{ "paper_id": "W01-0512", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T06:00:56.037147Z" }, "title": "The Unknown Word Problem: a Morphological Analysis of Japanese Using Maximum Entropy Aided by a Dictionary", "authors": [ { "first": "Kiyotaka", "middle": [], "last": "Uchimotoy", "suffix": "", "affiliation": {}, "email": "uchimoto@crl.go.jp" }, { "first": "Satoshi", "middle": [], "last": "Sekinez", "suffix": "", "affiliation": {}, "email": "sekine@cs.nyu.edu" }, { "first": "Hitoshi", "middle": [], "last": "Isaharay", "suffix": "", "affiliation": {}, "email": "isahara]@crl.go.jp" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "In this paper we describe a morphological analysis method based on a maximum entropy m o d e l. This method uses a model that can not only consult a dictionary with a large amount of lexical information but can also identify unknown words by learning certain characteristics. The model has the potential to overcome the unknown word problem.", "pdf_parse": { "paper_id": "W01-0512", "_pdf_hash": "", "abstract": [ { "text": "In this paper we describe a morphological analysis method based on a maximum entropy m o d e l. This method uses a model that can not only consult a dictionary with a large amount of lexical information but can also identify unknown words by learning certain characteristics. The model has the potential to overcome the unknown word problem.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Morphological analysis is one of the basic techniques used in Japanese sentence analysis. A morpheme is a minimal grammatical unit, such as a word or a sux, and morphological analysis is the process segmenting a given sentence into a r o w of morphemes and assigning to each morpheme grammatical attributes such as a partof-speech (POS) and an inection type. One of the most important problems in morphological analysis is that posed by unknown words, which are words found in neither a dictionary nor a training corpus, and there have been two statistical approaches to this problem. One is to acquire unknown words from corpora and put them into a dictionary (e.g., (Mori and Nagao, 1996) ), and the other is to estimate a model that can identify unknown words correctly (e.g., (Kashioka et al., 1997; Nagata, 1999) ). We would like t o b e a b l e t o m a k e good use of both approaches. If words acquired by the former method could be added to a dictionary and a model developed by the latter method could consult the amended dictionary, then the model could be the best statistical model which has the potential to overcome the unknown word problem. Mori and Nagao proposed a statistical model that can consult a dictionary (Mori and Nagao, 1998) . In their model the probability that a string of letters or characters is a morpheme is augmented when the string is found in a dictionary. The improvement o f the accuracy was slight, however, so we think that it is dicult to eciently integrate the mechanism for consulting a dictionary into an n-gram model. In this paper we therefore describe a morphological analysis method based on a maximum entropy (M.E.) model. This method uses a model that can not only consult a dictionary but can also identify unknown words by learning certain characteristics. To learn these characteristics, we focused on such information as whether or not a string is found in a dictionary and what types of characters are used in a string. The model estimates how l i k ely a string is to be a morpheme according to the information on hand. When our method was used to identify morpheme segments in sentences in the Kyoto University corpus and to identify the major parts-of-speech of these morphemes, the recall and precision were respectively 95.80% and 95.09%.", "cite_spans": [ { "start": 668, "end": 690, "text": "(Mori and Nagao, 1996)", "ref_id": "BIBREF11" }, { "start": 780, "end": 803, "text": "(Kashioka et al., 1997;", "ref_id": "BIBREF5" }, { "start": 804, "end": 817, "text": "Nagata, 1999)", "ref_id": "BIBREF14" }, { "start": 1230, "end": 1252, "text": "(Mori and Nagao, 1998)", "ref_id": "BIBREF12" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "This section describes a model which estimates how likely a string is to be a morpheme. We implemented this model within an M.E. framework. Given a tokenized test corpus, the problem of Japanese morphological analysis can be reduced to the problem of assigning one of two tags to each string in a sentence. A string is tagged with a 1 or a 0 to indicate whether or not it is a morpheme. When a string is a m o rpheme, a grammatical attribute is assigned to it. The 1 tag is thus divided into the number, n, of grammatical attributes assigned to morphemes, and the problem is to assign an attribute (from 0 to n) t o e v ery string in a given sentence. The (n+1) tags form the space of \\fu-tures\" in the M.E. formulation of our problem of morphological analysis. The M.E. model, as well as other similar models, enables the computation of P (fjh) for any future f from the space of possible futures, F , and for every history, h, from the space of possible histories, H.