysis somewhat by observing that when parsing d1 dn the cky algorithm only considers nonterminals of the form adi by restricting to the relevant productions we obtain 02 we observe that in practical applications we always have n aa parse is just a derivation of lhn and its probabilitylike that of any derivation we findis defined as the product of the probabilities of all productions used to condition inference rules in the proof treethe highestprobability derivation for any item can be reconstructed recursively at the end of the parse provided that each item maintains not only a bit indicating whether it can be derived but also the probability and instantiated root rule of its highestprobability derivation treewe now give a variant of the algorithm of 4 the variant has the same asymptotic complexity but will often be faster in practicenotice that the attachleft rule of figure 1 tries to combine the nonterminal label bdhd of a previously derived constituent with every possible nonterminal label of the form cdhthe improved version shown in figure 2 restricts cdh to be the label of a previously derived adjacent constituentthis improves speed if there are not many such constituents and we can enumerate them in 0 time apiece it is necessary to use an agenda data structure when implementing the declarative algorithm of figure 2deriving narrower items before wider ones as before will not work here because the rule halve derives narrow items from wide onesrather than parsing an input string directly it is often desirable to parse another string related by a transductionlet t be a finitestate transducer that maps a morpheme sequence w e vit to its orthographic realization a grapheme sequence fo t may realize arbitrary morphological processes including affixation local clitic movement deletion of phonological nulls forbidden or dispreferred kgrams typographical errors and mapping of multiple senses onto the same graphemegiven grammar g and an input ti we ask whether e twe have extended all the algorithms in this paper to this case the items simply keep track of the transducer state as welldue to space constraints we sketch only the special case of multiple sensessuppose that the input is id d1 dn and each d2 has up to g possible senseseach item now needs to track its head sense along with its head position in idwherever an item formerly recorded a head position h it must now record a pair where dh e vt is a specific sense of dhno rule in figures 12 will mention more than two such pairsso the time complexity increases by a factor of 07 head automaton grammars in time 0 in this section we show that a lengthn string generated by a head automaton grammar can be parsed in time 0we do this by providing a translation from head automaton grammars to bilexical cfgs4 this result improves on the headautomaton parsing algorithm given by alshawi which is analogous to the cky algorithm on bilexical cfgs and is likewise 0 in practice a head automaton grammar is a function h a 14 ha that defines a head automaton for each element of its domainlet vt domain and d a single head automaton is an acceptor for a language of string pairs e v x v informally if b is the leftmost symbol of zr and q e a then ha can move from state q to state q matching symbol b and removing it from the left end of zrsymmetrically if b is the rightmost symbol of zi and q e sa only by reading all of yimmediately after which it is said to be in a flip stateand then reading all of xformally a flip state is one that allows entry on a transition and that either allows exit on a transition or is a final statewe are concerned here with head automaton grammars h such that every ha is splitthese correspond to bilexical cfgs in which any derivation aa xay has the form aa xba xaythat is a word left dependents are more oblique than its right dependents and ccommand themsuch grammars are broadly applicableeven if ha is not split there usually exists a split head automaton hquot recognizing the same languageh la exists if xy e l is regular in particular lei must exist unless ha has a cycle that includes both and 4 transitionssuch cycles would be necessary for ha itself to accept a formal language such as n 0 where word a takes 2n dependents but we know of no naturallanguage motivation for ever using them in a hagone more definition will help us bound the complexitya split head automaton ha is said to be gsplit if its set of flip states denoted c qa has size g the languages that can be recognized by gsplit has are those that can be written as 1g 1 li x r where the li and ri are regular languages over vt eisner actually defined bilexical grammars in terms of the latter property6 we now present our result figure 3 specifies an 0 recognition algorithm for a head automaton grammar h in which every h is gsplitfor deterministic automata the runtime is 0a considerable improvement on the 0 result of which also assumes deterministic automataas in 4 a simple bottomup implementation will sufficefor a practical speedup add hj as an antecedent to the mid rule like our previous algorithms this one takes two steps to attach a child constituent to a parent constituentbut instead of full constituentsstrings xdy e d it uses only halfconstituents like xdi and the other halves of these constituents can be attached later because to find an accepting path for in a split head automaton one can separately find the halfpath before the flip state and the halfpath after the flip state these two halfpaths can subsequently be joined into an accepting path if they have the same flip state s ie one path starts where the other endsannotating our left halfconstituents with s makes this check possiblewe have formally described and given faster parsing algorithms for three practical grammatical rewriting systems that capture dependencies between pairs of wordsall three systems admit naive 0 algorithmswe give the first 0 results for the natural formalism of bilexical contextfree grammar and for alshawi head automaton grammarsfor the usual case split head automaton grammars or equivalent bilexical cfgs we replace the 0 algorithm of by one with a smaller grammar constantnote that eg all senses would restore the g 2 factorindeed this approach gives added flexibility a word sense unlike its choice of flip state is visible to the ha that reads it three models in are susceptible to the 0 method our dynamic programming techniques for cheaply attaching head information to derivations can also be exploited in parsing formalisms other than rewriting systemsthe authors have developed an 0time parsing algorithm for bilexicalized tree adjoining grammars improving the naive 0 methodthe results mentioned in 6 are related to the closure property of cfgs under generalized sequential machine mapping this property also holds for our class of bilexical cfgs	P99-1059	efficient parsing for bilexical contextfree grammars and head automaton grammarsseveral recent stochastic parsers use bilexical grammars where each word type idiosyncratically prefers particular complements with particular head wordswe present o parsing algorithms for two bilexical formalisms improving the prior upper bounds of ofor a common special case that was known to allow o parsing we present an o algorithm with an improved grammar constantwe show that the dynamic programming algorithms for lexicalized pcfgs require o states
a statistical parser for czech this paper considers statistical parsing of czech which differs radically from english in at least two it is a inflected and it has relatively word order differences are likely to pose new problems for techniques that have been developed on english we describe our experience in building on the parsing model of our final results 80 dependency accuracy represent good progress towards the 91 accuracy of the parser on english text much of the recent research on statistical parsing has focused on english languages other than english are likely to pose new problems for statistical methodsthis paper considers statistical parsing of czech using the prague dependency treebank as a source of training and test data also show these characteristicsmany european languages exhibit fwo and hi phenomena to a lesser extentthus the techniques and results found for czech should be relevant to parsing several other languagesthis paper first describes a baseline approach based on the parsing model of which recovers dependencies with 72 accuracywe then describe a series of refinements to the model giving an improvement to 80 accuracy with around 82 accuracy on newspaperbusiness articles textthe prague dependency treebank pdt has been modeled after the penn treebank with one important exception following the praguian linguistic tradition the syntactic annotation is based on dependencies rather than phrase structuresthus instead of quotnonterminalquot symbols used at the nonleaves of the tree the pdt uses socalled analytical functions capturing the type of relation between a dependent and its governing nodethus the number of nodes is equal to the number of tokens plus one the pdt contains also a traditional morphosyntactic annotation at each word position as czech is a hi language the size of the set of possible tags is unusually high more than 3000 tags may be assigned by the czech morphological analyzerthe pdt also contains machineassigned tags and lemmas for each word for evaluation purposes the pdt has been divided into a training set and a developmentevaluation test set pair parsing accuracy is defined as the ratio of correct dependency links vs the total number of dependency links in a sentence as usual with the development test set being available during the development phase all final results has been obtained on the evaluation test set which nobody could see beforehandthe parsing model builds on model 1 of this section briefly describes the modelthe parser uses a lexicalized grammar each nonterminal has an associated headword and partofspeech we write nonterminals as x x is the nonterminal label and x is a pair where w is the associated headword and t as the pos tagsee figure 1 for an example lexicalized tree and a list of the lexicalized rules that it containseach rule has the fonnl with the exception of the top rule in the tree which has the form top h h is the headchild of the phrase which inherits the headword