Dataset Viewer
Auto-converted to Parquet
preprocessed_text
stringlengths
43
152k
preprocessed_summary
stringlengths
55
3.02k
tnt - a statistical part-of-speech tagger trigrams'n'tags ( tnt ) is an efficient statistical part-of-speech tagger . contrary to claims found elsewhere in the literature , we argue that a tagger based on markov models performs at least as well as other current approaches , including the maximum entropy framework . a recent comparison has even shown that tnt performs significantly better for the tested corpora . we describe the basic model of tnt , the techniques used for smoothing and for handling unknown words . furthermore , we present evaluations on two corpora . a large number of current language processing systems use a part-of-speech tagger for pre-processing . the tagger assigns a ( unique or ambiguous ) part-ofspeech tag to each token in the input and passes its output to the next processing level , usually a parser . furthermore , there is a large interest in part-ofspeech tagging for corpus annotation projects , who create valuable linguistic resources by a combination of automatic processing and human correction . for both applications , a tagger with the highest possible accuracy is required . the debate about which paradigm solves the part-of-speech tagging problem best is not finished . recent comparisons of approaches that can be trained on corpora ( van halteren et al. , 1998 ; volk and schneider , 1998 ) have shown that in most cases statistical aproaches ( cutting et al. , 1992 ; schmid , 1995 ; ratnaparkhi , 1996 ) yield better results than finite-state , rule-based , or memory-based taggers ( brill , 1993 ; daelemans et al. , 1996 ) . they are only surpassed by combinations of different systems , forming a & quot ; voting tagger & quot ; . among the statistical approaches , the maximum entropy framework has a very strong position . nevertheless , a recent independent comparison of 7 taggers ( zavrel and daelemans , 1999 ) has shown that another approach even works better : markov models combined with a good smoothing technique and with handling of unknown words . this tagger , tnt , not only yielded the highest accuracy , it also was the fastest both in training and tagging . the tagger comparison was organized as a & quot ; blackbox test & quot ; : set the same task to every tagger and compare the outcomes . this paper describes the models and techniques used by tnt together with the implementation . the reader will be surprised how simple the underlying model is . the result of the tagger comparison seems to support the maxime & quot ; the simplest is the best & quot ; . however , in this paper we clarify a number of details that are omitted in major previous publications concerning tagging with markov models . as two examples , ( rabiner , 1989 ) and ( charniak et al. , 1993 ) give good overviews of the techniques and equations used for markov models and part-ofspeech tagging , but they are not very explicit in the details that are needed for their application . we argue that it is not only the choice of the general model that determines the result of the tagger but also the various & quot ; small & quot ; decisions on alternatives . the aim of this paper is to give a detailed account of the techniques used in tnt . additionally , we present results of the tagger on the negra corpus ( brants et al. , 1999 ) and the penn treebank ( marcus et al. , 1993 ) . the penn treebank results reported here for the markov model approach are at least equivalent to those reported for the maximum entropy approach in ( ratnaparkhi , 1996 ) . for a comparison to other taggers , the reader is referred to ( zavrel and daelemans , 1999 ) . tnt uses second order markov models for part-ofspeech tagging . the states of the model represent tags , outputs represent the words . transition probabilities depend on the states , thus pairs of tags . output probabilities only depend on the most recent category . to be explicit , we calculate for a given sequence of words w1 of length t. t1 tr are elements of the tagset , the additional tags t_1 , to , and t7- , ±1 are beginning-of-sequence and end-of-sequence markers . using these additional tags , even if they stem from rudimentary processing of punctuation marks , slightly improves tagging results . this is different from formulas presented in other publications , which just stop with a & quot ; loose end & quot ; at the last word . if sentence boundaries are not marked in the input , tnt adds these tags if it encounters one of [ . ! ? ; ] as a token . transition and output probabilities are estimated from a tagged corpus . as a first step , we use the maximum likelihood probabilities p which are derived from the relative frequencies : for all t1 , t2 , t3 in the tagset and w3 in the lexicon . n is the total number of tokens in the training corpus . we define a maximum likelihood probability to be zero if the corresponding nominators and denominators are zero . as a second step , contextual frequencies are smoothed and lexical frequences are completed by handling words that are not in the lexicon ( see below ) . trigram probabilities generated from a corpus usually can not directly be used because of the sparsedata problem . this means that there are not enough instances for each trigram to reliably estimate the probability . furthermore , setting a probability to zero because the corresponding trigram never occured in the corpus has an undesired effect . it causes the probability of a complete sequence to be set to zero if its use is necessary for a new text sequence , thus makes it impossible to rank different sequences containing a zero probability . the smoothing paradigm that delivers the best results in tnt is linear interpolation of unigrams , bigrams , and trigrams . therefore , we estimate a trigram probability as follows : p are maximum likelihood estimates of the probabilities , and a1 + a2 ± a3 = 1 , so p again represent probability distributions . we use the context-independent variant of linear interpolation , i.e. , the values of the as do not depend on the particular trigram . contrary to intuition , this yields better results than the context-dependent variant . due to sparse-data problems , one can not estimate a different set of as for each trigram . therefore , it is common practice to group trigrams by frequency and estimate tied sets of as . however , we are not aware of any publication that has investigated frequency groupings for linear interpolation in part-of-speech tagging . all groupings that we have tested yielded at most equivalent results to contextindependent linear interpolation . some groupings even yielded worse results . the tested groupings included a ) one set of as for each frequency value and b ) two classes ( low and high frequency ) on the two ends of the scale , as well as several groupings in between and several settings for partitioning the classes . the values of a1 , a2 , and a3 are estimated by deleted interpolation . this technique successively removes each trigram from the training corpus and estimates best values for the as from all other ngrams in the corpus . given the frequency counts for uni- , bi- , and trigrams , the weights can be very efficiently determined with a processing time linear in the number of different trigrams . the algorithm is given in figure 1 . note that subtracting 1 means taking unseen data into account . without this subtraction the model would overfit the training data and would generally yield worse results . currently , the method of handling unknown words that seems to work best for inflected languages is a suffix analysis as proposed in ( samuelsson , 1993 ) . tag probabilities are set according to the word 's ending . the suffix is a strong predictor for word classes , e.g. , words in the wall street journal part of the penn treebank ending in able are adjectives ( .11 ) in 98 % of the cases ( e.g . fashionable , variable ) , the rest of 2 % are nouns ( e.g . cable , variable ) . the probability distribution for a particular suffix is generated from all words in the training set that share the same suffix of some predefined maximum length . the term suffix as used here means & quot ; final sequence of characters of a word & quot ; which is not necessarily a linguistically meaningful suffix . probabilities are smoothed by successive abstraction . this calculates the probability of a tag t given the last m letters i of an n letter word : p ( t1/7 „ , +1 , , .. ln ) . the sequence of increasingly more general contexts omits more and more characters of the suffix , such that p ( tlin-m+2 , • • • p ( tlin_m+3 , , i ) , , p ( t ) are used for smoothing . the recursion formula is set a = a2 = a3 = 0 foreach trigram t1 , t2 , t3 with f ( ti , t2 , t3 ) > 0 depending on the maximum of the following three values : for i = m ... 0 , using the maximum likelihood estimates p from frequencies in the lexicon , weights oi and the initialization for the markov model , we need the inverse conditional probabilities p ( 1,2_1+1 , ... /tilt ) which are obtained by bayesian inversion . a theoretical motivated argumentation uses the standard deviation of the maximum likelihood probabilities for the weights 0 , ( samuelsson , 1993 ) . this leaves room for interpretation . we use the longest suffix that we can find in the training set ( i.e. , for which the frequency is greater than or equal to 1 ) , but at most 10 characters . this is an empirically determined choice . 2 ) we use a context-independent approach for 0 „ as we did for the contextual weights a . it turned out to be a good choice to set all 0 , to the standard deviation of the unconditioned maximum likelihood probabilities of the tags in the training corpus , i.e. , we set for all i = 0 ... m — 1 , using a tagset of s tags and the average ( 11 ) this usually yields values in the range 0,03 ... 0.10 . 3 ) we use different estimates for uppercase and lowercase words , i.e. , we maintain two different suffix tries depending on the capitalization of the word . this information improves the tagging results . 