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "A \\history\" in M.E. is all of the conditioning data that enable us to make a decision in the space of futures. In the problem of morphological analysis, we can reformulate this in terms of nding the probability o f f associated with the relationship at index t in the test corpus: P (fjh t ) = P (fjInformation derivable from the test corpus related to relationship t)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "The computation of P (fjh) i n a n y M.E. models is dependent on a set of \\features\" which w ould be helpful in making a prediction about the future. Like most current M.E. models in computational linguistics, our model is restricted to those features which are binary functions of the history and future. For instance, one of our features is g(h; f ) = 8 > < > :", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "1 : if has(h; x) = true;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "x = \\POS(01)(Major) : verb; 00 & f = 1 0 : otherwise:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "(1)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "Here \\has(h,x)\" is a binary function that returns true if the history h has feature x. In our experiments, we focused on such information as whether or not a string is found in a dictionary, the length of the string, what types of characters are used in the string, and the part-of-speech o f the adjacent morpheme.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "Given a set of features and some training data, the M.E. estimation process produces a model in which every feature g i has an associated parameter i . This enables us to compute the conditional probability a s f o l l o ws (Berger et al., 1996) :", "cite_spans": [ { "start": 224, "end": 245, "text": "(Berger et al., 1996)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "P (fjh) = Q i g i (h;f) i Z (h) (2) Z (h) = X f Y i g i (h;f) i : (3)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "The M.E. estimation process guarantees that for every feature g i , the expected value of g i according to the M.E. model will equal the empirical expectation of g i in the training corpus. In other", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "words, X h;fP (h; f) 1 g i (h; f) = X hP (h) 1 X f P M: E: (fjh) 1 g i (h; f): (4)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "HereP is an empirical probability a n d P M: E: is the probability assigned by the model. We dene part-of-speech and bunsetsu boundaries as grammatical attributes. Here a bunsetsu is a phrasal unit consisting of one or more morphemes. When there are m types of parts-of-speech, and the left-hand side of each morpheme may o r m a y not be a bunsetsu boundary, the number, n, of grammatical attributes assigned to morphemes is 2 2 m. 1 We propose a model which estimates the likelihood that a given string is a morpheme and has the grammatical attribute i(1 i n). We call it a morpheme model. This model is represented by Eq. 2, in which f can be one of (n + 1 ) tags from 0 to n.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "A given sentence is divided into morphemes, and a grammatical attribute is assigned to each morpheme so as to maximize the sentence probability estimated by our morpheme model. Sentence probability is dened as the product of the probabilities estimated for a particular division of morphemes in a sentence. We use the Viterbi algorithm to nd the optimal set of morphemes in a s e n tence and we use the method proposed by Nagata (Nagata, 1994) to search for the Nbest sets.", "cite_spans": [ { "start": 429, "end": 443, "text": "(Nagata, 1994)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "A Morpheme Model", "sec_num": "2" }, { "text": "Experiments and Discussion", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "3", "sec_num": null }, { "text": "The part-of-speech categories that we used follow those of JUMAN (Kurohashi and Nagao, 1999) . There are 53 categories covering all possible combinations of major and minor categories as dened in JUMAN. The number of grammatical attributes is 106 if we include the detection of whether or not the left side of a morpheme is a bunsetsu boundary. We do not identify inection types probabilistically since they can be almost perfectly identied by c hecking the spelling of the current morpheme after a part-of-speech has been assigned to it. Therefore, f in Eq. (2) can be one of 107 tags from 0 to 106. We used the Kyoto University text corpus (Version 2) (Kurohashi and Nagao, 1997) , a tagged corpus of the Mainichi newspaper. For training, we used 7,958 sentences from newspaper articles appearing from January 1 to January 8, 1995, and for testing, we used 1,246 sentences from articles appearing on January 9, 1995.", "cite_spans": [ { "start": 65, "end": 92, "text": "(Kurohashi and Nagao, 1999)", "ref_id": "BIBREF8" }, { "start": 654, "end": 681, "text": "(Kurohashi and Nagao, 1997)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "Given a sentence, for every string consisting of ve or less characters and every string appearing in the JUMAN dictionary (Kurohashi and Nagao, 1999) , whether or not the string is a morpheme was determined and then the grammatical attribute of each string determined to be a morpheme was identied and assigned to that string. The maximum length was set at ve because morphemes consisting of six or more characters are mostly compound words or words consisting of katakana characters. The stipulation that strings consisting of six or more characters appear in the JUMAN dictionary was set because long strings not present in the JUMAN dictionary were rarely found to be morphemes in our training corpus. Here we assume that compound words that do not appear in the JU-MAN dictionary can be divided into strings consisting of ve or less characters because compound words tend not to appear in dictionaries, and in fact, compound words which consist of six or more characters and do not appear in the dictionary were not found in our training corpus. Katakana strings that are not found in the JUMAN dictionary were assumed to be included in the dictionary as an entry having the part-of-speech \\Unknown(Major), Katakana(Minor).