h from its parent p l1ln and ri are left and right modifiers of h either n or m may be zero and n the model can be considered to be a variant of probabilistic contextfree grammar in pcfgs each rule a in the cfg underlying the pcfg has an associated probability pin p is defined as a product of terms by assuming that the righthandside of the rule is generated in three steps probability i p h h where lni stopthe stop symbol is added to the vocabulary of nonterminals and the model stops generating left modifiers when it is generatedother rules in the tree contribute similar sets of probabilitiesthe probability for the entire tree is calculated as the product of all these terms describes a series of refinements to this basic model the addition of quotdistancequot the addition of subcategorization parameters and parameters that model whmovement estimation techniques that smooth various levels of backoff search for the highest probability tree for a sentence is achieved using a ckystyle parsing algorithmmany statistical parsing methods developed for english use lexicalized trees as a representation several emphasize the use of parameters associated with dependencies between pairs of wordsthe czech pdt contains dependency annotations but no tree structuresfor parsing czech we considered a strategy of converting dependency structures in training data to lexicalized trees then running the parsing algorithms originally developed for englisha key point is that the mapping from lexicalized trees to dependency structures is manytooneas an example figure 2 shows an input dependency structure and three different lexicalized trees with this dependency structurethe choice of tree structure is crucial in determining the independence assumptions that the parsing model makesthere are at least 3 degrees of freedom when deciding on the tree structures to provide a baseline result we implemented what is probably the simplest possible conversion scheme the baseline approach gave a result of 719 accuracy on the development test setwhile the baseline approach is reasonably successful there are some linguistic phenomena that lead to clear problemsthis section describes some tree transformations that are linguistically motivated and lead to improvements in parsing accuracyin the pdt the verb is taken to be the head of both sentences and relative clausesfigure 4 illustrates how the baseline transformation method can lead to parsing errors in relative clause casesfigure 4 shows the solution to the problem the label of the relative clause is changed to sbar and an additional vp level is added to the right of the relative pronounsimilar transformations were applied for relative clauses involving whpps whnps and whadverbials the pdt takes the conjunct to be the head of coordination structures in these cases the baseline approach gives tree structures such as that in figure 5the nonterminal label for the phrase is jp this choice of nonterminal is problematic for two reasons the jp label is assigned to all coordinated phrases for example hiding the fact that the constituent in figure 5 is an np the model assumes that left and right modifiers are generated independently of each other and as it stands will give unreasonably high probability to two unlike phrases being coordinatedto fix these problems the nonterminal label in coordination cases was altered to be the same as that of the second conjunct see figure 5a similar transformation was made for cases where a comma was the head of a phrasefigure 6 shows an additional change concerning commasthis change increases the sensitivity of the model to punctuationthis section describes some modifications to the parameterization of the model guish main clauses from relative clauses both have a verb as the head so both are labeled vp a typical parsing error due to relative and main clauses not being distinguished the solution to the problem a modification to relative clause structures in training datathe model of had conditioning variables that allowed the model to learn a preference for dependencies which do not cross verbsfrom the results in table 3 adding this condition improved accuracy by about 09 on the development setthe parser of used punctuation as an indication of phrasal boundariesit was found that if a constituent z has two children x and y separated by a punctuation mark then y is generally followed by a punctuation mark or the end of sentence markerthe parsers of encoded this as a hard constraintin the czech parser we added a cost of 25 2 to structures that violated this constraintthe model of section 3 made the assumption that modifiers are generated independently of each otherthis section describes a hi gram model where the context is increased to consider the previously generated modifier also describes use of bigram statisticsthe righthandside of a rule is now assumed to be generated in the following three step process where lo is defined as a special null symbolthus the previous modifier li_1 is added to the conditioning context part of speech tags serve an important role in statistical parsing by providing the model with a level of generalization as to how classes of words tend to behave what roles they play in sentences and what other classes they tend to combine withstatistical parsers of english typically make use of the roughly 50 pos tags used in the penn treebank corpus but the czech pdt corpus provides a much richer set of pos tags with over 3000 possible tags defined by the tagging system and over 1000 tags actually found in the corpususing that large a tagset with a training corpus of only 19000 sentences would lead to serious sparse data problemsit is also clear that some of the distinctions being made by the tags are more important than others for parsingwe therefore explored different ways of extracting smaller but still maximally informative pos tagsetsthe pos tags in the czech pdt corpus are encoded in 13character stringstable 1 shows the role of each characterfor example the tag nnmp1 a would be used for a word that had quotnounquot as both its main and detailed part of speech that was masculine plural nominative and whose negativeness value was quotaffirmativequotwithin the corpus each word was annotated with all of the pos tags that would be possible given its spelling using the output of a morphological analysis program and also with the single one of those tags that a statistical pos tagging program had predicted to be the correct tag table 2 shows a phrase from the corpus with the alternative possible tags and machineselected tag for each wordin the training portion of the corpus the correct tag as judged by human annotators was also providedin the baseline approach the first letter or quotmain part of speechquot of the full pos strings was used as the tagthis resulted in a tagset with 13 possible valuesa number of alternative richer tagsets were explored using various combinations of character positions from the tag stringthe most successful alternative was a twoletter tag whose first letter was always the main pos and whose second letter was the case field if the main pos was one that displays case while otherwise the second letter was the detailed posthis twoletter scheme resulted in 58 tags and provided about a 11 parsing improvement over the baseline on the development seteven richer tagsets that also included the person gender and number values were tested without yielding any further improvement presumably because the damage from sparse data outweighed the value of the additional information presentan entirely different approach rather than searching by hand for effective tagsets would be to use clustering to derive them automaticallywe explored two different methods bottomup and topdown for automatically deriving pos tag sets based on counts of governing and dependent tags extracted from the parse trees that the parser constructs from the training dataneither tested approach resulted in any improvement in parsing performance compared to the handdesigned quottwo letterquot tagset but the implementations of each were still only preliminary and a clustered tagset more adroitly derived might do betterone final issue regarding pos tags was how to deal with the ambiguity between possible tags both in training and testin the training data there was a choice between using the output of the pos tagger or the human annotator judgment as to the correct tagin test data the correct answer was not available but the pos tagger output could be used if desiredthis turns out to matter only for unknown words as the parser is designed to do its own tagging for words that it has seen in training at least 5 times ignoring any tag supplied with the inputfor quotunknownquot words the parser can be set either to believe the tag supplied by the pos tagger or to allow equally any of the dictionaryderived possible tags for the word effectively allowing the parse context to make the choice our tests indicated that if unknown words are treated by believing the pos tagger suggestion then scores are better if the parser is also trained on the pos tagger suggestions rather than on the human annotator correct tagstraining on the correct tags results in 1 worse performanceeven though the pos tagger tags are less accurate they are more like what the parser will be using in the test data and that turns out to be the key pointon the other hand if the parser allows all possible dictionary tags for unknown words in test material then it pays to train on the actual correct tagsin initial tests this combination of training on the correct tags and allowing all dictionary tags for unknown test words somewhat outperformed the alternative of using the pos tagger predictions both for training and for unknown test wordswhen tested with the final version of the parser on the full development set those two strategies performed at the same levelwe ran three versions of the parser over the final test set the baseline version the full model with all additions and the full model with everything but the bigram modelthe baseline system on the fithat although the science section only contributes 25 of the sentences in test data it contains much longer sentences than the other sections and therefore accounts for 38 of the dependencies in test data nal test set achieved 723 accuracythe