4 ) another freedom concerns the choice of the words in the lexicon that should be used for suffix handling . should we use all words , or are some of them better suited than others ? accepting that unknown words are most probably infrequent , one can argue that using suffixes of infrequent words in the lexicon is a better approximation for unknown words than using suffixes of frequent words . therefore , we restrict the procedure of suffix handling to words with a frequency smaller than or equal to some threshold value . empirically , 10 turned out to be a good choice for this threshold . additional information that turned out to be useful for the disambiguation process for several corpora and tagsets is capitalization information . tags are usually not informative about capitalization , but probability distributions of tags around capitalized words are different from those not capitalized . the effect is larger for english , which only capitalizes proper names , and smaller for german , which capitalizes all nouns . we use flags ci that are true if wi is a capitalized word and false otherwise . these flags are added to the contextual probability distributions . instead of and equations ( 3 ) to ( 5 ) are updated accordingly . this is equivalent to doubling the size of the tagset and using different tags depending on capitalization . the processing time of the viterbi algorithm ( rabiner , 1989 ) can be reduced by introducing a beam search . each state that receives a 6 value smaller than the largest 6 divided by some threshold value 0 is excluded from further processing . while the viterbi algorithm is guaranteed to find the sequence of states with the highest probability , this is no longer true when beam search is added . nevertheless , for practical purposes and the right choice of 0 , there is virtually no difference between the algorithm with and without a beam . empirically , a value of 0 = 1000 turned out to approximately double the speed of the tagger without affecting the accuracy . the tagger currently tags between 30,000 and 60,000 tokens per second ( including file i/o ) on a pentium 500 running linux . the speed mainly depends on the percentage of unknown words and on the average ambiguity rate . we evaluate the tagger 's performance under several aspects . first of all , we determine the tagging accuracy averaged over ten iterations . the overall accuracy , as well as separate accuracies for known and unknown words are measured . second , learning curves are presented , that indicate the performance when using training corpora of different sizes , starting with as few as 1,000 tokens and ranging to the size of the entire corpus ( minus the test set ) . an important characteristic of statistical taggers is that they not only assign tags to words but also probabilities in order to rank different assignments . we distinguish reliable from unreliable assignments by the quotient of the best and second best assignmentsl . all assignments for which this quotient is larger than some threshold are regarded as reliable , the others as unreliable . as we will see below , accuracies for reliable assignments are much higher . the tests are performed on partitions of the corpora that use 90 % as training set and 10 % as test set , so that the test data is guaranteed to be unseen during training . each result is obtained by repeating the experiment 10 times with different partitions and averaging the single outcomes . in all experiments , contiguous test sets are used . the alternative is a round-robin procedure that puts every 10th sentence into the test set . we argue that contiguous test sets yield more realistic results because completely unseen articles are tagged . using the round-robin procedure , parts of an article are already seen , which significantly reduces the percentage of unknown words . therefore , we expect even 'by definition , this quotient is oo if there is only one possible tag for a given word . higher results when testing on every 10th sentence instead of a contiguous set of 10 % . in the following , accuracy denotes the number of correctly assigned tags divided by the number of tokens in the corpus processed . the tagger is allowed to assign exactly one tag to each token . we distinguish the overall accuracy , taking into account all tokens in the test corpus , and separate accuracies for known and unknown tokens . the latter are interesting , since usually unknown tokens are much more difficult to process than known tokens , for which a list of valid tags can be found in the lexicon . the german negra corpus consists of 20,000 sentences ( 355,000 tokens ) of newspaper texts ( frankfurter rundschau ) that are annotated with parts-ofspeech and predicate-argument structures ( skut et al. , 1997 ) . it was developed at the saarland university in saarbriicken2 . part of it was tagged at the ims stuttgart . this evaluation only uses the partof-speech annotation and ignores structural annotations . tagging accuracies for the negra corpus are shown in table 2 . figure 3 shows the learning curve of the tagger , i.e. , the accuracy depending on the amount of training data . training length is the number of tokens used for training . each training length was tested ten times , training and test sets were randomly chosen and disjoint , results were averaged . the training length is given on a logarithmic scale . it is remarkable that tagging accuracy for known words is very high even for very small training corpora . this means that we have a good chance of getting the right tag if a word is seen at least once during training . average percentages of unknown tokens are shown in the bottom line of each diagram . we exploit the fact that the tagger not only determines tags , but also assigns probabilities . if there is an alternative that has a probability & quot ; close to & quot ; that of the best assignment , this alternative can be viewed as almost equally well suited . the notion of & quot ; close to & quot ; is expressed by the distance of probabilities , and this in turn is expressed by the quotient of probabilities . so , the distance of the probabilities of a best tag tbest and an alternative tag tau is expressed by p ( tbest ) /p ( talt ) 7 which is some value greater or equal to 1 since the best tag assignment has the highest probability . figure 4 shows the accuracy when separating assignments with quotients larger and smaller than the threshold ( hence reliable and unreliable assignments ) . as expected , we find that accuracies for percentage known unknown • overall unknowns acc . acc . acc . a table 5 : part-of-speech tagging accuracy for the penn treebank . the table shows the percentage of unknown tokens , separate accuracies and standard deviations for known and unknown tokens , as well as the overall accuracy . percentage known unknown overall unknowns acc . acc . acc . reliable assignments are much higher than for unreliable assignments . this distinction is , e.g. , useful for annotation projects during the cleaning process , or during pre-processing , so the tagger can emit multiple tags if the best tag is classified as unreliable . we use the wall street journal as contained in the penn treebank for our experiments . the annotation consists of four parts : 1 ) a context-free structure augmented with traces to mark movement and discontinuous constituents , 2 ) phrasal categories that are annotated as node labels , 3 ) a small set of grammatical functions that are annotated as extensions to the node labels , and 4 ) part-of-speech tags ( marcus et al. , 1993 ) . this evaluation only uses the part-ofspeech annotation . the wall street journal part of the penn treebank consists of approx . 50,000 sentences ( 1.2 million tokens ) . tagging accuracies for the penn treebank are shown in table 5 . figure 6 shows the learning curve of the tagger , i.e. , the accuracy depending on the amount of training data . training length is the number of tokens used for training . each training length was tested ten times . training and test sets were disjoint , results are averaged . the training length is given on a logarithmic scale . as for the negra corpus , tagging accuracy is very high for known tokens even with small amounts of training data . we exploit the fact that the tagger not only determines tags , but also assigns probabilities . figure 7 shows the accuracy when separating assignments with quotients larger and smaller than the threshold ( hence reliable and unreliable assignments ) . again , we find that accuracies for reliable assignments are much higher than for unreliable assignments . average part-of-speech tagging accuracy is between 96 % and 97 % , depending on language and tagset , which is at least on a par with state-of-the-art results found in the literature , possibly better . for the penn treebank , ( ratnaparkhi , 1996 ) reports an accuracy of 96.6 % using the maximum entropy approach , our much simpler and therefore faster hmm approach delivers 96.7 % . this comparison needs to be re-examined , since we use a ten-fold crossvalidation and averaging of results while ratnaparkhi only makes one test run . the accuracy for known tokens is significantly higher than for unknown tokens . for the german newspaper data , results are 8.7 % better when the word was seen before and therefore is in the lexicon , than when it was not seen before ( 97.7 % vs. 89.0 % ) . accuracy for known tokens is high even with very small amounts of training data . as few as 1000 tokens are sufficient to achieve 95 % -96 % accuracy for them . it is important for the tagger to have seen a word at least once during training . stochastic taggers assign probabilities to tags . we exploit the probabilities to determine reliability of assignments . for a subset that is determined during processing by the tagger we achieve accuracy rates of over 99 % . the accuracy of the complement set is much lower . this information can , e.g. , be exploited in an annotation project to give an additional treatment to the unreliable assignments , or to pass selected ambiguities to a subsequent processing step . we have shown that a tagger based on markov models yields state-of-the-art results , despite contrary claims found in the literature . for example , the markov model tagger used in the comparison of ( van halteren et al. , 1998 ) yielded worse results than all other taggers . in our opinion , a reason for the wrong claim is that the basic algorithms leave several decisions to the implementor . the rather large amount of freedom was not handled in detail in previous publications : handling of start- and end-of-sequence , the exact smoothing technique , how to determine the weights for context probabilities , details on handling unknown words , and how to determine the weights for unknown words . note that the decisions we made yield good results for both the german and the english corpus . they do so for several other corpora as well . the architecture remains applicable to a large variety of languages . according to current tagger comparisons ( van halteren et al. , 1998 ; zavrel and daelemans , 1999 ) , and according to a comparsion of the results presented here with those in ( ratnaparkhi , 1996 ) , the maximum entropy framework seems to be the only other approach yielding comparable results to the one presented here . it is a very interesting future research topic to determine the advantages of either of these approaches , to find the reason for their high accuracies , and to find a good combination of both . tnt is freely available to universities and related organizations for research purposes ( see http : //www.coli.uni-sb.derthorstenant ) . many thanks go to hans uszkoreit for his support during the development of tnt . most of the work on tnt was carried out while the author received a grant of the deutsche forschungsgemeinschaft in the graduiertenkolleg kognitionswissenschaft saarbriicken . large annotated corpora are the pre-requisite for developing and testing part-ofspeech taggers , and they enable the generation of high-quality language models . therefore , i would like to thank all the people who took the effort to annotate the penn treebank , the susanne corpus , the stuttgarter referenzkorpus , the negra corpus , the verbmobil corpora , and several others . and , last but not least , i would like to thank the users of tnt who provided me with bug reports and valuable suggestions for improvements .
tnt - a statistical part-of-speech tagger trigrams'n'tags ( tnt ) is an efficient statistical part-of-speech tagger . contrary to claims found elsewhere in the literature , we argue that a tagger based on markov models performs at least as well as other current approaches , including the maximum entropy framework . a recent comparison has even shown that tnt performs significantly better for the tested corpora . we describe the basic model of tnt , the techniques used for smoothing and for handling unknown words . furthermore , we present evaluations on two corpora . we achieve the automated tagging of a syntactic-structure-based set of grammatical function tags including phrase-chunk and syntactic-role modifiers trained in supervised mode from a tree bank of german .