\" An optimal set of morphemes in a sentence is searched for by employing the Viterbi algorithm under the condition that connectivity rules dened between parts-of-speech in JUMAN must be met. The assigned part-of-speech in the optimal set is not always selected from the parts-of-speech attached to entries in the JUMAN dictionary, b u t may also be selected from the 53 categories of the M.E. model. It is dicult to select an appropriate category from the 53 when there is little training data, so we assume that every entry in the JUMAN dictionary has all possible parts-ofspeech, and the part-of-speech assigned to each morpheme is selected from those attached to the entry corresponding to the morpheme string.", "cite_spans": [ { "start": 122, "end": 149, "text": "(Kurohashi and Nagao, 1999)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "The features used in our experiments are listed in Table 1 . Each r o w i n T able 1 contains a feature type, feature values, and an experimental result that will be explained later. Each feature consists of a type and a value. The features are basically some attributes of the morpheme itself or those of the morpheme to the left of it. We used the 31,717 features that were found three or more times in the training corpus. The notations \\(0)\" and \\(-1)\" used in the feature type column in Table 1 respectively indicate a target string and the morpheme on the left of it.", "cite_spans": [], "ref_spans": [ { "start": 51, "end": 58, "text": "Table 1", "ref_id": "TABREF0" }, { "start": 492, "end": 499, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "The terms used in the table are the following:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "String: Strings which appeared as a morpheme ve or more times in the training corpus", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "Length: Length of a string POS: Part-of-speech. \\Major\" and \\Minor\" respectively indicate major and minor partof-speech categories as dened in JUMAN.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "Inf: Inection ty p e a s d e n e d i n J U M A N Dic: We use the JUMAN dictionary, which has about 200,000 entries (Kurohashi and Nagao, 1999) . \\Major&Minor\" indicates possible combinations between major and minor part-of-speech categories. When the target string is in the dictionary, the partof-speech a t t a c hed to the entry corresponding to the string is used as a feature value.", "cite_spans": [ { "start": 115, "end": 142, "text": "(Kurohashi and Nagao, 1999)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "If an entry has two or more parts-of-speech, the part-of-speech w h i c h leads to the highest probability in a sentence estimated from our model is selected as a feature value. JUMAN has another type of dictionary, which is called a phrase dictionary. Each entry in the phrase dictionary consists of one or more morphemes such as \\ (to, case marker), (wa, topic marker), (ie, say).\" JUMAN uses this dictionary to detect morphemes which need a longer context to be identied correctly. When the target string corresponds to the string of the left most morpheme in the phrase dictionary in JUMAN, the part-of-speech at- 106tached to the entry plus the information that it is in the phrase dictionary (such a s \\Verb&Phrase\") is used as a feature value.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "TOC: Types of characters used in a string. \\(Beginning)\" and \\(End)\" respectively represent the leftmost and rightmost characters of a string. When a string consists of only one character, the \\(Beginning)\" and \\(End)\" are the same character. \\TOC(0)(Transition)\" represents the transition from the leftmost character to the rightmost one in a string. \\TOC(-1)(Transition)\" represents the transition from the rightmost character in the adjacent morpheme on the left to the leftmost one in the target string. For example, when the adjacent morpheme on the left is \\ (sensei, teacher)\" and the target string is \\ (ni, c a s e marker),\" the feature value \\Kanji!Hiragana\" is selected.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "BB: Indicates whether or not the left side of a morpheme is a bunsetsu boundary.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "3.1" }, { "text": "Some results of the morphological analysis are listed in Table 2 . Recall is the percentage of morphemes in the test corpus whose segmentation and major POS tag are identied correctly. Precision is the percentage of all morphemes identied by the system that are identied correctly. F represents the F-measure and is dened by the following equation.", "cite_spans": [], "ref_spans": [ { "start": 57, "end": 64, "text": "Table 2", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Results and Discussion", "sec_num": "3.2" }, { "text": "F 0 measure = 2 2 Recall 2 P recision Recall + P recision Table 2 shows results obtained by using our method, by using JUMAN, and by using JU-MAN plus KNP (Kurohashi, 1998) . We show the result obtained using JUMAN plus KNP because JUMAN alone assigns an \\Unknown\" tag to katakana strings when they are not in the dictionary. All katakana strings not found (30,830/31,417) 98.31 in the dictionary are therefore evaluated as errors. KNP improves on JUMAN by replacing the \\Unknown\" tag with a \\Noun\" tag and disambiguating part-of-speech ambiguities which arise during the process of parsing when there is more than one JUMAN analysis with the same score.", "cite_spans": [ { "start": 155, "end": 172, "text": "(Kurohashi, 1998)", "ref_id": "BIBREF9" }, { "start": 357, "end": 372, "text": "(30,830/31,417)", "ref_id": null } ], "ref_spans": [ { "start": 58, "end": 65, "text": "Table 2", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Results and Discussion", "sec_num": "3.2" }, { "text": "The accuracy in segmentation and major POS tagging obtained with our method and that obtained with JUMAN were about 3% worse than that obtained with JUMAN plus KNP. We think the main reason for this was an insucient amount of training data and feature sets and the inconsistency of the corpus. The number of sentences in the training