final system achieved 800 accuracy3 a 77 absolute improvement and a 278 relative improvementthe development set showed very similar results a baseline accuracy of 719 and a final accuracy of 793table 3 shows the relative improvement of each component of the mode14table 4 shows the results on the development set by genreit is interesting to see that the performance on newswire text is over 2 better than the averaged performancethe science section of the development set is considerably harder to parse the main piece of previous work on parsing czech that we are aware of is described in this is a rulebased system which is based on a manually designed set of rulesthe system accuracy is not evaluated on a test corpus so it is difficult to compare our results to theirswe can however make some comparison of the results in this paper to those on parsing english describes results of 91 accuracy in recovering dependencies on section 0 of the penn wall street journal treebank using model 2 of this task is almost certainly easier for a number of reasons there was more training data wall street journal may be an easier domain than the pdt as a reasonable proportion of sentences come from a subdomain financial news which is relatively restrictedunlike model 1 model 2 of the parser takes subcategorization information into account which gives some improvement on english and might well also improve results on czechgiven these differences it is difficult to make a direct comparison but the overall conclusion seems to be that the czech accuracy is approaching results on english although it is still somewhat behindthe 80 dependency accuracy of the parser represents good progress towards english parsing performancea major area for future work is likely to be an improved treatment of morphology a natural approach to this problem is to consider more carefully how pos tags are used as word classes by the modelwe have begun to investigate this issue through the automatic derivation of pos tags through clustering or quotsplittingquot approachesit might also be possible to exploit the internal structure of the pos tags for example through incremental prediction of the pos tag being generated or to exploit the use of word lemmas effectively splitting wordword relations into syntactic dependencies and more semantic dependencies	P99-1065	a statistical parser for czechthis paper considers statistical parsing of czech which differs radically from english in at least two respects it is a highly inflected language and it has relatively free word orderthese differences are likely to pose new problems for techniques that have been developed on englishwe describe our experience in building on the parsing model of our final results 80 dependency accuracy represent good progress towards the 91 accuracy of the parser on english textwe use a transformed tree bank from the prague dependency treebank for constituent parsing on czech
automatic identification of word translations from unrelated english and german corpora algorithms for the alignment of words in translated texts are well established however only recently new approaches have been proposed to identify word translations from nonparallel or even unrelated texts this task is more difficult because most statistical clues useful in the processing of parallel texts cannot be applied to nonparallel texts whereas for parallel texts in some studies up to 99 of the word alignments have been shown to be correct the accuracy for nonparallel texts has been around 30 up to now the current study which is based on the assumption that there is a correlation between the patterns of word cooccurrences in corpora of different languages makes a significant improvement to about 72 of word translations identified correctly starting with the wellknown paper of brown et al on statistical machine translation there has been much scientific interest in the alignment of sentences and words in translated textsmany studies show that for nicely parallel corpora high accuracy rates of up to 99 can be achieved for both sentence and word alignment of course in practice due to omissions transpositions insertions and replacements in the process of translation with real texts there may be all kinds of problems and therefore robustness is still an issue nevertheless the results achieved with these algorithms have been found useful for the cornpilation of dictionaries for checking the consistency of terminological usage in translations for assisting the terminological work of translators and interpreters and for examplebased machine translationby now some alignment programs are offered commercially translation memory tools for translators such as ibm translation manager or trados translator workbench are bundled or can be upgraded with programs for sentence alignmentmost of the proposed algorithms first conduct an alignment of sentences that is they locate those pairs of sentences that are translations of each otherin a second step a word alignment is performed by analyzing the correspondences of words in each pair of sentencesthe algorithms are usually based on one or several of the following statistical clues all these clues usually work well for parallel textshowever despite serious efforts in the compilation of parallel corpora the availability of a largeenough parallel corpus in a specific domain and for a given pair of languages is still an exceptionsince the acquisition of monolingual corpora is much easier it would be desirable to have a program that can determine the translations of words from comparable or possibly unrelated monolingual texts of two languagesthis is what translators and interpreters usually do when preparing terminology in a specific field they read texts corresponding to this field in both languages and draw their conclusions on word correspondences from the usage of the termsof course the translators and interpreters can understand the texts whereas our programs are only considering a few statistical cluesfor nonparallel texts the first clue which is usually by far the strongest of the three mentioned above is not applicable at allthe second clue is generally less powerful than the first since most words are ambiguous in natural languages and many ambiguities are different across languagesnevertheless this clue is applicable in the case of comparable texts although with a lower reliability than for parallel textshowever in the case of unrelated texts its usefulness may be near zerothe third clue is generally limited to the identification of word pairs with similar spellingfor all other pairs it is usually used in combination with the first cluesince the first clue does not work with nonparallel texts the third clue is useless for the identification of the majority of pairsfor unrelated languages it is not applicable anywayin this situation rapp proposed using a clue different from the three mentioned above his cooccurrence clue is based on the assumption that there is a correlation between cooccurrence patterns in different languagesfor example if the words teacher and school cooccur more often than expected by chance in a corpus of english then the german translations of teacher and school lehrer and schule should also cooccur more often than expected in a corpus of germanin a feasibility study he showed that this assumption actually holds for the language pair englishgerman even in the case of unrelated textswhen comparing an english and a german cooccurrence matrix of corresponding words he found a high correlation between the cooccurrence patterns of the two matrices when the rows and columns of both matrices were in corresponding word order and a low correlation when the rows and columns were in random orderthe validity of the cooccurrence clue is obvious for parallel corpora but as empirically shown by rapp it also holds for nonparallel corporait can be expected that this clue will work best with parallel corpora secondbest with comparable corpora and somewhat worse with unrelated corporain all three cases the problem of robustness as observed when applying the wordorder clue to parallel corpora is not severetranspositions of text segments have virtually no negative effect and omissions or insertions are not criticalhowever the cooccurrence clue when applied to comparable corpora is much weaker than the wordorder clue when applied to parallel corpora so larger corpora and wellchosen statistical methods are requiredafter an attempt with a context heterogeneity measure for identifying word translations fung based her later work also on the cooccurrence assumption by presupposing a lexicon of seed words she avoids the prohibitively expensive computational effort encountered by rapp the method described here although developed independently of fung work goes in the same directionconceptually it is a trivial case of rapp matrix permutation methodby simply assuming an initial lexicon the large number of permutations to be considered is reduced to a much smaller number of vector comparisonsthe main contribution of this paper is to describe a practical implementation based on the cooccurrence clue that yields good resultsas mentioned above it is assumed that across languages there is a correlation between the cooccurrences of words that are translations of each otherif for example in a text of one language two words a and b cooccur more often than expected by chance then in a text of another language those words that are translations of a and b should also cooccur more frequently than expectedthis is the only statistical clue used throughout this paperit is further assumed that there is a small dictionary available at the beginning and that our aim is to expand this base lexiconusing a corpus of the target language we first compute a cooccurrence matrix whose rows are all word types occurring in the corpus and whose columns are all target words appearing in the base lexiconwe now select a word of the source language whose translation is to be determinedusing our sourcelanguage corpus we compute a cooccurrence vector for this wordwe translate all known words in this vector to the target languagesince our base lexicon is small only some of the translations are knownall unknown words are discarded from the vector and the vector positions are sorted in order to match the vectors of the