mildly non-projective dependency structures syntactic parsing requires a fine balance between expressivity and complexity , so that naturally occurring structures can be accurately parsed without compromising efficiency . in dependency-based parsing , several constraints have been proposed that restrict the class of permissible structures , such as projectivity , planarity , multi-planarity , well-nestedness , gap degree , and edge degree . while projectivity is generally taken to be too restrictive for natural language syntax , it is not clear which of the other proposals strikes the best balance between expressivity and complexity . in this paper , we review and compare the different constraints theoretically , and provide an experimental evaluation using data from two treebanks , investigating how large a proportion of the structures found in the treebanks are permitted under different constraints . the results indicate that a combination of the well-nestedness constraint and a parametric constraint on discontinuity gives a very good fit with the linguistic data . dependency-based representations have become increasingly popular in syntactic parsing , especially for languages that exhibit free or flexible word order , such as czech ( collins et al. , 1999 ) , bulgarian ( marinov and nivre , 2005 ) , and turkish ( eryi˘git and oflazer , 2006 ) . many practical implementations of dependency parsing are restricted to projective structures , where the projection of a head word has to form a continuous substring of the sentence . while this constraint guarantees good parsing complexity , it is well-known that certain syntactic constructions can only be adequately represented by non-projective dependency structures , where the projection of a head can be discontinuous . this is especially relevant for languages with free or flexible word order . however , recent results in non-projective dependency parsing , especially using data-driven methods , indicate that most non-projective structures required for the analysis of natural language are very nearly projective , differing only minimally from the best projective approximation ( nivre and nilsson , 2005 ; hall and novák , 2005 ; mcdonald and pereira , 2006 ) . this raises the question of whether it is possible to characterize a class of mildly non-projective dependency structures that is rich enough to account for naturally occurring syntactic constructions , yet restricted enough to enable efficient parsing . in this paper , we review a number of proposals for classes of dependency structures that lie between strictly projective and completely unrestricted non-projective structures . these classes have in common that they can be characterized in terms of properties of the dependency structures themselves , rather than in terms of grammar formalisms that generate the structures . we compare the proposals from a theoretical point of view , and evaluate a subset of them empirically by testing their representational adequacy with respect to two dependency treebanks : the prague dependency treebank ( pdt ) ( hajiˇc et al. , 2001 ) , and the danish dependency treebank ( ddt ) ( kromann , 2003 ) . the rest of the paper is structured as follows . in section 2 , we provide a formal definition of dependency structures as a special kind of directed graphs , and characterize the notion of projectivity . in section 3 , we define and compare five different constraints on mildly non-projective dependency structures that can be found in the literature : planarity , multiplanarity , well-nestedness , gap degree , and edge degree . in section 4 , we provide an experimental evaluation of the notions of planarity , well-nestedness , gap degree , and edge degree , by investigating how large a proportion of the dependency structures found in pdt and ddt are allowed under the different constraints . in section 5 , we present our conclusions and suggestions for further research . for the purposes of this paper , a dependency graph is a directed graph on the set of indices corresponding to the tokens of a sentence . we write [ n ] to refer to the set of positive integers up to and including n. throughout this paper , we use standard terminology and notation from graph theory to talk about dependency graphs . in particular , we refer to the elements of the set v as nodes , and to the elements of the set e as edges . we write i -- > j to mean that there is an edge from the node i to the node j ( i.e. , ( i , j ) e e ) , and i -- > * j to mean that the node i dominates the node j , i.e. , that there is a ( possibly empty ) path from i to j . for a given node i , the set of nodes dominated by i is the yield of i . we use the notation 3r ( i ) to refer to the projection of i : the yield of i , arranged in ascending order . most of the literature on dependency grammar and dependency parsing does not allow arbitrary dependency graphs , but imposes certain structural constraints on them . in this paper , we restrict ourselves to dependency graphs that form forests . definition 2 a dependency forest is a dependency graph with two additional properties : figure 1 shows a dependency forest taken from pdt . it has two roots : node 2 ( corresponding to the complementizer proto ) and node 8 ( corresponding to the final punctuation mark ) . some authors extend dependency forests by a special root node with position 0 , and add an edge ( 0 , i ) for every root node i of the remaining graph ( mcdonald et al. , 2005 ) . this ensures that the extended graph always is a tree . although such a definition can be useful , we do not follow it here , since it obscures the distinction between projectivity and planarity to be discussed in section 3 . in contrast to acyclicity and the indegree constraint , both of which impose restrictions on the dependency relation as such , the projectivity constraint concerns the interaction between the dependency relation and the positions of the nodes in the sentence : it says that the nodes in a subtree of a dependency graph must form an interval , where an interval ( with endpoints i and j ) is the set [ i , j ] : = { kev i i < k and k < j } . definition 3 a dependency graph is projective , if the yields of its nodes are intervals . since projectivity requires each node to dominate a continuous substring of the sentence , it corresponds to a ban on discontinuous constituents in phrase structure representations . projectivity is an interesting constraint on dependency structures both from a theoretical and a practical perspective . dependency grammars that only allow projective structures are closely related to context-free grammars ( gaifman , 1965 ; obre¸bski and grali´nski , 2004 ) ; among other things , they have the same ( weak ) expressivity . the projectivity constraint also leads to favourable parsing complexities : chart-based parsing of projective dependency grammars can be done in cubic time ( eisner , 1996 ) ; hard-wiring projectivity into a deterministic dependency parser leads to linear-time parsing in the worst case ( nivre , 2003 ) . while the restriction to projective analyses has a number of advantages , there is clear evidence that it can not be maintained for real-world data ( zeman , 2004 ; nivre , 2006 ) . for example , the graph in figure 1 is non-projective : the yield of the node 1 ( marked by the dashed rectangles ) does not form an interval—the node 2 is ‘ missing ’ . in this section , we present several proposals for structural constraints that relax projectivity , and relate them to each other . the notion of planarity appears in work on link grammar ( sleator and temperley , 1993 ) , where it is traced back to mel ’ ˇcuk ( 1988 ) . informally , a dependency graph is planar , if its edges can be drawn above the sentence without crossing . we emphasize the word above , because planarity as it is understood here does not coincide with the standard graph-theoretic concept of the same name , where one would be allowed to also use the area below the sentence to disentangle the edges . figure 2a shows a dependency graph that is planar but not projective : while there are no crossing edges , the yield of the node 1 ( the set 11 , 3 } ) does not form an interval . using the notation linked ( i , j ) as an abbreviation for the statement ‘ there is an edge from i to j , or vice versa ’ , we formalize planarity as follows : definition 4 a dependency graph is planar , if it does not contain nodes a , b , c , d such that linked ( a , c ) a linked ( b , d ) a a < b < c < d . yli-jyrä ( 2003 ) proposes multiplanarity as a generalization of planarity suitable for modelling dependency analyses , and evaluates it experimentally using data from ddt . definition 5 a dependency graph g = ( v ; e ) is m-planar , if it can be split into m planar graphs such that e = e1u- - -uem . the planar graphs gi are called planes . as an example of a dependency forest that is 2planar but not planar , consider the graph depicted in figure 2b . in this graph , the edges ( 1 , 4 ) and ( 3 , 5 ) are crossing . moving either edge to a separate graph partitions the original graph into two planes . bodirsky et al . ( 2005 ) present two structural constraints on dependency graphs that characterize analyses corresponding to derivations in tree adjoining grammar : the gap degree restriction and the well-nestedness constraint . a gap is a discontinuity in the projection of a node in a dependency graph ( plátek et al. , 2001 ) . more precisely , let 7ri be the projection of the node i . then a gap is a pair ( jk , jk+1 ) of nodes adjacent in 7ri such that definition 6 the gap degree of a node i in a dependency graph , gd ( i ) , is the number of gaps in 7ri . as an example , consider the node labelled i in the dependency graphs in figure 3 . in graph 3a , the projection of i is an interval ( ( 2 , 3 , 4 ) ) , so i has gap degree 0 . in graph 3b , 7ri = ( 2 , 3 , 6 ) contains a single gap ( ( 3 , 6 ) ) , so the gap degree of i is 1 . in the rightmost graph , the gap degree of i is 2 , since 7ri = ( 2 , 4 , 6 ) contains two gaps ( ( 2 , 4 ) and ( 4 , 6 ) ) . definition 7 the gap degree of a dependency graph g , gd ( g ) , is the maximum among the gap degrees of its nodes . thus , the gap degree of the graphs in figure 3 is 0 , 1 and 2 , respectively , since the node i has the maximum gap degree in all three cases . the well-nestedness constraint restricts the positioning of disjoint subtrees in a dependency forest . two subtrees are called disjoint , if neither of their roots dominates the other . definition 8 two subtrees t1 , t2 interleave , if there are nodes l1 , r1 e t1 and l2 , r2 e t2 such that l1 < l2 < r1 < r2 . a dependency graph is well-nested , if no two of its disjoint subtrees interleave . both graph 3a and graph 3b are well-nested . graph 3c is not well-nested . to see this , let t1 be the subtree rooted at the node labelled i , and let t2 be the subtree rooted at j . these subtrees interleave , as t1 contains the nodes 2 and 4 , and t2 contains the nodes 3 and 5 . the notion of edge degree was introduced by nivre ( 2006 ) in order to allow mildly non-projective structures while maintaining good parsing efficiency in data-driven dependency parsing.2 define the span of an edge ( i , j ) as the interval s ( ( i , j ) ) w= [ min ( i , j ) , max ( i , j ) ] . definition 9 let g = ( v i e ) be a dependency forest , let e = ( i , j ) be an edge in e , and let ge be the subgraph of g that is induced by the nodes contained in the span of e. • the degree of an edge e 2 e , ed ( e ) , is the number of connected components c in ge such that the root of c is not dominated by the head of e. • the edge degree of g , ed ( g ) , is the maximum among the degrees of the edges in g. to illustrate the notion of edge degree , we return to figure 3 . graph 3a has edge degree 0 : the only edge that spans more nodes than its head and its dependent is ( 1 , 5 ) , but the root of the connected component f2 , 3 , 4g is dominated by 1 . both graph 3b and 3c have edge degree 1 : the edge ( 3 , 6 ) in graph 3b and the edges ( 2 , 4 ) , ( 3 , 5 ) and ( 4 , 6 ) in graph 3c each span a single connected component that is not dominated by the respective head . apart from proposals for structural constraints relaxing projectivity , there are dependency frameworks that in principle allow unrestricted graphs , but provide mechanisms to control the actually permitted forms of non-projectivity in the grammar . the non-projective dependency grammar of kahane et al . ( 1998 ) is based on an operation on dependency trees called lifting : a ‘ lift ’ of a tree t is the new tree that is obtained when one replaces one 2we use the term edge degree instead of the original simple term degree from nivre ( 2006 ) to mark the distinction from the notion of gap degree . or more edges ( i , k ) in t by edges ( j , k ) , where j ! * i . the exact conditions under which a certain lifting may take place are specified in the rules of the grammar . a dependency tree is