corpus was only about 8,000, and we did not use as many combined features as were proposed in Ref. (Uchimoto et al., 1999) . We w ere unable to use more training data or more feature sets because every string consisting of ve or less characters in our training corpus was used to train our model, so the amount of tokenized training data would have become too large and the training would not have been completed on the available machine if we had used more training data or more feature sets. The inconsistency of the corpus was due to the way the corpus was made. The Kyoto University corpus was made by manually correcting the output of JU-MAN plus KNP, and it is dicult to manually correct all of the inconsistencies in the output. The use of JUMAN plus KNP thus has an advantage over the use of our method when we evaluate a system's accuracy by using the Kyoto University corpus. For example, the number of morphemes whose rightmost character is \\ \" w as 153 in the test corpus, and they were all the s a m e a s t h o s e in the output of JUMAN plus KNP. There were three errors (about 2%) in the output of our system. There were several inconsistencies in the test corpus such a s \\ (seisan, Noun), (sha, Sux)(producer),\" and \\ (shouhi-sha, Noun)(consumer).\" They should have been corrected in the corpusmaking process to \\ (seisan, Noun), (sha, Sux)(producer),\" and \\ (shouhi, Noun), (sha, Sux)(consumer).\" It is dicult for our model to discriminate among these without over-training when there are such inconsistencies in the corpus. Other similar inconsistencies were, for example, \\ (geijutsuka, Noun)(artist)\" and \\ (kougei, Noun), (ka, Sux)(craftsman),\" \\ (keishi-cho, Noun)(the Metropolitan Police Board)\" and \\ (kensatsu, Noun), (cho, Noun)(the Public Prosecutor's Oce),\" and \\ (genjitsuteki, Adjective)(realistic)\" and \\ (risou, Noun), (teki, Sux)(ideal).\". If these had been corrected consistently when making the corpus, the accuracy obtained by our method could have b e e n better than that shown in Table 2. A study on corpus revision should be undertaken to resolve this issue. We believe i t c a n be resolved by using our trained model. There is a high possibility that a morpheme lacks consistency in the training corpus when its probability, re-estimated by our model, is low. Thus a method which detects morphemes having a low probability can identify those lacking consistency in the training corpus. We intend to try this in the future.", "cite_spans": [ { "start": 433, "end": 456, "text": "(Uchimoto et al., 1999)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Results and Discussion", "sec_num": "3.2" }, { "text": "In our model, dictionary information and certain characteristics of unknown words are reected as features, as shown in Table 1 . \\String\" and \\Dic\" reect the dictionary information, 2 and \\Length\" and \\TOC\"(types of characters) reect the characteristics of unknown words. Therefore, our model can not only consult a dictionary but can also detect unknown words. analysis without the complete feature set. Almost all of the feature sets improved accuracy. The contribution of the dictionary information was especially signicant. There were cases, however, in which the use of dictionary information led to a decrease in the accuracy. For example, we found these erroneous segmentations: \\ (umi, sea) (ni, case marker) (kaketa, bet) (romanha, the Romantic school) \" and \\ (aranami, raging waves) (ni, case marker) (make, lose) (naishin, one's inmost heart) (to, case marker) \" (Underlined strings were errors.) when the correct segmentations were: \\ (umi, sea) (ni, case marker) (kaketa, bet) (roman, romance) (wa, topic marker) \" and \\ (aranami, raging waves) (ni, case marker) (makenai, not to lose) (kokoro, heart) (to, case marker) \" ( \\ \" i n d i c a t e s a m o r p h ological boundary.).", "cite_spans": [], "ref_spans": [ { "start": 119, "end": 126, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Features and Accuracy", "sec_num": "3.3" }, { "text": "These errors were caused by nonstandard en-tries in the JUMAN dictionary. The dictionary had not only the usual notation using kanji characters, \\ \" and \\ ,\" but also the uncommon notation using hiragana strings, \\ \" and \\ \". To prevent this type of error, it is necessary to remove nonstandard entries from the dictionary or to investigate the frequency of such entries in large corpora and to use it as a feature.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features and Accuracy", "sec_num": "3.3" }, { "text": "Accuracy and the Amount of", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "3.4", "sec_num": null }, { "text": "The accuracies (F-measures) for the training corpus and the test corpus are shown in Figure 1 plotted against the number of sentences used for training. The learning curve s h o ws that we can expect improvement i f w e use more training data.", "cite_spans": [], "ref_spans": [ { "start": 85, "end": 93, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Training Data", "sec_num": null }, { "text": "The strength of our method is that it can identify morphemes when they are unknown words and can assign appropriate parts-of-speech to them. For example, the nouns \\ (Souseki)\" and \\ (Rohan)\" are not found in the JU- MAN dictionary. JUMAN plus KNP analyzes them simply as \\ (Noun) (Noun)\" and \\ (Adverb) (Noun),\" whereas our system analyzes both of them correctly. Our system correctly identied them as names of people even though they were not in the dictionary and did not appear as features in our M.E. model. Since these names, or proper nouns, are newly coined and can be represented by a variety of expressions, no proper nouns can be included in a dictionary, nor can they appear in a training corpus; this means that proper nouns could