targetlanguage matrixwith the resulting vector we now perform a similarity computation to all vectors in the cooccurrence matrix of the target languagethe vector with the highest similarity is considered to be the translation of our sourcelanguage wordto conduct the simulation a number of resources were requiredthese are as the german corpus we used 135 million words of the newspaper frankfurter allgemeine zeitung and as the english corpus 163 million words of the guardian since the orientation of the two newspapers is quite different and since the time spans covered are only in part overlapping the two corpora can be considered as more or less unrelatedfor testing our results we started with a list of 100 german test words as proposed by russell which he used for an association experiment with german subjectsby looking up the translations for each of these 100 words we obtained a test set for evaluationour germanenglish base lexicon is derived from the collins gem german dictionary with about 22300 entriesfrom this we eliminated all multiword entries so 16380 entries remainedbecause we had decided on our test word list beforehand and since it would not make much sense to apply our method to words that are already in the base lexicon we also removed all entries belonging to the 100 test wordssince our corpora are very large to save disk space and processing time we decided to remove all function words from the textsthis was done on the basis of a list of approximately 600 german and another list of about 200 english function wordsthese lists were compiled by looking at the closed class words in an english and a german morphological lexicon and at word frequency lists derived from our corporaby eliminating function words we assumed we would lose little information function words are often highly ambiguous and their cooccurrences are mostly based on syntactic instead of semantic patternssince semantic patterns are more reliable than syntactic patterns across language families we hoped that eliminating the function words would give our method more generalitywe also decided to lemmatize our corporasince we were interested in the translations of base forms only it was clear that lemmatization would be usefulit not only reduces the sparsedata problem but also takes into account that german is a highly inflectional language whereas english is notfor both languages we conducted a partial lemmatization procedure that was based only on a morphological lexicon and did not take the context of a word form into accountthis means that we could not lemmatize those ambiguous word forms that can be derived from more than one base formhowever this is a relatively rare casealthough we had a contextsensitive lemmatizer for german available this was not the case for english so for reasons of symmetry we decided not to use the context featurei in cases in which an ambiguous word can be both a content and a function word preference was given to those interpretations that appeared to occur more frequentlyfor counting word cooccurrences in most other studies a fixed window size is chosen and it is determined how often each pair of words occurs within a text window of this sizehowever this approach does not take word order within a window into accountsince it has been empirically observed that word order of content words is often similar between languages and since this may be a useful statistical clue we decided to modify the common approach in the way proposed by rapp instead of computing a single cooccurrence vector for a word a we compute several one for each position within the windowfor example if we have chosen the window size 2 we would compute a first cooccurrence vector for the case that word a is two words ahead of another word b a second vector for the case that word a is one word ahead of word b a third vector for a directly following b and a fourth vector for a following two words after bif we added up these four vectors the result would be the cooccurrence vector as obtained when not taking word order into accounthowever this is not what we doinstead we combine the four vectors of length n into a single vector of length 4nsince preliminary experiments showed that a window size of 3 with consideration of word order seemed to give somewhat better results than other window types the results reported here are based on vectors of this kindhowever the computational methods described below are in the same way applicable to window sizes of any length with or without consideration of word orderour method is based on the assumption that there is a correlation between the patterns of word cooccurrences in texts of different languageshowever as rapp proposed this correlation may be strengthened by not using the cooccurrence counts directly but association strengths between words insteadthe idea is to eliminate wordfrequency effects and to emphasize significant word pairs by comparing their observed cooccurrence counts with their expected cooccurrence countsin the past for this purpose a number of measures have been proposedthey were based on mutual information conditional probabilities or on some standard statistical tests such as the chisquare test or the loglikelihood ratio for the purpose of this paper we decided to use the loglikelihood ratio which is theoretically well justified and more appropriate for sparse data than chisquarein preliminary experiments it also led to slightly better results than the conditional probability measureresults based on mutual information or cooccurrence counts were significantly worsefor efficient computation of the loglikelihood ratio we used the following formula2 where with parameters kj expressed in terms of corpus frequencies kli frequency of common occurrence of word a and word b k22 size of corpus corpus frequency of a corpus frequency of b all cooccurrence vectors were transformed using this formulathereafter they were normalized in such a way that for each vector the sum of its entries adds up to onein the rest of the paper we refer to the transformed and normalized vectors as association vectorsto determine the english translation of an unknown german word the association vector of the german word is computed and compared to all association vectors in the english association matrixfor comparison the correspondences between the vector positions and the columns of the matrix are determined by using the base lexiconthus for each vector in the english matrix a similarity value is computed and the english words are ranked according to these valuesit is expected that the correct translation is ranked first in the sorted listfor vector comparison different similarity measures can be consideredsalton mcgill proposed a number of measures such as the cosine coefficient the jaccard coefficient and the dice coefficient for the computation of related terms and synonyms ruge landauer and dumais and fung and mckeown used the cosine measure whereas grefenstette used a weighted jaccard measurewe propose here the cityblock metric which computes the similarity between two vectors x and y as the sum of the absolute differences of corresponding vector positions in a number of experiments we compared it to other similarity measures such as the cosine measure the jaccard measure the euclidean distance and the scalar product and found that the cityblock metric yielded the best resultsthis may seem surprising since the formula is very simple and the computational effort smaller than with the other measuresit must be noted however that the other authors applied their similarity measures directly to the cooccurrence vectors whereas we applied the measures to the association vectors based on the loglikelihood ratioaccording to our observations estimates based on the loglikelihood ratio are generally more reliable across different corpora and languagesthe results reported in the next section were obtained using the following procedure 1based on the word cooccurrences in the german corpus for each of the 100 german test words its association vector was computedin these vectors all entries belonging to words not found in the english part of the base lexicon were deleted2based on the word cooccurrences in the english corpus an association matrix was computed whose rows were all word types of the corpus with a frequency of 100 or higher3 and whose columns were all english words occurring as first translations of the german words in the base lexicon3using the similarity function each of the german vectors was compared to all vectors of the english matrixthe mapping between vector positions was based on the first translations given in the base lexiconfor each of the german source words the english vocabulary was ranked according to the resulting similarity value3 the limitation to words with frequencies above 99 was introduced for computational reasons to reduce the number of vector comparisons and thus speed up the programthe purpose of this limitation was not to limit the number of translation candidates consideredexperiments with lower thresholds showed that this choice has little effect on the results to our set of test words4 this means that alternative translations of a word were not consideredanother approach as conducted by fung yee would be to consider all possible translations listed in the lexicon and to give them equal weightour decision was motivated by the observation that many words have a salient first translation and that this translation is listed first in the collins gem dictionary germanenglishwe did not explore this issue further since in a small pocket dictionary only few ambiguities are listedtable 1 shows the results for 20 of the 100 german test wordsfor each of these test words the top five translations as automatically generated are listedin addition for each word its expected english translation from the test set is given together with its position in the ranked lists of computed translationsthe positions in the ranked lists are a measure for the quality of the predictions with a 1 meaning that the prediction is correct and a high value meaning that the program was far from predicting the correct wordif we look at the table we see that in many cases the program predicts the expected word with other possible translations immediately followingfor example for the german word