acceptable , if it can be lifted to form a projective graph.3 a similar design is pursued in topological dependency grammar ( duchier and debusmann , 2001 ) , where a dependency analysis consists of two , mutually constraining graphs : the id graph represents information about immediate dominance , the lp graph models the topological structure of a sentence . as a principle of the grammar , the lp graph is required to be a lift of the id graph ; this lifting can be constrained in the lexicon . the structural conditions we have presented here naturally fall into two groups : multiplanarity , gap degree and edge degree are parametric constraints with an infinite scale of possible values ; planarity and well-nestedness come as binary constraints . we discuss these two groups in turn . parametric constraints with respect to the graded constraints , we find that multiplanarity is different from both gap degree and edge degree in that it involves a notion of optimization : since every dependency graph is m-planar for some sufficiently large m ( put each edge onto a separate plane ) , the interesting question in the context of multiplanarity is about the minimal values for m that occur in real-world data . but then , one not only needs to show that a dependency graph can be decomposed into m planar graphs , but also that this decomposition is the one with the smallest number of planes among all possible decompositions . up to now , no tractable algorithm to find the minimal decomposition has been given , so itis not clear how to evaluate the significance of the concept as such . the evaluation presented by yli-jyrä ( 2003 ) makes use of additional constraints that are sufficient to make the decomposition unique . the fundamental difference between gap degree and edge degree is that the gap degree measures the number of discontinuities within a subtree , while the edge degree measures the number of intervening constituents spanned by a single edge . this difference is illustrated by the graphs displayed in figure 4 . graph 4a has gap degree 2 but edge degree 1 : the subtree rooted at node 2 ( marked by the solid edges ) has two gaps , but each of its edges only spans one connected component not dominated by 2 ( marked by the squares ) . in contrast , graph 4b has gap degree 1 but edge degree 2 : the subtree rooted at node 2 has one gap , but this gap contains two components not dominated by 2 . nivre ( 2006 ) shows experimentally that limiting the permissible edge degree to 1 or 2 can reduce the average parsing time for a deterministic algorithm from quadratic to linear , while omitting less than 1 % of the structures found in ddt and pdt . it can be expected that constraints on the gap degree would have very similar effects . binary constraints for the two binary constraints , we find that well-nestedness subsumes planarity : a graph that contains interleaving subtrees can not be drawn without crossing edges , so every planar graph must also be well-nested . to see that the converse does not hold , consider graph 3b , which is well-nested , but not planar . since both planarity and well-nestedness are proper extensions of projectivity , we get the following hierarchy for sets of dependency graphs : projective c planar c well-nested c unrestricted the planarity constraint appears like a very natural one at first sight , as it expresses the intuition that ‘ crossing edges are bad ’ , but still allows a limited form of non-projectivity . however , many authors use planarity in conjunction with a special representation of the root node : either as an artificial node at the sentence boundary , as we mentioned in section 2 , or as the target of an infinitely long perpendicular edge coming ‘ from the outside ’ , as in earlier versions of word grammar ( hudson , 2003 ) . in these situations , planarity reduces to projectivity , so nothing is gained . even in cases where planarity is used without a special representation of the root node , it remains a peculiar concept . when we compare it with the notion of gaps , for example , we find that , in a planar dependency tree , every gap .i ; j/ must contain the root node r , in the sense that i < r < j : if the gap would only contain non-root nodes k , then the two paths from r to k and from i to j would cross . this particular property does not seem to be mirrored in any linguistic prediction . in contrast to planarity , well-nestedness is independent from both gap degree and edge degree in the sense that for every d > 0 , there are both wellnested and non-well-nested dependency graphs with gap degree or edge degree d. all projective dependency graphs ( d = 0 ) are trivially well-nested . well-nestedness also brings computational benefits . in particular , chart-based parsers for grammar formalisms in which derivations obey the well-nestedness constraint ( such as tree adjoining grammar ) are not hampered by the ‘ crossing configurations ’ to which satta ( 1992 ) attributes the fact that the universal recognition problem of linear context-free rewriting systems is x30-complete . in this section , we present an experimental evaluation of planarity , well-nestedness , gap degree , and edge degree , by examining how large a proportion of the structures found in two dependency treebanks are allowed under different constraints . assuming that the treebank structures are sampled from naturally occurring structures in natural language , this provides an indirect evaluation of the linguistic adequacy of the different proposals . the experiments are based on data from the prague dependency treebank ( pdt ) ( hajiˇc et al. , 2001 ) and the danish dependency treebank ( ddt ) ( kromann , 2003 ) . pdt contains 1.5m words of newspaper text , annotated in three layers according to the theoretical framework of functional generative description ( böhmová et al. , 2003 ) . our experiments concern only the analytical layer , and are based on the dedicated training section of the treebank . ddt comprises 100k words of text selected from the danish parole corpus , with annotation property all structures gap degree 0 gap degree 1 gap degree 2 gap degree 3 gap degree 4 edge degree 0 edge degree 1 edge degree 2 edge degree 3 edge degree 4 edge degree 5 edge degree 6 projective planar well-nested of primary and secondary dependencies based on discontinuous grammar ( kromann , 2003 ) . only primary dependencies are considered in the experiments , which are based on the entire treebank.4 the results of our experiments are given in table 1 . for the binary constraints ( planarity , well-nestedness ) , we simply report the number and percentage of structures in each data set that satisfy the constraint . for the parametric constraints ( gap degree , edge degree ) , we report the number and percentage of structures having degree d ( d > 0 ) , where degree 0 is equivalent ( for both gap degree and edge degree ) to projectivity . for ddt , we see that about 15 % of all analyses are non-projective . the minimal degree of non-projectivity required to cover all of the data is 2 in the case of gap degree and 4 in the case of edge degree . for both measures , the number of structures drops quickly as the degree increases . ( as an example , only 7 or 0.17 % of the analyses in ddt have gap 4a total number of 17 analyses in ddt were excluded because they either had more than one root node , or violated the indegree constraint . ( both cases are annotation errors . ) degree 2 . ) regarding the binary constraints , we find that planarity accounts for slightly more than the projective structures ( 86.41 % of the data is planar ) , while almost all structures in ddt ( 99.89 % ) meet the well-nestedness constraint . the difference between the two constraints becomes clearer when we base the figures on the set of non-projective structures only : out of these , less than 10 % are planar , while more than 99 % are well-nested . for pdt , both the number of non-projective structures ( around 23 % ) and the minimal degrees of non-projectivity required to cover the full data ( gap degree 4 and edge degree 6 ) are higher than in ddt . the proportion of planar analyses is smaller than in ddt if we base it on the set of all structures ( 82.16 % ) , but significantly larger when based on the set of non-projective structures only ( 22.93 % ) . however , this is still very far from the well-nestedness constraint , which has almost perfect coverage on both data sets . as a general result , our experiments confirm previous studies on non-projective dependency parsing ( nivre and nilsson , 2005 ; hall and novák , 2005 ; mcdonald and pereira , 2006 ) : the phenomenon of non-projectivity can not be ignored without also ignoring a significant portion of real-world data ( around 15 % for ddt , and 23 % for pdt ) . at the same time , already a small step beyond projectivity accounts for almost all of the structures occurring in these treebanks . more specifically , we find that already an edge degree restriction of d < 1 covers 98.24 % of ddt and 99.54 % of pdt , while the same restriction on the gap degree scale achieves a coverage of 99.84 % ( ddt ) and 99.57 % ( pdt ) . together with the previous evidence that both measures also have computational advantages , this provides a strong indication for the usefulness of these constraints in the context of non-projective dependency parsing . when we compare the two graded constraints to each other , we find that the gap degree measure partitions the data into less and larger clusters than the edge degree , which may be an advantage in the context of using the degree constraints as features in a data-driven approach towards parsing . however , our purely quantitative experiments can not answer the question , which of the two measures yields the more informative clusters . the planarity constraint appears to be of little use as a generalization of projectivity : enforcing it excludes more than 75 % of the non-projective data in pdt , and 90 % of the data in ddt . the relatively large difference in coverage between the two treebanks may at least partially be explained with their different annotation schemes for sentence-final punctuation . in ddt , sentence-final punctuation marks are annotated as dependents of the main verb of a dependency nexus . this , as we have discussed above , places severe restrictions on permitted forms of non-projectivity in the remaining sentence , as every discontinuity that includes the main verb must also include the dependent punctuation marks . on the other hand , in pdt , a sentencefinal punctuation mark is annotated as a separate root node with no dependents . this scheme does not restrict the remaining discontinuities at all . in contrast to planarity , the well-nestedness constraint appears to constitute a very attractive extension of projectivity . for one thing , the almost perfect coverage of well-nestedness on ddt and pdt ( 99.89 % ) could by no means be expected on purely combinatorial grounds—only 7 % of all possible dependency structures for sentences of length 17 ( the average sentence length in pdt ) , and only slightly more than 5 % of all possible dependency structures for sentences of length 18 ( the average sentence length in ddt ) are well-nested.5 moreover , a cursory inspection of the few problematic cases in ddt indicates that violations of the wellnestedness constraint may , at least in part , be due to properties of the annotation scheme , such as the analysis of punctuation in quotations . however , a more detailed analysis of the data from both treebanks is needed before any stronger conclusions can be drawn concerning well-nestedness . in this paper , we have reviewed a number of proposals for the characterization of mildly non-projective dependency structures , motivated by the need to find a better balance between expressivity and complexity than that offered by either strictly projective or unrestricted non-projective structures . experimental evaluation based on data from two treebanks shows , that a combination of the wellnestedness constraint and parametric constraints on discontinuity ( formalized either as gap degree or edge degree ) gives a very good fit with the empirical linguistic data . important goals for future work are to widen the empirical basis by investigating more languages , and to perform a more detailed analysis of linguistic phenomena that violate certain constraints . another important line of research is the integration of these constraints into parsing algorithms for non-projective dependency structures , potentially leading to a better trade-off between accuracy and efficiency than that obtained with existing methods . acknowledgements we thank three anonymous reviewers of this paper for their comments . the work of marco kuhlmann is funded by the collaborative research centre 378 ‘ resource-adaptive cognitive processes ’ of the deutsche forschungsgemeinschaft . the work of joakim nivre is partially supported by the swedish research council .