easily beunknown words. We investigated the accuracy of our method in identifying morphemes when they are unknown words, and the results are listed in Table 3 . The rst row in each section shows the recall for the morphemes that were unknown words. The second row in each section shows the percentage of morphemes whose segmentation and \\minor\" POS tag were identied correctly. The dierence between the rst and second lines, the third and fourth lines, and fth and sixth lines is the denition of unknown words. Unknown words were dened respectively as words not found in the dictionary nor in our training corpus, as words not found in the dictionary nor in our features, and as words not found in the dictionary. Our accuracy, shown as the second rows in Table 3 was more than 5% better than that of JUMAN plus KNP for each denition. These results show that our model can eciently learn the characteristics of unknown words, especially those of proper nouns such as the names of people, organizations, and locations.", "cite_spans": [], "ref_spans": [ { "start": 895, "end": 902, "text": "Table 3", "ref_id": "TABREF3" }, { "start": 1500, "end": 1507, "text": "Table 3", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Unknown Words and Accuracy", "sec_num": "3.5" }, { "text": "Several methods based on statistical models have been proposed for the morphological analysis of Japanese sentences. An F-measure of about 96% was achieved by a method based on a hidden Markov model (HMM) (Takeuchi and Matsumoto, 1997) and by one based on a v ariable-memory Markov model (Haruno and Matsumoto, 1997; Kitauchi et al., 1999) . Although the accuracy obtained with these methods was better than that obtained with ours, their accuracy cannot be compared directly with that of our method because their partof-speech categories dier from ours. And an advantage of our model is that it can handle unknown words, whereas their models do not handle unknown words well. In their models, unknown words are divided into a combination of a word consisting of one character and known words. Haruno and Matsumoto (Haruno and Matsumoto, 1997) achieved a recall of about 96% when using trigram or greater information, but achieved a recall of only 94% when using bigram information. This leads us to believe t h a t we could obtain better accuracy if we use trigram or greater information. We plan to do so in future work. Two approaches have b e e n u s e d t o d e a l w i t h unknown words: acquiring unknown words from corpora and putting them into a dictionary (e.g., (Mori and Nagao, 1996) ) and developing a model that can identify unknown words correctly (e.g., (Kashioka et al., 1997; Nagata, 1999) ). Nagata reported a recall of about 40% for unknown words (Nagata, 1999) . As shown in Table 3 , our method achieved a recall of 69.90% for unknown words. Our accuracy was about 30% better than his. It is dicult to compare his method with ours directly because he used a dierent corpus (the EDR corpus), but the part-of-speech categories and the definition of morphemes he used were similar to ours. Thus, this comparison is helpful in evaluating our method. There are no spaces between morphemes in Japanese. In general, therefore, detecting whether a given string is an unknown word or is not a morpheme is dicult when it is not found in the dictionary, nor in the training corpus. However, our model learns whether or not a given string is a morpheme and has a huge amount of data for learning what in a corpus is not a morpheme. Therefore, we believe that the characteristics of our model led to its good results for identifying unknown words.", "cite_spans": [ { "start": 205, "end": 235, "text": "(Takeuchi and Matsumoto, 1997)", "ref_id": "BIBREF19" }, { "start": 288, "end": 316, "text": "(Haruno and Matsumoto, 1997;", "ref_id": "BIBREF4" }, { "start": 317, "end": 339, "text": "Kitauchi et al., 1999)", "ref_id": "BIBREF6" }, { "start": 794, "end": 843, "text": "Haruno and Matsumoto (Haruno and Matsumoto, 1997)", "ref_id": "BIBREF4" }, { "start": 1273, "end": 1295, "text": "(Mori and Nagao, 1996)", "ref_id": "BIBREF11" }, { "start": 1370, "end": 1393, "text": "(Kashioka et al., 1997;", "ref_id": "BIBREF5" }, { "start": 1394, "end": 1407, "text": "Nagata, 1999)", "ref_id": "BIBREF14" }, { "start": 1467, "end": 1481, "text": "(Nagata, 1999)", "ref_id": "BIBREF14" } ], "ref_spans": [ { "start": 1496, "end": 1503, "text": "Table 3", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Related Work", "sec_num": "4" }, { "text": "Mori and Nagao proposed a model that can consult a dictionary (Mori and Nagao, 1998) ; they reported an F-measure of about 92 when using the EDR corpus and of about 95 when using the Kyoto University corpus. Their slight improvement in accuracy by using dictionary information resulted in an F-measure of about 0.2, while our improvement was about 1.7. Their accuracy of 95% when using the Kyoto University corpus is similar to ours, but they added to their dictionary all of the words appearing in the training corpus. Therefore, their experiment had to deal with fewer unknown words than ours did.", "cite_spans": [ { "start": 62, "end": 84, "text": "(Mori and Nagao, 1998)", "ref_id": "BIBREF12" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "4" }, { "text": "With regard to the morphological analysis of English sentences, methods for part-ofspeech tagging based on an HMM (Cutting et al., 1992) , a variable-memory Markov model (Sch utze and Singer, 1994) , a decision tree model (Daelemans et al., 1996) , an M.E. model (Ratnaparkhi, 1996) , a neural network model (Schmid, 1994) , and a transformation-based error-driven learning model (Brill, 1995) have been proposed, as well as a combined method (M arquez and Padr o, 1997; van Halteren et al., 1998) . On available machines, however, these models cannot