hauschen the correct translations bungalow cottage house and hut are listedin other cases typical associates follow the correct translationfor example the correct translation of madchen girl is followed by boy man brother and ladythis behavior can be expected from our associationist approachunfortunately in some cases the correct translation and one of its strong associates are mixed up as for example with frau where its correct translation woman is listed only second after its strong associate mananother example of this typical kind of error is pfeifen where the correct translation whistle is listed third after linesman and refereelet us now look at some cases where the program did particularly badlyfor kohl we had expected its dictionary translation cabbage but given that a substantial part of our newspaper corpora consists of political texts we do not need to further explain why our program lists major kohl thatcher gorbachev and bush state leaders who were in office during the time period the texts were writtenin other cases such as krankheit and whisky the simulation program simply preferred the british usage of the guardian over the american usage in our test set instead of sickness the program predicted disease and illness and instead of whiskey it predicted whiskya much more severe problem is that our current approach cannot properly handle ambiguities for the german word weifi it does not predict white but instead knowthe reason is that weifi can also be third person singular of the german verb wissen which in newspaper texts is more frequent than the color whitesince our lemmatizer is not contextsensitive this word was left unlemmatized which explains the resultto be able to compare our results with other work we also did a quantitative evaluationfor all test words we checked whether the predicted translation was identical to our expected translationthis was true for 65 of the 100 test wordshowever in some cases the choice of the expected translation in the test set had been somewhat arbitraryfor example for the german word strafie we had expected street but the system predicted road which is a translation quite as goodtherefore as a better measure for the accuracy of our system we counted the number of times where an acceptable translation of the source word is ranked firstthis was true for 72 of the 100 test words which gives us an accuracy of 72in another test we checked whether an acceptable translation appeared among the top 10 of the ranked liststhis was true in 89 cases5 for comparison fung mckeown report an accuracy of about 30 when only the top candidate is countedhowever it must be emphasized that their result has been achieved under very different circumstanceson the one hand their task was more difficult because they worked on a pair of unrelated languages using smaller corpora and a random selection of test words many of which were multiword termsalso they predetermined a single translation as being correcton the other hand when conducting their evaluation fung mckeown limited the vocabulary they considered as translation candidates to a few hundred terms which obviously facilitates the taskthe method described can be seen as a simple case of the gradient descent method proposed by rapp which does not need an initial lexicon but is computationally prohibitively expensiveit can also be considered as an extension from the monolingual to the bilingual case of the wellestablished methods for semantic or syntactic word clustering as proposed by schiitze grefenstette ruge rapp lin and otherssome of these authors perform a shallow or full syntactical analysis before constructing the cooccurrence vectorsothers reduce the size of the cooccurrence matrices by performing a singular value decompositionhowever in yet unpublished work we found that at least for the computation of synonyms and related words neither syntactical analysis nor singular value decomposition lead to significantly better results than the approach described here when applied to the monolingual case so we did not try to include these methods in our systemnevertheless both methods are of technical value since they lead to a reduction in the size of the cooccurrence matricesfuture work has to approach the difficult problem of ambiguity resolution which has not been dealt with hereone possibility would be to semantically disambiguate the words in the corpora beforehand another to look at cooccurrences between significant word sequences instead of cooccurrences between single wordsto conclude with let us add some speculation by mentioning that the ability to identify word translations from nonparallel texts can be seen as an indicator in favor of the associationist view of human language acquisition it gives us an idea of how it is possible to derive the meaning of unknown words from texts by only presupposing a limited number of known words and then iteratively expanding this knowledge baseone possibility to get the process going would be to learn vocabulary lists as in school another to simply acquire the names of items in the physical worldi thank manfred wettler gisela zunkerrapp wolfgang lezius and anita todd for their support of this work	P99-1067	automatic identification of word translations from unrelated english and german corporaalgorithms for the alignment of words in translated texts are well establishedhowever only recently new approaches have been proposed to identify word translations from nonparallel or even unrelated textsthis task is more difficult because most statistical clues useful in the processing of parallel texts cannot be applied to nonparallel textswhereas for parallel texts in some studies up to 99 of the word alignments have been shown to be correct the accuracy for nonparallel texts has been around 30 up to nowthe current study which is based on the assumption that there is a correlation between the patterns of word cooccurrences in corpora of different languages makes a significant improvement to about 72 of word translations identified correctlywe create bagofwords context vectors around both the source and target language words and then project the source into the english target space via the current small translation dictionarywe filter out bilingual term pairs with low monolingual frequencies we show that accurate translations can be learned for 100 german nouns that are not contained in the seed bilingual dictionary
mining the web for bilingual text strand is a languageindependent system for automatic discovery of text in parallel translation on the world wide web this paper extends the preliminary strand results by adding automatic language identification scaling up by orders of magnitude and formally evaluating performance the most recent endproduct is an automatically acquired parallel corpus comprising 2491 englishfrench document pairs approximately 15 million words per language text in parallel translation is a valuable resource in natural language processingstatistical methods in machine translation typically rely on large quantities of bilingual text aligned at the document or sentence level and a number of approaches in the burgeoning field of crosslanguage information retrieval exploit parallel corpora either in place of or in addition to mappings between languages based on information from bilingual dictionaries despite the utility of such data however sources of bilingual text are subject to such limitations as licensing restrictions usage fees restricted domains or genres and dated text or such sources simply may not exist for language pairs of interestalthough the majority of web content is in english it also shows great promise as a source of multilingual contentusing figures from the babel survey of multilinguality on the web it is possible to estimate that as of june 1997 there were on the order of 63000 primarily nonenglish web servers ranging over 14 languagesmoreover a followup investigation of the nonenglish servers suggests that nearly a third contain some useful crosslanguage data such as parallel english on the page or links to parallel english pages the followup also found pages in five languages not identified by the babel study given the continued explosive increase in the size of the web the trend toward business organizations that cross national boundaries and high levels of competition for consumers in a global marketplace it seems impossible not to view multilingual content on the web as an expanding resourcemoreover it is a dynamic resource changing in content as the world changesfor example diekema et al in a presentation at the 1998 trec7 conference observed that the performance of their crosslanguage information retrieval was hurt by lexical gaps such as bosnial bosnie this illustrates a highly topical missing pair in their static lexical resource and gey et al also at trec7 observed that in doing crosslanguage retrieval using commercial machine translation systems gaps in the lexicon could make the difference between precision of 008 and precision of 083 on individual queriesresnik presented an algorithm called strand designed to explore the web as a source of parallel text demonstrating its potential with a smallscale evaluation based on the author judgmentsafter briefly reviewing the strand architecture and preliminary results this paper goes beyond that preliminary work in two significant waysfirst the framework is extended to include a filtering stage that uses automatic language identification to eliminate an important class of false positives documents that appear structurally to be parallel translations but are in fact not in the languages of interestthe system is then run on a somewhat larger scale and evaluated formally for english and spanish using measures of agreement with independent human judges precision and recall second the algorithm is scaled up more seriously to generate large numbers of parallel documents this time for english and french and again subjected to formal evaluation the concrete end result reported here is an automatically acquired englishfrench parallel corpus of web documents comprising 2491 document pairs approximately 15 million words per language containing little or no noisethis section is a brief summary of the strand system and previously reported preliminary results the strand architecture is organized