mildly non-projective dependency structures syntactic parsing requires a fine balance between expressivity and complexity , so that naturally occurring structures can be accurately parsed without compromising efficiency . in dependency-based parsing , several constraints have been proposed that restrict the class of permissible structures , such as projectivity , planarity , multi-planarity , well-nestedness , gap degree , and edge degree . while projectivity is generally taken to be too restrictive for natural language syntax , it is not clear which of the other proposals strikes the best balance between expressivity and complexity . in this paper , we review and compare the different constraints theoretically , and provide an experimental evaluation using data from two treebanks , investigating how large a proportion of the structures found in the treebanks are permitted under different constraints . the results indicate that a combination of the well-nestedness constraint and a parametric constraint on discontinuity gives a very good fit with the linguistic data .
using corpus statistics and wordnet relations for sense identification corpus-based approaches to word sense identification have flexibility and generality but suffer from a knowledge acquisition bottleneck . we show how knowledge-based techniques can be used to open the bottleneck by automatically locating training corpora . we describe a statistical classifier that combines topical context with local cues to identify a word sense . the classifier is used to disambiguate a noun , a verb , and an adjective . a knowledge base in the form of wordnet 's lexical relations is used to automatically locate training examples in a general text corpus . test results are compared with those from manually tagged training examples . corpus-based approaches to word sense identification have flexibility and generality but suffer from a knowledge acquisition bottleneck . we show how knowledge-based techniques can be used to open the bottleneck by automatically locating training corpora . we describe a statistical classifier that combines topical context with local cues to identify a word sense . the classifier is used to disambiguate a noun , a verb , and an adjective . a knowledge base in the form of wordnet 's lexical relations is used to automatically locate training examples in a general text corpus . test results are compared with those from manually tagged training examples . an impressive array of statistical methods have been developed for word sense identification . they range from dictionary-based approaches that rely on definitions ( vdronis and ide 1990 ; wilks et al . 1993 ) to corpus-based approaches that use only word cooccurrence frequencies extracted from large textual corpora ( schtitze 1995 ; dagan and itai 1994 ) . we have drawn on these two traditions , using corpus-based co-occurrence and the lexical knowledge base that is embodied in the wordnet lexicon . the two traditions complement each other . corpus-based approaches have the advantage of being generally applicable to new texts , domains , and corpora without needing costly and perhaps error-prone parsing or semantic analysis . they require only training corpora in which the sense distinctions have been marked , but therein lies their weakness . obtaining training materials for statistical methods is costly and timeconsuming—it is a & quot ; knowledge acquisition bottleneck & quot ; ( gale , church , and yarowsky 1992a ) . to open this bottleneck , we use wordnet 's lexical relations to locate unsupervised training examples . section 2 describes a statistical classifier , tlc ( topical/local classifier ) , that uses topical context ( the open-class words that co-occur with a particular sense ) , local context ( the open- and closed-class items that occur within a small window around a word ) , or a combination of the two . the results of combining the two types of context to disambiguate a noun ( line ) , a verb ( serve ) , and an adjective ( hard ) are presented . the following questions are discussed : when is topical context superior to local context ( and vice versa ) ? is their combination superior to either type alone ? do the answers to these questions depend on the size of the training ? do they depend on the syntactic category of the target ? manually tagged training materials were used in the development of tlc and the experiments in section 2 . the cognitive science laboratory at princeton university , with support from nsf-arpa , is producing textual corpora that can be used in developing and evaluating automatic methods for disambiguation . examples of the different meanings of one thousand common , polysemous , open-class english words are being manually tagged . the results of this effort will be a useful resource for training statistical classifiers , but what about the next thousand polysemous words , and the next ? in order to identify senses of these words , it will be necessary to learn how to harvest training examples automatically . section 3 describes wordnet 's lexical relations and the role that monosemous & quot ; relatives & quot ; of polysemous words can play in creating unsupervised training materials . tlc is trained with automatically extracted examples , its performance is compared with that obtained from manually tagged training materials . work on automatic sense identification from the 1950s onward has been well summarized by hirst ( 1987 ) and dagan and itai ( 1994 ) . the discussion below is limited to work that is closely related to our research . hearst ( 1991 ) represents local context with a shallow syntactic parse in which the context is segmented into prepositional phrases , noun phrases , and verb groups . the target noun is coded for the word it modifies , the word that modifies it , and the prepositions that precede and follow it . open-class items within ±3 phrase segments of the target are coded in terms of their relation to the target ( modifier or head ) or their role in a construct that is adjacent to the target . evidence is combined in a manner similar to that used by the local classifier component of tlc . with supervised training of up to 70 sentences per sense , performance on three homographs was quite good ( 88-100 % correct ) ; with fewer training examples and semantically related senses , performance on two additional words was less satisfactory ( 73-77 % correct ) . gale , church , and yarowsky ( 1992a ) developed a topical classifier based on bayesian decision theory . the only information the classifier uses is an unordered list of words that co-occur with the target in training examples . no other cues , such as part-of-speech tags or word order , are used . leacock , towel ! , and voorhees ( 1993 ) compared this bayesian classifier with a content vector classifier as used in information retrieval and a neural network with backpropagation . the classifiers were compared using different numbers of senses ( two , three , or six manually tagged senses of line ) and different amounts of training material ( 50 , 100 , and 200 examples ) . on the sixsense task , the classifiers averaged 74 % correct answers . leacock , towel ! , and voorhees ( 1993 ) found that the response patterns of the three classifiers converged , suggesting that each of the classifiers was extracting as much data as is available in purely topical approaches that look only at word counts from training examples . if this is the case , any technique that uses only topical information will not be significantly more accurate than the three classifiers tested . leacock , towell , and voorhees ( 1996 ) showed that performance of the content vector topical classifier could be improved with the addition of local templates— specific word patterns that were recognized as being indicative of a particular sense— in an extension of an idea initially suggested by weiss ( 1973 ) . although the templates proved to be highly reliable when they occurred , all too often , none were found . yarowsky ( 1993 ) also found that template-like structures are very powerful indicators of sense . he located collocations by looking at adjacent words or at the first word to the left or right in a given part of speech and found that , with binary ambiguity , a word has only one sense in a given collocation with a probability of 90-99 % .1 however , he had an average of only 29 % recall ( i.e. , the collocations were found in only 29 % of the cases ) . when local information occurred it was highly reliable , but all too often , it did not occur . bruce and wiebe ( 1994a , 1994b ) have developed a classifier that represents local context by morphology ( the inflection on the target word ) , the syntactic category of words within a window of ±2 words from the target , and collocation-specific items found in the sentence . the collocation-specific items are those determined to be the most informative , where an item is considered informative if the model for independence between it and a sense tag provided a poor fit to the training data . the relative probabilities of senses , available from the training corpus , are used in the decision process as prior probabilities . for each test example , the evidence in its local context is combined in a bayesian-type model of the probability of each sense , and the most probable sense is selected . performance ranges from 77-84 % correct on the test words , where a lower bound for performance based on always selecting the most frequent sense for the same words ( i.e. , the sense with the greatest prior probability ) would yield 53-80 % correct . yarowsky ( 1994 ) , building on his earlier work , designed a classifier that looks at words within ±k positions from the target ; lemma forms are obtained through morphological analysis ; and a coarse part-of-speech assignment is performed by dictionary lookup . context is represented by collocations based on words or parts of speech at specific positions within the window or , less specifically , in any position . also coded are some special classes of words , such as weekday , that might serve to distinguish among word senses . for each type of local-context evidence found in the corpus , a log-likelihood ratio is constructed , indicating the strength of the evidence for one form of the homograph versus the other . these ratios are then arranged in a sorted decision list with the largest values ( strongest evidence ) first . a decision is made for a test sentence by scanning down the decision list until a match is found . thus , only the single best piece of evidence is used . the classifier was tested on disambiguating the homographs that result from accent removal in spanish and french ( e.g. , seria , seria ) . in tests with the number of training examples ranging from a few hundred to several thousand , overall accuracy was high , above 90 % . clearly , sense identification is an active area of research , and considerable ingenuity is apparent . but despite the promising results reported in this literature , the reality is that there still are no large-scale , operational systems for tagging the senses of words in text . the statistical classifier , tlc , uses topical context , local context , or a combination of the two , for word sense identification . tlc 's flexibility in using both forms is an important asset for our investigations . a noun , a verb , and an adjective were tested in this study . table 1 provides a synonym or brief gloss for each of the senses used . training corpora and testing corpora were collected as follows : wall street journal corpus and from the american printing house for the blind corpus . ' examples for hard were taken from the ldc san jose mercury news ( sjm ) corpus . each consisted of the sentence containing the target and one sentence preceding it . the resulting strings had an average length of 49 items . 