handle a large amount of lexical information. We think that our model, which can not only consult a dictionary with a large amount of lexical information, but can also identify unknown words by learning certain characteristics, has the potential to achieve good accuracy for part-of-speech tagging in English. We plan to apply our model to English sentences.", "cite_spans": [ { "start": 114, "end": 136, "text": "(Cutting et al., 1992)", "ref_id": null }, { "start": 170, "end": 197, "text": "(Sch utze and Singer, 1994)", "ref_id": null }, { "start": 222, "end": 246, "text": "(Daelemans et al., 1996)", "ref_id": "BIBREF3" }, { "start": 263, "end": 282, "text": "(Ratnaparkhi, 1996)", "ref_id": "BIBREF15" }, { "start": 308, "end": 322, "text": "(Schmid, 1994)", "ref_id": "BIBREF16" }, { "start": 380, "end": 393, "text": "(Brill, 1995)", "ref_id": "BIBREF1" }, { "start": 443, "end": 470, "text": "(M arquez and Padr o, 1997;", "ref_id": null }, { "start": 471, "end": 497, "text": "van Halteren et al., 1998)", "ref_id": "BIBREF21" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "4" }, { "text": "This paper described a method for morphological analysis based on a maximum entropy (M.E.) model. This method uses a model that can not only consult a dictionary but can also identify unknown words by learning certain characteristics. To learn these characteristics, we focused on such information as whether or not a string is found in a dictionary and what types of characters are used in a string. The model estimates how likely a string is to be a morpheme according to the information on hand. When our method was used to identify morpheme segments in sentences in the Kyoto University corpus and to identify the major parts-of-speech of these morphemes, the recall and precision were respectively 95.80% and 95.09%. In our experiments without each feature set shown in Tables 1, we found that dictionary information signicantly contributes to improving accuracy. We also found that our model can eciently learn the characteristics of unknown words, especially proper nouns such as the names of people, organizations, and locations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5" }, { "text": "Not only morphemes but also bunsetsus can be identied by considering the information related to their bunsetsu boundaries.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "\\String\" indicates strings that make up a morpheme and were found ve or more times in the training corpus. Using this information as features in our M.E. model corresponds to consulting a dictionary constructed from the training corpus.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "A Maximum Entropy Approach to Natural Language Processing", "authors": [ { "first": "Adam", "middle": [ "L" ], "last": "Berger", "suffix": "" }, { "first": "Stephen", "middle": [ "A" ], "last": "Della Pietra", "suffix": "" }, { "first": "Vincent", "middle": [ "J Della" ], "last": "Pietra", "suffix": "" } ], "year": 1996, "venue": "Computational Linguistics", "volume": "22", "issue": "1", "pages": "39--71", "other_ids": {}, "num": null, "urls": [], "raw_text": "Adam L. Berger, Stephen A. Della Pietra, and Vincent J. Della Pietra. 1996. A Max- imum Entropy Approach to Natural Lan- guage Processing. Computational Linguis- tics, 22(1):39{71.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Transformation-Based Error-Driven Learning and Natural Language Processing: A Case Study in Part-of-Speech T agging", "authors": [ { "first": "Eric", "middle": [], "last": "Brill", "suffix": "" } ], "year": 1995, "venue": "Computational Linguistics", "volume": "21", "issue": "4", "pages": "543--565", "other_ids": {}, "num": null, "urls": [], "raw_text": "Eric Brill. 1995. Transformation-Based Error- Driven Learning and Natural Language Pro- cessing: A Case Study in Part-of-Speech T ag- ging. Computational Linguistics, 21(4):543{ 565.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Pedersen, and Penelope Sibun. 1992. A Practical Part-of-Speech Tagger", "authors": [ { "first": "Doung", "middle": [], "last": "Cutting", "suffix": "" }, { "first": "Julian", "middle": [], "last": "Kupiec", "suffix": "" } ], "year": null, "venue": "Proceedings of the Third Conference on Applied Natural Language Processing", "volume": "", "issue": "", "pages": "133--140", "other_ids": {}, "num": null, "urls": [], "raw_text": "Doung Cutting, Julian Kupiec, Jan Peder- sen, and Penelope Sibun. 1992. A Practical Part-of-Speech Tagger. In Proceedings of the Third Conference on Applied Natural Lan- guage Processing, pages 133{140.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "MBT: A Memory-Based Part-of-Speech Tagger-Generator", "authors": [ { "first": "Walter", "middle": [], "last": "Daelemans", "suffix": "" }, { "first": "Jakub", "middle": [], "last": "Zavrel", "suffix": "" }, { "first": "Peter", "middle": [], "last": "Berck", "suffix": "" }, { "first": "Steven", "middle": [], "last": "Gills", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 4th Workshop on Very Large Corpora", "volume": "", "issue": "", "pages": "1--14", "other_ids": {}, "num": null, "urls": [], "raw_text": "Walter Daelemans, Jakub Zavrel, Peter Berck, and Steven Gills. 1996. MBT: A Memory- Based Part-of-Speech Tagger-Generator. In Proceedings of the 4th Workshop on Very Large Corpora, pages 1{14.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Mistake-Driven Mixture of Hierarchical-Tag Context Trees", "authors": [ { "first": "Masahiko", "middle": [], "last": "Haruno", "suffix": "" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 1997, "venue": "Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics (ACL)", "volume": "", "issue": "", "pages": "230--237", "other_ids": {}, "num": null, "urls": [], "raw_text": "Masahiko Haruno and Yuji Matsumoto. 