as a pipeline beginning with a candidate generation stage that generates candidate pairs of documents that might be parallel translationsthe first implementation of the generation stage used a query to the altavista search engine to generate pages that could be viewed as quotparentsquot of pages in parallel translation by asking for pages containing one portion of anchor text containing the string quotenglishquot within a fixed distance of another anchor text containing the string quotspanishquotthis generated many good pairs of pages such as those pointed to by hyperlinks reading click here for english version and click here for spanish version as well as many bad pairs such as university pages containing links to english literature in close proximity to spanish literaturethe candidate generation stage is followed by a candidate evaluation stage that represents the core of the approach filtering out bad candidates from the set of generated page pairsit employs a structural recognition algorithm exploiting the fact that web pages in parallel translation are invariably very similar in the way they are structured hence the in strandfor example see figure 2the structural recognition algorithm first runs both documents through a transducer that reduces each to a linear sequence of tokens corresponding to html markup elements interspersed with tokens representing undifferentiated quotchunksquot of textfor example the transducer would replace the html source text acl 99 conference home page with the three tokens begin title chunk24 and end titlethe number inside the chunk token is the length of the text chunk not counting whitespace from this point on only the length of the text chunks is used and therefore the structural filtering algorithm is completely language independentgiven the transducer output for each document the structural filtering stage aligns the two streams of tokens by applying a standard widely available dynamic programming algorithm for finding an optimal alignment between two linear sequences1 this alignment matches identical markup tokens to each other as much as possible identifies runs of unmatched tokens that appear to exist only in one sequence but not the other and marks pairs of nonidentical tokens that were forced to be matched to each other in order to obtain the best alignment posknown to many programmers as different1111111111111110111111 le vmdredi 25 cacao 199640 membres nit de la seemananon oat assisth au stninaiste sr les psalm exemplars en maniac de reglentation qui visait i aida la rnininbto 6 familiarises avec pereglemsitation cannina de rechange asa progrunes de rtglanentasionlanimanurm zane bmw airedmp gannet direction 4061 de wasoonnation owes la mance en mktit aphasia dtindusthe caned pe rune de introduction de nif animus de diversification 496 6 a anation da services comma les coda yokota l snort lonstion de lindastrie a add que pseud mochainement un document stu i divenifirstse des modes de postai da service qui intrait de divers suj comas les anfmnisnes de premvion de rethange des semi les phi aggroprifs que a problemes leafs par nunn cada volontalms us code volontaire un ensemble dengageme scernaliffs ne faisant pas eaplidtanent pink dun mgirrio logialatif ou reglementaire cone0 pom lateen faxonnex conneller ou israluer is canpartement de ceux qui la ont pan iiistiliminent pax a pnnsuivi 69 gins analyse thincipalaffair ag6 au racentariat du conseil du treeceadroit du gaivemement de makatea n alma simplanent aux mmiciparas une solution de red i reglemenmtioe p gomernernmat au moment a is thglementation f objes true examen adru du public ks gouvornemerns i lishelle town van des approchen volonmirsat en 519014 a la rigkoausion et mans comae submits i celled les codes volontaires prima40 tin cumin menthe manages notanstent sible2 at this point if there were too many unmatched tokens the candidate pair is taken to be prima facie unacceptable and immediately filtered outotherwise the algorithm extracts from the alignment those pairs of chunk tokens that were matched to each other in order to obtain the best alignments3 it then computes the correlation between the lengths of these nonmarkup text chunksas is well known there is a reliably linear relationship in the lengths of text translations small pieces of source text translate to small pieces of target text medium to medium and large to largetherefore we can apply a standard statistical hypothesis test and if p pr and pr prfor english and spanish this translates as a simple requirement that the quotenglishquot page look more like english than spanish and that the quotspanishquot page look more like spanish than englishlanguage identification is performed on the plaintext versions of the pagescharacter 5gram models for languages under consideration are constructed using 100k characters of training data from the european corpus initiative available from the linguistic data consortium in a formal evaluation strand with the new language identification stage was run for english and spanish starting from the top 1000 hits yielded up by altavista in the candidate generation stage leading to a set of 913 candidate pairsa test set of 179 items was generated for annotation by human judges containing it was impractical to manually evaluate all pairs filtered out structurally owing to the time required for judgments and the desire for two independent judgments per pair in order to assess interjudge reliabilitythe two judges were both native speakers of spanish with high proficiency in english neither previously familiar with the projectthey worked independently using a web browser to access test pairs in a fashion that allowed them to view pairs side by sidethe judges were told they were helping to evaluate a system that identifies pages on the web that are translations of each other and were instructed to make decisions according to the following criterion is this pair of pages intended to show the same material to two different users one a reader of english and the other a reader of spanishthe phrasing of the criterion required some consideration since in previous experience with human judges and translations i have found that judges are frequently unhappy with the quality of the translations they are looking atfor present purposes it was required neither that the document pair represent a perfect translation nor even necessarily a good one strand was being tested not on its ability to determine translation quality which might or might not be a criterion for inclusion in a parallel corpus but rather its ability to facilitate the task of locating page pairs that one might reasonably include in a corpus undifferentiated by quality the judges were permitted three responses when computing evaluation measures page pairs classified in the third category by a human judge for whatever reason were excluded from considerationtable 1 shows agreement measures between the two judges between strand and each individual judge and the agreement between strand and the intersection of the two judges annotations that is strand evaluated against only those cases where the two judges agreed which are therefore the items we can regard with the highest confidencethe table also shows cohen lc an agreement measure that corrects for chance agreement the most important value in the table is the value of 07 for the two human judges which can be interpreted as sufficiently high to indicate that the task is reasonably well defined for every language l in that range and that d2 meet the corresponding requirement for frenchdoing so leads to the results in table 3this translates into an estimated 100 precision against 641 recall with a yield of 2491 documents approximately 15 million words per language as counted after removal of html markupthat is with a reasonable though admittedly posthoc revision of the language identification criterion comparison with human subjects suggests the acquired corpus is nontrivial and essentially noise free and moreover that the system excludes only a third of the pages that should have been keptnaturally this will need to be verified in a new evaluation on fresh data6language id across a wide range of languages is not difficult to obtaineg see the 13language set of the freely available cmu stochastic language identhis paper places acquisition of parallel text from the web on solid empirical footing making a number of contributions that go beyond the preliminary studythe system has been extended with automated language identification and scaled up to the point where a nontrivial parallel corpus of english and french can be produced completely automatically from the world wide webin the process it was discovered that the most lightweight use of language identification restricted to just the the language pair of interest needed to be revised in favor of a strategy that includes identification over a wide range of languagesrigorous evaluation using human judges suggests that the technique produces an extremely clean corpus noise estimated at between 0 and 8 even without human intervention requiring no more resources per language than a relatively small sample of text used to train automatic language identificationtwo directions for future work are apparentfirst experiments need to be done using languages that are less common on the weblikely first pairs to try include englishkorean englishitalian and englishgreekinspection of web sites those with bilingual text identified by strand and those without suggests that the strategy of using altavista to generate candidate pairs could be improved upon significantly by adding a true web crawler to quotminequot sites where bilingual text is known to be available eg sites uncovered by a first pass of the system using the altavista enginei would conjecture that for englishfrench there is an order of magnitude more bilingual text on the web than that uncovered in this early stage of researcha second natural direction is the application of webbased parallel text in applications such as lexical acquisition and crosslanguage information retrieval especially since a sideeffect of the core strand algorithm is aligned quotchunksquot ie nonmarkup segments found to correspond to each other based on alignment of the markuppreliminary experiments using even small amounts of these data suggest that standard techniques such as crosslanguage lexical association can uncover useful data	P99-1068	mining the web for bilingual textstrand is a languageindependent system for automatic discovery of text in parallel translation on the world wide webthis paper extends the preliminary strand results by adding automatic language identification scaling up by orders of magnitude and formally evaluating performancethe most recent endproduct is an automatically acquired parallel corpus comprising 2491 englishfrench document pairs approximately 15 million words per languagewe use structure markup information from pages without looking at their content to attempt to align them