2 . examples where the target was the head of an unambiguous collocation were removed from the files . being unambiguous , they do not need to be disambiguated . these collocations , for example , product line and hard candy were found using wordnet . in section 3 , we consider how they can be used for unsupervised training . examples where the target was part of a proper noun were also removed ; for example , japan air lines was not taken as an example of line . first 25 , 50 , 100 , and 200 examples of the least frequent sense , and examples from the other senses in numbers that reflected their relative frequencies in the corpus . as an illustration , in the smallest training set for hard , there were 25 examples of the least frequent sense , 37 examples of the second most frequent sense , and 256 examples of the most frequent sense . the test sets were of fixed size : each contained 150 of the least frequent sense and examples of the other senses in numbers that reflected their relative frequencies . the operation of tlc consists of preprocessing , training , and testing . during preprocessing , examples are tagged with a part-of-speech tagger ( brill 1994 ) ; special tags are inserted at sentence breaks ; and each open-class word found in wordnet is replaced with its base form . this step normalizes across morphological variants without resorting to the more drastic measure of stemming . morphological information is not lost , since the part-of-speech tag remains unchanged . training consists of counting the frequencies of various contextual cues for each sense . testing consists of taking a new example of the polysemous word and computing the most probable sense , based on the cues present in the context of the new item . a comparison is made to the sense assigned by a human judge , and the classifier 's decision is scored as correct or incorrect . tlc uses a bayesian approach to find the sense s , that is the most probable given the cues ci contained in a context window of ±k positions around the polysemous target word . for each s „ the probability is computed with bayes ' rule : as golding ( 1995 ) points out , the term p ( c_k , . • • s , ) is difficult to estimate because of the sparse data problem , but if we assume , as is often done , that the occurrence of each cue is independent of the others , then this term can be replaced with : in tlc , we have made this assumption and have estimated p ( ci i si ) from the training . of course , the sparse data problem affects these probabilities too , and so tlc uses the good-turing formula ( good 1953 ; chiang , lin , and su 1995 ) , to smooth the values of p ( cj s , ) , including providing probabilities for cues that did not occur in the training . tlc actually uses the mean of the good-turing value and the training-derived value for p ( cj s , ) . when cues do not appear in training , it uses the mean of the goodturing value and the global probability of the cue p ( ci ) , obtained from a large text corpus . this approach to smoothing has yielded consistently better performance than relying on the good-turing values alone . tuation . for this cue type , p ( cj i s , ) is the probability that item cl appears precisely at location j for sense si . positions j = —2 , —1 , 1,2 are used . the global probabilities , for example p ( the_i ) , are based on counts of closed-class items found at these positions relative to the nouns in a large text corpus . the local window width of ±2 was selected after pilot testing on the semantically tagged brown corpus . as in ( 2 ) above , the local window does not extend beyond a sentence boundary . 4 . part-of-speech tags in the positions j = —2 , —1,0 , 1,2 are also used as cues . the probabilities for these tags are computed for specific positions ( e.g. , p ( dt_i i p ( dta ) in a manner similar to that described in ( 3 ) above . when tlc is configured to use only topical information , cue type ( 1 ) is employed . when it is configured for local information , cue types ( 2 ) , ( 3 ) , and ( 4 ) are used . finally , in combined mode , the set of cues contains all four types . 2.3 results figures 1 to 3 show the accuracy of the classifier as a function of the size of the training set when using local context , topical context , and a combination of the two , averaged across three runs for each training set . to the extent that the words used are representative , some clear differences appear as a function of syntactic category . with the verb serve , local context was more reliable than topical context at all levels of training ( 78 % versus 68 % with 200 training examples for the least frequent sense ) . the combination of local and topical context showed improvement ( 83 % ) over either form alone ( see figure 1 ) . with the adjective hard , local context was much more reliable as an indicator of sense than topical context for all training sizes ( 83 % versus 60 % with 200 training examples ) and the combined classifier 's performance ( at 83 % ) was the same as for local ( see figure 2 ) . in the case of the noun line , topical was slightly better than local at all set sizes , but with 200 training examples , their combination yielded 84 % accuracy , greater than either topical ( 78 % ) or local ( 67 % ) alone ( see figure 3 ) . to summarize , local context was more reliable than topical context as an indicator of sense for this verb and this adjective , but slightly less reliable for this noun . the combination of local and topical context showed improved or equal performance for all three words . performance for all of the classifiers improved with increased training size . all classifiers performed best with at least 200 training examples per sense , but the learning curve tended to level off beyond a minimum 100 training examples . these results are consistent with those of yarowsky ( 1993 ) , based on his experiments with pseudowords , homophones , and homonyms ( discussed below ) . he observed that performance for verbs and adjectives dropped sharply as the window increased , while distant context remained useful for nouns . thus one is tempted to conclude that nouns depend more on topic than do verbs and adjectives . but such a conclusion is probably an overgeneralization , inasmuch as some noun senses are clearly nontopical . thus , leacock , towell , and voorhees ( 1993 ) found that some senses of the noun line are not susceptible to disambiguation with topical context . for example , the 'textual ' sense of line can appear with any topic , whereas the 'product ' sense of line can not . when it happens that a nontopical sense accounts for a large proportion of occurrences ( in our study , all senses of hard are nontopical ) , then adding topical context to local will have little benefit and may even reduce accuracy . one should not conclude from these results that the topical classifiers and tlc are inferior to the classifiers reviewed in section 2 . in our experiments , monosemous collocations in wordnet that contain the target word were systematically removed from the training and testing materials . this was done on the assumption that these words are not ambiguous . removing them undoubtedly made the task more difficult than it would normally be . how much more difficult ? an estimate is possible . we classifier performance on four senses of the verb serve . percentage accounted for by most frequent sense = 41 % . searched through 7,000 sentences containing line and found 1,470 sentences contained line as the head of a monosemous collocation in wordnet , i.e. , line could be correctly disambiguated in some 21 % of those 7,000 sentences simply on the basis of the wordnet entries in which it occurred . in other words , if these sentences had been included in the experiment—and had been identified by automatic lookup—overall accuracy would have increased from 83 % to 87 % . using topical context alone , tlc performs no worse than other topical classifiers . leacock , towell , and voorhees ( 1993 ) report that the three topical classifiers tested averaged 74 % accuracy on six senses of the noun line . with these same training and testing data , tlc performed at 73 % accuracy . similarly , when the content vector and neural network classifiers were run on manually tagged training and testing examples of the verb serve , they averaged 74 % accuracy—as did tlc using only topical context . when local context is combined with topical , tlc is superior to the topical classifiers compared in the leacock , towel ! , and voorhees ( 1993 ) study . just how useful is a sense classifier whose accuracy is 85 % or less ? probably not very useful if it is part of a fully automated nlp application , but its performance might be adequate in an interactive application ( e.g. , machine-assisted translation , on-line thesaurus functions in word processing , interactive information retrieval ) . in fact , when recall does not have to be 100 % ( as when a human is in the loop ) the precision of the classifier can be improved considerably . the classifier described above always selects the sense that has the highest probability . we have observed that when classifier performance on three senses of the adjective hard . percentage accounted for by most frequent sense = 80 % . the difference between the probability of this sense and that of the second highest is relatively small , the classifier 's choice is often incorrect . one way to improve the precision of the classifier , though at the price of reduced recall , is to identify these situations and allow it to respond do not know rather than forcing a decision . what is needed is a measure of the difference in the probabilities of the two senses . following the approach of dagan and itai ( 1994 ) , we use the log of the ratio of the probabilities in ( pi/p2 ) for this purpose . based on this value , a threshold e can be set to control when the classifier selects the most probable sense . for example , if e = 2 , then ln ( pi/p2 ) must be 2 or greater for a decision to be made . dagan and itai ( 1994 ) also describe a way to make the threshold dynamic so that it adjusts for the amount of evidence used to estimate pi and p2 . the basic idea is to create a one-tailed confidence interval so that we can state with probability 1 — a that the true value of the difference measure is greater than o . when the amount of evidence is small , the value of the measure must be larger in order to insure that e is indeed exceeded . table 2 shows precision and recall values for serve , hard , and line at eight different settings of 0 using a 60 % confidence interval . tlc was first trained on 100 examples of each sense , and it was then tested on separate 100-example sets . in all cases , precision was positively correlated with the square root of 0 ( all r values > .97 ) , and recall was negatively correlated with the square root of 0 ( r values < —.96 ) . as cross-validation , the equations of the lines that fit the precision and recall results on the test sample were used to predict the precision and recall at the various values of 0 on a second test sample . they provided a good fit to the new data , accounting for an average of 93 % of the variance . the standard errors of estimate for hard , serve , and line were .028 , .030 , and .029 for precision , and .053 , .068 , and .041 for recall . this demonstrates that it is possible to produce accurate predictions of precision and recall as a function of for new test sets . when the threshold is set to a large value , precision approaches 100 % . the criterion thus provides a way to locate those cases that can be identified automatically with very high accuracy . when tlc uses a high criterion for assigning senses , it can be used to augment the training examples by automatically collecting new examples from the test corpus . in summary , the results obtained with tlc support the following preliminary conclusions : ( a ) improvement with training levels off after about 100 training examples for the least frequent sense ; ( b ) the high predictive power of local context for the verb and adjective indicate that the local parameters effectively capture syntactically mediated relations , e.g. , the subject and object or complement of verbs , or the noun that an adjective modifies ; ( c ) nouns may be more & quot ; topical & quot ; than verbs and adjectives , and therefore benefit more from the combination of topical and local context ; ( d ) the precision of tlc can be considerably improved at the price of recall , a trade-off that may be desirable in some interactive nlp applications . a final observation we can make is that when topical and local information is combined , what we have called & quot ; nontopical senses & quot ; can reduce overall accuracy . for example , the 'textual ' sense of line is relatively topic-independent . the results of the line experiment were not affected too adversely because the nontopical sense of line accounted for only 10 % of the training examples . the effects of nontopical senses will be more serious when most senses are nontopical , as in the case of many adjectives and verbs . the generality of these conclusions must , of course , be tested with additional words , which brings us to the problem of obtaining training and testing corpora . on one hand , it is surprising that a purely statistical classifier can & quot ; learn & quot ; how to identify a sense of a polysemous word with as few as 100 example contexts . on the other hand , anyone who has manually built such sets knows that even collecting 100 examples of