1997. Mistake-Driven Mixture of Hierarchical-Tag Context Trees. In Proceedings of the 35th An- nual Meeting of the Association for Compu- tational Linguistics (ACL), pages 230{237.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Decision-Tree Morphological Analysis without a Dictionary for Japanese", "authors": [ { "first": "Hideki", "middle": [], "last": "Kashioka", "suffix": "" }, { "first": "Stephen", "middle": [ "G" ], "last": "Eubank", "suffix": "" }, { "first": "Ezra", "middle": [ "W" ], "last": "Black", "suffix": "" } ], "year": 1997, "venue": "Proceedings of the Natural Language Processing Pacic Rim Symposium", "volume": "", "issue": "", "pages": "541--544", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hideki Kashioka, Stephen G. Eubank, and Ezra W. Black. 1997. Decision-Tree Mor- phological Analysis without a Dictionary for Japanese. In Proceedings of the Natural Lan- guage Processing Pacic Rim Symposium, pages 541{544.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Probabilistic Model Learning for Japanese Morphological Analysis by Error-driven Feature Selection", "authors": [ { "first": "Akira", "middle": [], "last": "Kitauchi", "suffix": "" }, { "first": "Takehito", "middle": [], "last": "Utsuro", "suffix": "" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 1999, "venue": "", "volume": "40", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Akira Kitauchi, Takehito Utsuro, and Yuji Mat- sumoto. 1999. Probabilistic Model Learn- ing for Japanese Morphological Analysis by Error-driven Feature Selection. Transactions of Information Processing Society of Japan, 40(5):2325{2337. (in Japanese).", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Building a Japanese Parsed Corpus while Improving the Parsing System", "authors": [ { "first": "Sadao", "middle": [], "last": "Kurohashi", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1997, "venue": "Proceedings of the Natural Language Processing Pacic Rim Symposium", "volume": "", "issue": "", "pages": "451--456", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sadao Kurohashi and Makoto Nagao. 1997. Building a Japanese Parsed Corpus while Im- proving the Parsing System. In Proceedings of the Natural Language Processing Pacic Rim Symposium, pages 451{456.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Japanese Morphological Analysis System JU-MAN Version 3.61", "authors": [ { "first": "Sadao", "middle": [], "last": "Kurohashi", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1999, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sadao Kurohashi and Makoto Nagao, 1999. Japanese Morphological Analysis System JU- MAN Version 3.61. Department of Informat- ics, Kyoto University.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Japanese Dependency/Case Structure Analyzer KNP Version 2.0b6", "authors": [ { "first": "Sadao", "middle": [], "last": "Kurohashi", "suffix": "" } ], "year": 1998, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sadao Kurohashi, 1998. Japanese Depen- dency/Case Structure Analyzer KNP Ver- sion 2.0b6. Department of Informatics, Ky- oto University.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "A Flexible POS Tagger Using an Automatically Acquired Language Model", "authors": [], "year": 1997, "venue": "Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics (ACL)", "volume": "", "issue": "", "pages": "238--252", "other_ids": {}, "num": null, "urls": [], "raw_text": "Llu is M arquez and Llu is Padr o. 1997. A Flexi- ble POS Tagger Using an Automatically Ac- quired Language Model. In Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics (ACL), pages 238{252.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Word Extraction from Corpora and Its Partof-Speech Estimation Using Distributional Analysis", "authors": [ { "first": "Shinsuke", "middle": [], "last": "Mori", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 16th International Conference on Computational Linguistics (COLING96)", "volume": "", "issue": "", "pages": "1119--1122", "other_ids": {}, "num": null, "urls": [], "raw_text": "Shinsuke Mori and Makoto Nagao. 1996. Word Extraction from Corpora and Its Part- of-Speech Estimation Using Distributional Analysis. In Proceedings of the 16th Interna- tional Conference on Computational Linguis- tics (COLING96), pages 1119{1122.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "An Improvement of a Morphological Analysis by a Morpheme Clustering", "authors": [ { "first": "Shinsuke", "middle": [], "last": "Mori", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1998, "venue": "Journal of Natural Language Processing", "volume": "5", "issue": "2", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Shinsuke Mori and Makoto Nagao. 1998. An Improvement of a Morphological Analysis by a Morpheme Clustering. Journal of Nat- ural Language Processing, 5(2):75{103. (in Japanese).", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "A Stochastic Japanese Morphological Analyzer Using a Forward-DP Backward-A 3 N-Best Search Algorithm", "authors": [ { "first": "Masaaki", "middle": [], "last": "Nagata", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 15th International Conference on Computational Linguistics (COL-ING94)", "volume": "", "issue": "", "pages": "201--207", "other_ids": {}, "num": null, "urls": [], "raw_text": "Masaaki Nagata. 1994. A Stochastic Japanese Morphological Analyzer Using a Forward-DP Backward-A 3 N-Best Search Algorithm. In Proceedings of the 15th International Con- ference on Computational Linguistics (COL- ING94), pages 201{207.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "A Pa r t o f S p e e c h Estimation Method for Japanese Unknown Words using a Statistical Model of Morphology and Context", "authors": [ { "first": "Masaaki", "middle": [], "last": "Nagata", "suffix": "" } ], "year": 1999, "venue": "Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics (ACL)", "volume": "", "issue": "", "pages": "277--284", "other_ids": {}, "num": null, "urls": [], "raw_text": "Masaaki Nagata. 