estimators for stochastic unificationbased grammars loglinear models provide a statistically sound framework for stochastic quotunificationbasedquot grammars and stochastic versions of other kinds of grammars we describe two computationallytractable ways of estimating the parameters of such grammars from a training corpus of syntactic analyses and apply these to estimate a stochastic version of lexical probabilistic methods have revolutionized computational linguisticsthey can provide a systematic treatment of preferences in parsinggiven a suitable estimation procedure stochastic models can be quottunedquot to reflect the properties of a corpuson the other hand quotunificationbasedquot grammars can express a variety of linguisticallyimportant syntactic and semantic constraintshowever developing stochastic quotunificationbasedquot grammars has not proved as straightforward as might be hopedthe simple quotrelative frequencyquot estimator for pcfgs yields the maximum likelihood parameter estimate which is to say that it minimizes the kulbackliebler divergence between the training and estimated distributionson the other hand as abney points out the contextsensitive dependencies that quotunificationbasedquot constraints introduce render the relative frequency estimator suboptimal in general it does not maximize the likelihood and it is inconsistentabney proposes a markov random field or log linear model for subgs and the models described here are instances of abney general frameworkhowever the montecarlo parameter estimation procedure that abney proposes seems to be computationally impractical for reasonablesized grammarssections 3 and 4 describe two new estimation procedures which are computationally tractablesection 5 describes an experiment with a small lfg corpus provided to us by xerox parcthe log linear framework and the estimation procedures are extremely general and they apply directly to stochastic versions of hpsg and other theories of grammarwe follow the statistical literature in using the term feature to refer to the properties that parameters are associated with let sz be the set of all possible grammatical or wellformed analyseseach feature f maps a syntactic analysis to e s to a real value f the form of a syntactic analysis depends on the underlying linguistic theoryfor example for a pcfg co would be parse tree for a lfg co would be a tuple consisting of a cstructure an fstructure and a mapping from cstructure nodes to fstructure elements and for a chomskyian transformational grammar co would be a derivationloglinear models are models in which the log probability is a linear combination of feature values pcfgs gibbs distributions maximumentropy distributions and markov random fields are all examples of loglinear modelsa loglinear model associates each feature f3 with a realvalued parameter 03a loglinear model with m features is one in which the likelihood p of an analysis w is while the estimators described below make no assumptions about the range of the f in the models considered here the value of each feature f2 is the number of times a particular structural arrangement or configuration occurs in the analysis ranges over the natural numbersfor example the features of a pcfg are indexed by productions ie the value f of feature f is the number of times the ith production is used in the derivation w this set of features induces a treestructured dependency graph on the productions which is characteristic of markov branching processes this tree structure has the important consequence that simple quotrelativefrequenciesquot yield maximumlikelihood estimates for the 02extending a pcfg model by adding additional features not associated with productions will in general add additional dependencies destroy the tree structure and substantially complicate maximum likelihood estimationthis is the situation for a subg even if the features are production occurencesthe unification constraints create nonlocal dependencies among the productions and the dependency graph of a subg is usually not a treeconsequently maximum likelihood estimation is no longer a simple matter of computing relative frequenciesbut the resulting estimation procedures albeit more complicated have the virtue of applying to essentially arbitrary featuresof the production or nonproduction typethat is since estimators capable of finding maximumlikelihood parameter estimates for production features in a subg will also find maximumlikelihood estimates for nonproduction features there is no motivation for restricting features to be of the production typelinguistically there is no particular reason for assuming that productions are the best features to use in a stochastic language modelfor example the adjunct attachment ambiguity in results in alternative syntactic structures which use the same productions the same number of times in each derivation so a model with only production features would necessarily assign them the same likelihoodthus models that use production features alone predict that there should not be a systematic preference for one of these analyses over the other contrary to standard psycholinguistic resultsthere are many different ways of choosing features for a subg and each of these choices makes an empirical claim about possible distributions of sentencesspecifying the features of a subg is as much an empirical matter as specifying the grammar itselffor any given ubg there are a large number of subgs that can be constructed from it differing only in the features that each subg usesin addition to production features the stochastic lfg models evaluated below used the following kinds of features guided by the principles proposed by hobbs and bear adjunct and argument features indicate adjunct and argument attachment respectively and permit the model to capture a general argument attachment preferencein addition there are specialized adjunct and argument features corresponding to each grammatical function used in lfg there are features indicating both high and low attachment another feature indicates nonrightbranching nonterminal nodesthere is a feature for nonparallel coordinate structures each fstructure attributeatomic value pair which appears in any feature structure is also used as a featurewe also use a number of features identifying syntactic structures that seem particularly important in these corpora such as a feature identifying nps that are dates we would have liked to have included features concerning specific lexical items but we felt that our corpora were so small that the associated parameters could not be accurately estimatedsuppose wi wr is a training corpus of n syntactic analysesletting f3 f3 poi the log likelihood of the corpus cz and its derivatives are where e0 is the expected value of h under the distribution determined by the parameters 0the maximumlikelihood estimates are the 0 which maximize log 14 the chief difficulty in finding the maximumlikelihood estimates is calculating e9 which involves summing over the space of wellformed syntactic structures a there seems to be no analytic or efficient numerical way of doing this for a realistic subgabney proposes a gradient ascent based upon a monte carlo procedure for estimating eothe idea is to generate random samples of feature structures from the distribution po where 0 is the current parameter estimate and to use these to estimate eo and hence the gradient of the likelihoodsamples are generated as follows given a subg abney constructs a covering pcfg based upon the subg and 0 the current estimate of 0the derivation trees of the pcfg can be mapped onto a set containing all of the subg syntactic analysesmonte carlo samples from the pcfg are comparatively easy to generate and sample syntactic analyses that do not map to wellformed subg syntactic structures are then simply discardedthis generates a stream of syntactic structures but not distributed according to p abney proposes using a metropolis acceptancerejection method to adjust the distribution of this stream of feature structures to achieve detailed balance which then produces a stream of feature structures distributed according to pei while this scheme is theoretically sound it would appear to be computationally impractical for realistic subgsevery step of the proposed procedure requires a very large number of pcfg samples samples must be found that correspond to wellformed subgs many such samples are required to bring the metropolis algorithm to equilibrium many samples are needed at equilibrium to properly estimate ethe idea of a gradient ascent of the likelihood is appealinga simple calculation reveals that the likelihood is concave and therefore free of local maximabut the gradient is intractablethis motivates an alternative strategy involving a databased estimate of e9 where y is the yield belonging to the syntactic analysis w and yi y is the yield belonging to the ith sample in the training corpusthe point is that eoly yi is generally computablein fact if sz is the set of wellformed syntactic structures that have yield y then a l expectations only involves summing over the possible syntactic analyses or parses 12 of the strings in the training corpuswhile it is possible to construct ubgs for which the number of possible parses is unmanageably high for many grammars it is quite manageable to enumerate the set of possible parses and thereby directly evaluate eo iy yi therefore we propose replacing the gradient by and performing a gradient ascentof course is no longer the gradient of the likelihood function but fortunately it is the gradient of another criterion instead of maximizing the likelihood of the syntactic