each sense is a long and tedious process . the next section presents one way in which the lexical knowledge in wordnet can be used to extract training examples automatically . corpus-based word sense identifiers are data hungry—it takes them mere seconds to digest all of the information contained in training materials that take months to prepare manually . so , although statistical classifiers are undeniably effective , they are not feasible until we can obtain reliable unsupervised training data . in the gale , church , and yarowsky ( 1992a ) study , training and testing materials were automatically acquired using an aligned french-english bilingual corpus by searching for english words that have two different french translations . for example , english tokens of sentence were translated as either peine or phrase . they collected contexts of sentence translated as peine to build a corpus for the judicial sense , and collected contexts of sentence translated as phrase to build a corpus for the grammatical sense . one problem with relying on bilingual corpora for data collection is that bilingual corpora are rare , and aligned bilingual corpora are even rarer . another is that since french and english are so closely related , different senses of polysemous english words often translate to the same french word . for example , line is equally polysemous in french and english—and most senses of line translate into french as ligne . several artificial techniques have been used so that classifiers can be developed and tested without having to invest in manually tagging the data : yarowsky ( 1993 ) and schtitze ( 1995 ) have acquired training and testing materials by creating pseudowords from existing nonhomographic forms . for example , a pseudoword was created by combining abused/escorted . examples containing the string escorted were collected to train on one sense of the pseudoword and examples containing the string abused were collected to train on the other sense . in addition , yarowsky ( 1993 ) used homophones ( e.g. , cellar/seller ) and yarowsky ( 1994 ) created homographs by stripping accents from french and spanish words . although these latter techniques are useful in their own right ( e.g. , spoken language systems or corrupted transmissions ) , the resulting materials do not generalize to the acquisition of tagged training for real polysemous or even homographic words . the results of disambiguation strategies reported for pseudowords and the like are consistently above 95 % overall accuracy , far higher than those reported for disambiguating three or more senses of polysemous words ( wilks et al . 1993 ; leacock , towel ! , and voorhees 1993 ) . yarowsky ( 1992 ) used a thesaurus to collect training materials . he tested the unsupervised training materials on 12 nouns with almost perfect results on homonyms ( 95-99 % ) , 72 % accuracy for four senses of interest , and 77 % on three senses of cone . the training was collected in the following manner . take a roget 's category—his examples were tool and animal—and collect sentences from a corpus ( in this case , grolier 's encyclopedia ) using the words in each category . consider the noun crane , which appears in both the roget 's categories tool and animal . to represent the tool category , yarowsky extracted contexts from gro/ier 's encyclopedia . for example , contexts with the words adz , shovel , crane , sickle , and so on . similarly he collected sentences with names of animals from the animal category . in these samples , crane and drill appeared under both categories . yarowsky points out that the resulting noise will be a problem only when one of the spurious senses is salient , dominating the training set , and he uses frequency-based weights to minimize these effects . we propose to minimize spurious training by using monosemous words and collocations—on the assumption that , if a word has only one sense in wordnet , it is monosemous . schtitze ( 1995 ) developed a statistical topical approach to word sense identification that provides its own automatically extracted training examples . for each occurrence t of a polysemous word in a corpus , a context vector is constructed by summing all the vectors that represent the co-occurrence patterns of the open-class words in t 's context ( i.e. , topical information is expressed as a kind of second-order co-occurrence ) . these context vectors are clustered , and the centroid of each cluster is used to represent a & quot ; sense. & quot ; when given a new occurrence of the word , a vector of the words in its context is constructed , and this vector is compared to the sense representations to find the closest match . schulze has used the method to disambiguate pseudowords , homographs , and polysemous words . performance varies depending , in part , on the number of clusters that are created to represent senses , and on the degree to which the distinctions correspond to different topics . this approach performs very well , especially with pseudowords and homographs . however , there is no automatic means to map the sense representations derived from the system onto the more conventional word senses found in dictionaries . consequently , it does not provide disambiguated examples that can be used by other systems . yarowsky ( 1995 ) has proposed automatically augmenting a small set of experimenter-supplied seed collocations ( e.g. , manufacturing plant and plant life for two different senses of the noun plant ) into a much larger set of training materials . he resolved the problem of the sparseness of his collocations by iteratively bootstrapping acquisition of training materials from a few seed collocations for each sense of a homograph . he locates examples containing the seeds in the corpus and analyzes these to find new predictive patterns in these sentences and retrieves examples containing these patterns . he repeats this step iteratively . results for the 12 pairs of homographs reported are almost perfect . in his paper , yarowsky suggests wordnet as a source for the seed collocations—a suggestion that we pursue in the next section . wordnet is particularly well suited to the task of locating sense-relevant context because each word sense is represented as a node in a rich semantic lexical network with synonymy , hyponymy , and meronymy links to other words , some of them polysemous and others monosemous . these lexical & quot ; relatives & quot ; provide a key to finding relevant training sentences in a corpus . for example , the noun suit is polysemous , but one sense of it has business suit as a monosemous daughter and another has legal proceeding as a hypernym . by collecting sentences containing the unambiguous nouns business suit and legal proceeding we can build two corpora of contexts for the respective senses of the polysemous word . all the systems described in section 2.1 could benefit from the additional training materials that monosemous relatives can provide . the wordnet on-line lexical database ( miller 1990 , 1995 ) has been developed at princeton university over the past 10 years . ' like a standard dictionary , wordnet contains the definitions of words . it differs from a standard dictionary in that , instead of being organized alphabetically , wordnet is organized conceptually . the basic unit in wordnet is a synonym set , or synset , which represents a lexicalized concept . for example , wordnet version 1.5 distinguishes between two senses of the noun shot with the synsets { shot , snapshot } and { shot , injection } . in the context , & quot ; the photographer took a shot of mary , & quot ; the word snapshot can be substituted for one sense of shot . in the context , & quot ; the nurse gave mary a flu shot , & quot ; the word injection can be substituted for another sense of shot . nouns , verbs , adjectives , and adverbs are each organized differently in wordnet . all are organized in synsets , but the semantic relations among the synsets differ depending on the grammatical category , as can be seen in table 3 . nouns are organized in a hierarchical tree structure based on hypernymy/hyponymy . the hyponym of a noun is its subordinate , and the relation between a hyponym and its hypernym is an is a kind of relation . for example , maple is a hyponym of tree , which is to say that a maple is a kind of tree . hypernymy ( supername ) and its inverse , hyponymy ( subname ) , are transitive semantic relations between synsets . meronymy ( part-name ) , and its inverse holortymy ( whole-name ) , are complex semantic relations that distinguish component parts , substantive parts , and member parts . the verbal hierarchy is based on troponymy , the is a manner of relation . for example , stroll is a troponym of walk , which is to say that strolling is a manner of walking . entailment relations between verbs are also coded in wordnet . the organization of attributive adjectives is based on the antonymy relation . where direct antonyms exist , adjective synsets point to antonym synsets . a head adjective is one that has a direct antonym ( e.g. , hot versus cold or long versus short ) . many adjectives , like sultry , have no direct antonyms . when an adjective has no direct antonym , its synset points to a head that is semantically similar to it . thus sultry and torrid are similar in meaning to hot , which has the direct antonym of cold . so , although sultry has no direct antonym , it has cold as its indirect antonym . relational adjectives do not have antonyms ; instead they point to nouns . consider the difference between a nervous disorder and a nervous student . in the former , nervous pertains to a noun , as in nervous system , whereas the latter is defined by its relation to other adjectives—its synonyms ( e.g. , edgy ) and antonyms ( e.g. , relaxed ) . adverbs have synonymy and antonymy relations . when the adverb is morphologically related to an adjective ( when an -ly suffix is added to an adjective ) and semantically related to the adjective as well , the adverb points to the adjective . we have had some success in exploiting wordnet 's semantic relations for word sense identification . since the main problem with classifiers that use local context is the sparseness of the training data , leacock and chodorow ( 1998 ) used a proximity measure on the hypernym relation to replace the subject and complement of the verb serve in the testing examples with the subject and complement from training examples that were & quot ; closest & quot ; to them in the noun hierarchy . for example , one of the test sentences was & quot ; sauerbraten is usually served with dumplings , & quot ; where neither sauerbraten nor dumpling appeared in any training sentence . the similarity measures on wordnet found that sauerbraten was most similar to dinner in the training , and dumpling to bacon . these nouns were substituted for the novel ones in the test sets . thus the sentence & quot ; dinner is usually served with bacon & quot ; was substituted for the original sentence . augmentation of the local .context classifier with wordnet similarity measures showed a small but consistent improvement in the classifier 's performance . the improvement was greater with the smaller training sets . resnik ( 1992 ) uses an information-based measure , the most informative class , on the wordnet taxonomy . a class consists of the synonyms found at a node and the synonyms at all the nodes that it dominates ( all of its hyponyms ) . based on verb/object pairs collected from a corpus , resnik found , for example , that the objects for the verb open fall into two classes : receptacle and oral communication . conversely , the class of a verb 's object could be used to determine the appropriate sense of that verb . the experiments in the next section depend on a subset of the wordnet lexical relations , those involving monosemous relatives , so we were interested in determining just what proportion of word senses have such relatives . we examined 8,500 polysemous nouns that appeared in a moderate-size , 25-million-word corpus . in all , these 8,500 nouns have more than 24,000 wordnet senses . restricting the relations to synonyms , immediate hyponyms ( i.e. , daughters ) , and immediate hypernyms ( parents ) , we found that about 64 % ( 15,400 ) have monosemous relatives attested in the corpus . with larger corpora ( e.g. , with text obtained by web crawling ) and more lexical relations ( e.g. , meronymy ) , this percentage can be expected to increase . the approach we have used is related to that of yarowsky ( 1992 ) in that training materials are collected using a knowledge base , but it differs in other respects , notably in the selection of training and testing materials , the choice