1999. A Pa r t o f S p e e c h Esti- mation Method for Japanese Unknown Words using a Statistical Model of Morphology and Context. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics (ACL), pages 277{284.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "A Maximum Entropy Model for Part-Of-Speech Tagging", "authors": [ { "first": "Adwait", "middle": [], "last": "Ratnaparkhi", "suffix": "" } ], "year": 1996, "venue": "Conference o n Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "133--142", "other_ids": {}, "num": null, "urls": [], "raw_text": "Adwait Ratnaparkhi. 1996. A Maximum En- tropy Model for Part-Of-Speech Tagging. In Conference o n Empirical Methods in Natural Language Processing, pages 133{142.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Part-Of-Speech T agging with Neural Networks", "authors": [ { "first": "Helmut", "middle": [], "last": "Schmid", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 15th International Conference on Computational Linguistics (COLING94)", "volume": "", "issue": "", "pages": "172--176", "other_ids": {}, "num": null, "urls": [], "raw_text": "Helmut Schmid. 1994. Part-Of-Speech T agging with Neural Networks. In Proceedings of the 15th International Conference on Computa- tional Linguistics (COLING94), pages 172{ 176.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Partof-Speech T agging Using a Variable Memory Markov Model", "authors": [ { "first": "Yoram", "middle": [], "last": "Hinrich S C H Utze", "suffix": "" }, { "first": "", "middle": [], "last": "Singer", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 32nd", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hinrich S c h utze and Yoram Singer. 1994. Part- of-Speech T agging Using a Variable Memory Markov Model. In Proceedings of the 32nd", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Annual Meeting of the Association for Computational Linguistics (ACL)", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "181--187", "other_ids": {}, "num": null, "urls": [], "raw_text": "Annual Meeting of the Association for Com- putational Linguistics (ACL), pages 181{187.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "HMM Parameter Learning for Japanese Morphological Analyzer", "authors": [ { "first": "Koichi", "middle": [], "last": "Takeuchi", "suffix": "" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 1997, "venue": "", "volume": "83", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Koichi Takeuchi and Yuji Matsumoto. 1997. HMM Parameter Learning for Japanese Morphological Analyzer. Transactions of Information Processing Society of Japan, 83(3):500{509. (in Japanese).", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Japanese Dependency Structure Analysis Based on Maximum Entropy Models", "authors": [ { "first": "Kiyotaka", "middle": [], "last": "Uchimoto", "suffix": "" }, { "first": "Satoshi", "middle": [], "last": "Sekine", "suffix": "" }, { "first": "Hitoshi", "middle": [], "last": "Isahara", "suffix": "" } ], "year": 1999, "venue": "Proceedings of the Ninth Conference of the European Chapter of the Association for Computational Linguistics (EACL'99)", "volume": "", "issue": "", "pages": "196--203", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kiyotaka Uchimoto, Satoshi Sekine, and Hi- toshi Isahara. 1999. Japanese Dependency Structure Analysis Based on Maximum En- tropy Models. In Proceedings of the Ninth Conference of the European Chapter of the Association for Computational Linguistics (EACL'99), pages 196{203.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Improving Data Driven Wordclass Tagging by System Combination", "authors": [ { "first": "Jakub", "middle": [], "last": "Hans Van Halteren", "suffix": "" }, { "first": "Walter", "middle": [], "last": "Zavrel", "suffix": "" }, { "first": "", "middle": [], "last": "Daelemans", "suffix": "" } ], "year": 1998, "venue": "Proceedings of the COLING-ACL '98", "volume": "", "issue": "", "pages": "491--497", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hans van Halteren, Jakub Zavrel, and Walter Daelemans. 1998. Improving Data Driven Wordclass Tagging by System Combination. In Proceedings of the COLING-ACL '98, pages 491{497.", "links": null } }, "ref_entries": { "FIGREF0": { "uris": null, "num": null, "type_str": "figure", "text": "Relation between accuracy and the number of training sentences." }, "TABREF0": { "num": null, "html": null, "text": "Features.", "content": "
Feature
", "type_str": "table" }, "TABREF1": { "num": null, "html": null, "text": "Results of Experiments (Segmentation and major POS tagging).", "content": "
RecallPrecisionF-measure
Our method 95.80% (29,986/31,302) 95.09% (29,986/31,467) JUMAN 95.25% (29,814/31,302) 94.90% (29,814/31,417) JUMAN+KNP 98.49% (30,830/31,302) 98.13%95.44 95.07
", "type_str": "table" }, "TABREF2": { "num": null, "html": null, "text": "shows the results of an", "content": "
97\"training\"
\"testing\"
96.5
96
95.5
F-measure95
94.5
94
93.5
010002000300040005000600070008000
Number of Sentences
", "type_str": "table" }, "TABREF3": { "num": null, "html": null, "text": "Accuracy for unknown words (Recall).", "content": "
Segmentation and major POS tagging minor POS tagging Segmentation and
For words not found in the dictionary nor in our training corpus
Our method JUMAN+KNP 79.29% (490/618) 69.90% (432/618)27.51% (170/618) 20.55% (127/618)
For words not found in the dictionary nor in our features
Our method JUMAN+KNP 85.70% (809/944) 76.17% (719/944)32.20% (304/944) 27.22% (257/944)
For words not found in the dictionary
Our method 82.40% (1,138/1,381) 49.24% (680/1,381) JUMAN+KNP 89.79% (1,240/1,381) 38.60% (533/1,381)
", "type_str": "table" } } } }