analyses over the training corpus we maximize the conditional likelihood of these analyses given the observed yieldsin our experiments we have used a conjugategradient optimization program adapted from the one presented in press et al regardless of the pragmatic motivation one could perhaps argue that the conditional probabilities po are as useful as the full probabilities po at least in those cases for which the ultimate goal is syntactic analysisberger et al and jelinek make this same point and arrive at the same estimator albeit through a maximum entropy argumentthe problem of estimating parameters for loglinear models is not newit is especially difficult in cases such as ours where a large sample space makes the direct computation of expectations infeasiblemany applications in spatial statistics involving markov random fields are of this nature as wellin his seminal development of the mrf approach to spatial statistics besag introduced a quotpseudolikelihoodquot estimator to address these difficulties and in fact our proposal here is an instance of his methodin general the likelihood function is replaced by a more manageable product of conditional likelihoods what are the asymptotics of optimizing a pseudolikelihood functionlook first at the likelihood itselffor large n where 00 is the true parameter vectorup to a constant is the negative of the kullbackleibler divergence between the true and estimated distributions of syntactic analysesas sample size grows maximizing likelihood amounts to minimizing divergenceas for pseudolikelihood so that maximizing pseudolikelihood amounts to minimizing the average divergence between the true and estimated conditional distributions of analyses given yieldsmaximum likelihood estimation is consistent under broad conditions the sequence of distributions po associated with the maximum likelihood estimator for 00 given the samples con converges to poopseudolikelihood is also consistent but in the present implementation it is consistent for the conditional distributions poo and not necessarily for the full distribution poo it is not hard to see that pseudolikelihood will not always correctly estimate poosuppose there is a feature l which depends only on yields f2 fzin this case the derivative of plo contains no information about 0in fact in this case any value of 0i gives the same conditional distribution po y 0i is irrelevant to the problem of choosing good parsesdespite the assurance of consistency pseudolikelihood estimation is prone to over fitting when a large number of features is matched against a modestsized training corpusone particularly troublesome manifestation of over fitting results from the existence of features which relative to the training set we might term quotpseudomaximalquot let us say that a feature f is pseudomaximal for a yield y if vwf e 1 f f where w is any correct parse of y ie the feature value on every correct parse of y is greater than or equal to its value on any other parse of y pseudominimal features are defined similarlyit is easy to see that if h is pseudomaximal on each sentence of the training corpus then the parameter assignment 03 oo maximizes the corpus pseudolikelihoodsuch infinite parameter values indicate that the model treats pseudomaximal features categorically ie any parse with a nonmaximal feature value is assigned a zero conditional probabilityof course a feature which is pseudomaximal over the training corpus is not necessarily pseudomaximal for all yieldsthis is an instance of over fitting and it can be addressed as is customary by adding a regularization term that promotes small values of 0 to the objective functiona common choice is to add a quadratic to the loglikelihood which corresponds to multiplying the likelihood itself by a normal distributionin our experiments we multiplied the pseudolikelihood by a zeromean normal in 01 0 with diagonal covariance and with standard deviation ay for 03 equal to 7 times the maximum value of fi found in any parse in the training corpusthus instead of maximizing the log pseudolikelihood we choose 0 to maxithe pseudolikelihood estimator described in the last section finds parameter values which maximize the conditional probabilities of the observed parses given the observed sentences in the training corpusone of the empirical evaluation measures we use in the next section measures the number of correct parses selected from the set of all possible parsesthis suggests another possible objective function choose 0 to maximize the number co of times the maximum likelihood parse is in fact the correct parse in the training corpusc9 is a highly discontinuous function of 0 and most conventional optimization algorithms perform poorly on itwe had the most success with a slightly modified version of the simulated annealing optimizer described in press et al this procedure is much more computationally intensive than the gradientbased pseudolikelihood procedureits computational difficulty grows rapidly with the number of featuresron kaplan and hadar shemtov at xerox parc provided us with two lfg parsed corporathe verbmobil corpus contains appointment planning dialogs while the homecentre corpus is drawn from xerox printer documentationtable 1 summarizes the basic properties of these corporathese corpora contain packed cfstructure representations of the grammatical parses of each sentence with respect to lexicalfunctional grammarsthe corpora also indicate which of these parses is in fact the correct parse because slightly different grammars were used for each corpus we chose not to combine the two corpora although we used the set of features described in section 2 for both in the experiments described belowtable 2 describes the properties of the features used for each corpusin addition to the two estimators described above we also present results from a baseline estimator in which all parses are treated as equally likely we evaluated our estimators using heldout test corpus test we used two evaluation measuresin an actual parsing application a subg might be used to identify the correct parse from the set of grammatical parses so ouorfitrestevaluation measure counts the number c of sentences in the test corpus cahest whose maximum likelihood parse under the estimated model 0 is actually the correct parseif a sentence has 1 most likely parses and one of these parses is the correct parse then we score 1 for this sentencethe second evaluation measure is the pseudolikelihood of the test corpus is the likelihood of the correct parses given their yields so pseudolikelihood measures how much of the probability mass the model puts onto the correct analysesthis metric seems more relevant to applications where the system needs to estimate how likely it is that the correct analysis lies in a certain set of possible parses eg ambiguitypreserving translation and humanassisted disambiguationto make the numbers more manageable we actually present the negative logarithm of the pseudolikelihood rather than the pseudolikelihood itselfso smaller is betterbecause of the small size of our corpora we evaluated our estimators using a 10way crossvalidation paradigmwe randomly assigned sentences of each corpus into 10 approximately equalsized subcorpora each of which was used in turn as the test corpuswe evaluated on each subcorpus the parameters that were estimated from the 9 remaining subcorpora that served as the training corpus for this runthe evaluation scores from each subcorpus were summed in order to provide the scores presented heretable 3 presents the results of the empirical evaluationthe superior performance of both estimators on the verbmobil corpus probably reflects the fact that the nonrule features were designed to match both the grammar and content of that corpusthe pseudolikelihood estimator performed better than the correctparses estimator on both corpora under both evaluation metricsthere seems to be substantial over learning in all these models we routinely improved performance by discarding featureswith a small number of features the correctparses estimator typically scores better than the pseudolikelihood estimator on the correctparses evaluation metric but the pseudolikelihood estimator always scores better on the pseudolikelihood evaluation metricthis paper described a loglinear model for subgs and evaluated two estimators for such modelsbecause estimators that can estimate rule features for subgs can also estimate other kinds of features there is no particular reason to limit attention to rule features in a subgindeed the number and choice of features strongly influences the performance of the modelthe estimated models are able to identify the correct parse from the set of all possible parses approximately 50 of the timewe would have liked to introduce features corresponding to dependencies between lexical itemsloglinear models are wellsuited for lexical dependencies but because of the large number of such dependencies substantially larger corpora will probably be needed to estimate such mo dels 1 alternatively it may be possible to use a simpler nonsubg model of lexical dependencies estimated from a much larger corpus as the reference distribution with parses of the test corpus that were the correct parses and log pl is the negative logarithm of the pseudolikelihood of the test corpushowever there may be applications which can benefit from a model that performs even at this levelfor example in a machineassisted translation system a model like ours could be used to order possible translations so that more likely alternatives are presented before less likely onesin the ambiguitypreserving translation framework a model like this one could be used to choose between sets of analyses whose ambiguities cannot be preserved in translation	P99-1069	estimators for stochastic unificationbased grammarsloglinear models provide a statistically sound framework for stochastic unificationbased grammars and stochastic versions of other kinds of grammarswe describe two computationallytractable ways of estimating the parameters of such grammars from a training corpus of syntactic analyses and apply these to estimate a stochastic version of lexicalfunctional grammarwe incorporate general linguistic principles into a loglinear modelwe use parses generated by a lfg parser as input to a mrf approach