of a knowledge base , and use of both topical and local classifiers . yarowsky collects his training and testing materials from a specialized corpus , grolier 's encyclopedia . it remains to be seen whether a statistical classifier trained on a topically organized corpus such as an encyclopedia will perform in the same way when tested on general unrestricted text , such as newspapers , periodicals , and books . one of our goals is to determine whether automatic extraction of training examples is feasible using general corpora . in his experiment , yarowsky uses an updated on-line version of roget 's thesaurus that is not generally available to the research community . the only generally available version of roget 's is the 1912 edition , which contains many lexical gaps . we are using wordnet , which can be obtained via anonymous ftp . yarowsky 's classifier is purely topical , but we also examine local context . finally , we hope to avoid inclusion of spurious senses by using monosemous relatives . in this experiment we collected monosemous relatives of senses of 14 nouns . training sets are created in the following manner . a program called autotrain retrieves from wordnet all of the monosemous relatives of a polysemous word sense , samples and retrieves example sentences containing these monosemous relatives from a 30-million-word corpus of the san jose mercury news , and formats them for tlc . the sampling process retrieves the & quot ; closest & quot ; relatives first . for example , suppose that the system is asked to retrieve 100 examples for each sense of the noun court . the system first looks for the strongest or top-level relatives : for monosemous synonyms of the sense ( e.g. , tribunal ) and for daughter collocations that contain the target word as the head ( e.g. , superior court ) and tallies the number of examples in the corpus for each . if the corpus has 100 or more examples for these top-level relatives , it retrieves a sampling of them and formats them for tlc . if there are not enough top-level examples , the remainder of the target 's monosemous relatives are inspected in the order : all other daughters ; hyponym collocations that contain the target ; all other hyponyms ; hypernyms ; and , finally , sisters . autotrain takes as broad a sampling as possible across the corpus and never takes more than one example from an article . the number of examples for each relative is based on the relative proportion of its occurrences in the corpus . table 4 shows the monosemous relatives that were used to train five senses of the noun line—the monosemous relatives of the sixth sense in the original study , line as an abstract division , are not attested in the sim corpus . the purpose of the experiment was to see how well tlc performed using unsupervised training and , when possible , to compare this with its performance when training on the manually tagged materials being produced at princeton 's cognitive science laboratory . ' when a sufficient number of examples for two or more senses were available , 100 examples of each sense were set aside to use in training . the remainder were used for testing . only the topical and local open-class cues were used , since preliminary tests showed that performance declined when using local closed-class and part-of-speech cues obtained from the monosemous relatives . this is not surprising , as many of the relatives are collocations whose local syntax is quite different from that of the polysemous word in its typical usage . for example , the 'formation ' sense of line is often followed by an of-phrase as in a line of children , but its relative , picket line , is not . prior probabilities for the sense were taken from the manually tagged materials . table 5 shows the results when tlc was trained on monosemous relatives and on manually tagged training materials . baseline performance is when the classifier always chooses the most frequent sense . eight additional words had a sufficient number of manually tagged examples for testing but not for training tlc . these are shown in table 6 . for four of the examples in table 5 , training with relatives produced results within 1 % or 2 % of manually tagged training . line and work , however , showed a substantial decrease in performance . in the case of line , this might be due to overly specific training contexts . almost half of the training examples for the 'formation ' sense of line come from one relative , picket line . in fact , all of the monosemous relatives , except for rivet line and trap line , are human formations . this may have skewed training so that the classifier performs poorly on other uses of line as formation . in order to compare our results with those reported in yarowsky ( 1992 ) , we trained and tested on the same two senses of the noun duty that yarowsky had tested ( 'obligation ' and 'tax ' ) . he reported that his thesaurus-based approach yielded 96 % precision with 100 % recall . tlc used training examples based on monosemous wordnet relatives and correctly identified the senses with 93.5 % precision at 100 % recall . table 6 shows tlc 's performance on the other eight words after training with monosemous relatives and testing on manually tagged examples . performance is about the same as , or only slightly better than , the highest prior probability . in part , this is due to the rather high probability of the most frequent sense for this set . the values in the table are based on decisions made on all test examples . if a threshold is set for tlc ( see section 2.4 ) , precision of the classifier can be increased substantially , at the expense of recall . table 7 shows recall levels when tlc is trained on monosemous relatives and the value of e is set for 95 % precision . operating in this mode , the classifier can gather new training materials , automatically , and with high precision . this is a particularly good way to find clear cases of the most frequent sense . the results also show that not all words are well suited to this kind of operation . little can be gained for a word like work , where the two senses , 'activity ' and 'product , ' are closely related and therefore difficult for the classifier to distinguish , due to a high degree of overlap in the training contexts . problems of this sort can be detected even before testing , by computing correlations between the vectors of open-class words for the different senses . the cosine correlation between the 'activity ' and 'product ' senses of work is r = .49 , indicating a high degree of overlap . the mean correlation between pairs of senses for the other words in table 7 is r = .31 . our evidence indicates that local context is superior to topical context as an indicator of word sense when using a statistical classifier . the benefits of adding topical to local context alone depend on syntactic category as well as on the characteristics of the individual word . the three words studied yielded three different patterns ; a substantial benefit for the noun line , slightly less for the verb serve , and none for the adjective hard . some word senses are simply not limited to specific topics , and appear freely in many different domains of discourse . the existence of nontopical senses also limits the applicability of the & quot ; one sense per discourse & quot ; generalization of gale , church , and yarowsky ( 1992b ) , who observed that , within a document , a repeated word is almost always used in the same sense . future work should be directed toward developing methods for determining when a word has a nontopical sense . one approach to this problem is to look for a word that appears in many more topical domains than its total number of senses . because the supply of manually tagged training data will always be limited , we propose a method to obtain training data automatically using commonly available materials : exploiting wordnet 's lexical relations to harvest training examples from ldc corpora or even the world wide web . we found this method to be effective , although not as effective as using manually tagged training . we have presented the components of a system for acquiring unsupervised training materials that can be used with any statistical classifier . the components can be fit together in the following manner . for a polysemous word , locate the monosemous relatives for each of its senses in wordnet and extract examples containing these relatives from a large corpus . senses whose contexts greatly overlap can be identified with a simple cosine correlation . often , correlations are high between senses of a word that are systematically related , as we saw for the 'activity ' and 'product ' senses of work . in some cases , the contexts for the two closely related senses may be combined . since the frequencies of the monosemous relatives do not correlate with the frequencies of the senses , prior probabilities must be estimated for classifiers that use them . in the experiments of section 3.2 , these were estimated from the testing materials . they can also be estimated from a small manually tagged sample , such as the parts of the brown corpus that have been tagged with senses in wordnet . when the threshold is set to maximize precision , the results are highly reliable and can be used to support an interactive application , such as machine-assisted translation , with the goal of reducing the amount of interaction . although we have looked at only a few examples , it is clear that , given wordnet and a large enough corpus , the methods outlined for training on monosemous relatives can be generalized to build training materials for thousands of polysemous words . we are indebted to the other members of the wordnet group who have provided advice and technical support : christiane fellbaum , shari landes , and randee tengi . we are also grateful to paul bagyenda , ben johnson-laird and joshua schecter . we thank scott wayland , tim allison and jill hollifield for tagging the serve and hard corpora . finally we are grateful to the three anonymous cl reviewers for their comments and advice . this material is based upon work supported in part by the national science foundation under nsf award no . 1r19528983 and by the defense advanced research projects agency , grant no . n00014-91-1634 .
using corpus statistics and wordnet relations for sense identification corpus-based approaches to word sense identification have flexibility and generality but suffer from a knowledge acquisition bottleneck . we show how knowledge-based techniques can be used to open the bottleneck by automatically locating training corpora . we describe a statistical classifier that combines topical context with local cues to identity a word sense . the classifier is used to disambiguate a noun , a verb , and an adjective . a knowledge base in the form of wordnet 's lexical relations is used to automatically locate training examples in a general text corpus . test results are compared with those from manually tagged training examples . we present a method to obtain sense-tagged examples using monosemous relatives .
"automatic labeling of semantic roles present a system for identifying the semantic relationships , (...TRUNCATED)
"automatic labeling of semantic roles we present a system for identifying the semantic relationships(...TRUNCATED)
"generative models for statistical parsing with combinatory categorial grammar this paper compares a(...TRUNCATED)
"generative models for statistical parsing with combinatory categorial grammar this paper compares a(...TRUNCATED)
"corpus statistics meet the noun compound : some empirical results tagged dependency -.— tagged ad(...TRUNCATED)
"corpus statistics meet the noun compound : some empirical results a variety of statistical methods (...TRUNCATED)
"building a large annotated corpus of english : the penn treebank mitchell p. marcus * university of(...TRUNCATED)
building a large annotated corpus of english : the penn treebank
"the automated acquisit ion of topic signatures for text summarizat ion chin -yew l in and eduard ho(...TRUNCATED)
"the automated acquisition of topic signatures for text summarization in order to produce a good sum(...TRUNCATED)
"improving machine translation performance by exploiting non-parallel corpora we present a novel met(...TRUNCATED)
"improving machine translation performance by exploiting non-parallel corpora we present a novel met(...TRUNCATED)
"introduction to the special issue on word sense disambiguation : the state of the art the automatic(...TRUNCATED)
"introduction to the special issue on word sense disambiguation : the state of the art we present a (...TRUNCATED)
End of preview. Expand in Data Studio

No dataset card yet

Downloads last month
8