| { |
| "paper_id": "N07-1024", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T14:47:33.047791Z" |
| }, |
| "title": "Hybrid Models for Semantic Classification of Chinese Unknown Words", |
| "authors": [ |
| { |
| "first": "Xiaofei", |
| "middle": [], |
| "last": "Lu", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Pennsylvania State University University Park", |
| "location": { |
| "postCode": "16802", |
| "region": "PA", |
| "country": "USA" |
| } |
| }, |
| "email": "xxl13@psu.edu" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "This paper addresses the problem of classifying Chinese unknown words into fine-grained semantic categories defined in a Chinese thesaurus. We describe three novel knowledge-based models that capture the relationship between the semantic categories of an unknown word and those of its component characters in three different ways. We then combine two of the knowledge-based models with a corpus-based model which classifies unknown words using contextual information. Experiments show that the knowledge-based models outperform previous methods on the same task, but the use of contextual information does not further improve performance.", |
| "pdf_parse": { |
| "paper_id": "N07-1024", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "This paper addresses the problem of classifying Chinese unknown words into fine-grained semantic categories defined in a Chinese thesaurus. We describe three novel knowledge-based models that capture the relationship between the semantic categories of an unknown word and those of its component characters in three different ways. We then combine two of the knowledge-based models with a corpus-based model which classifies unknown words using contextual information. Experiments show that the knowledge-based models outperform previous methods on the same task, but the use of contextual information does not further improve performance.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "Research on semantic annotation has focused primarily on word sense disambiguation (WSD), i.e., the task of determining the appropriate sense for each instance of a polysemous word out of a set of senses defined for the word in some lexicon. Much less work has been done on semantic classification of unknown words, i.e., words that are not listed in the lexicon. However, real texts typically contain a large number of unknown words. Successful classification of unknown words is not only useful for lexical acquisition, but also necessary for natural language processing (NLP) tasks that require semantic annotation.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "This paper addresses the problem of classifying Chinese unknown words into fine-grained semantic categories defined in a Chinese thesaurus, Cilin (Mei et al., 1984) . This thesaurus classifies over 70,000 words into 12 major categories, including human (A), concrete object (B), time and space (C), abstract object (D), attributes (E), actions (F), mental activities (G), activities (H), physical states (I), relations (J), auxiliaries (K), and honorifics (L). The 12 major categories are further divided into 94 medium categories, which in turn are subdivided into 1428 small categories. Each small category contains synonyms that are close in meaning. For example, under the major category D, the medium category Dm groups all words that refer to institutions, and the small category Dm05 groups all words that refer to educational institutions, e.g., \u5b66\u6821 xu\u00e9xi\u00e0o 'school'. Unknown word classification involves a much larger search space than WSD. In classifying words into small categories in Cilin, the search space for a polysemous known word consists of all the categories the word belongs to, but that for an unknown word consists of all the 1428 small categories.", |
| "cite_spans": [ |
| { |
| "start": 146, |
| "end": 164, |
| "text": "(Mei et al., 1984)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Research on WSD has concentrated on using contextual information, which may be limited for infrequent unknown words. On the other hand, Chinese characters carry semantic information that is useful for predicting the semantic properties of the words containing them. We present three novel knowledge-based models that capture the relationship between the semantic categories of an unknown word and those of its component characters in three different ways, and combine two of them with a corpus-based model that uses contextual information to classify unknown words. Experiments show that the combined knowledge-based model achieves an accuracy of 61.6% for classifying unknown words into small categories in Cilin, but the use of contextual information does not further improve performance.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "The rest of the paper is organized as follows. Section 2 details the three novel knowledgebased models proposed for this task. Section 3 describes a corpus-based model. Section 4 reports the experiment results of the proposed models. Section 5 compares these results with previous results. Section 6 concludes the paper and points to avenues for further research.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "This section describes three novel knowledgebased models for semantic classification of Chinese unknown words, including an overlappingcharacter model, a character-category association model, and a rule-based model. These models model the relationship between the semantic category of an unknown word and those of its component characters in three different ways.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Knowledge-based Models", |
| "sec_num": "2" |
| }, |
| { |
| "text": "The baseline model predicts the category of an unknown word by counting the number of overlapping characters between the unknown word and the member words in each category. As words in the same category are similar in meaning and the meaning of a Chinese word is generally the composition of the meanings of its characters, it is common for words in the same category to share one or more character. This model tests the hypothesis that speakers draw upon the repertoire of characters that relate to a concept when creating new words to realize it.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "For each semantic category in Cilin, the set of unique characters in its member words are extracted, and the number of times each character occurs in word-initial, word-middle, and wordfinal positions is recorded. With this information, we develop two variants of the baseline model, which differ from each other in terms of whether it takes into consideration the positions in which the characters occur in words.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "In variant 1, the score of a category is the sum of the number of occurrences of each character of the target word in the category, as in (1), where t j denotes a category, w denotes the target word, c i denotes the ith character in w, n is the length of w, and f(c i ) is the frequency of c i in t j .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "(1)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "\u2211 = = n i i j c f w t Score 1 ) ( ) , (", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "In variant 2, the score of a category is the sum of the number of occurrences of each character of the unknown word in the category in its corresponding position, as in (2), where p i denotes the position of c i in w, which could be word-initial, word-middle, or word-final, and f(c i ,p i ) denotes the frequency of c i in position p i in t j .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "(2)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "= ) , ( w t Score j \u2211 = n i i i p c f 1 ) , (", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "In each variant, the category with the maximum score for a target word is proposed as the category of the word.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Baseline Model", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "The relationship between the semantic category of an unknown word and those of its component characters can also be captured in a more sophisticated way using information-theoretical models. We use two statistical measures, mutual information and \u03c7 2 , to compute character-category associations and word-category associations. Chen (2004) used the \u03c7 2 measure to compute character-character and word-word associations, but not word-category associations. We use word-category associations to directly predict the semantic categories of unknown words.", |
| "cite_spans": [ |
| { |
| "start": 328, |
| "end": 339, |
| "text": "Chen (2004)", |
| "ref_id": "BIBREF0" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "The mutual information and \u03c7 2 measures are calculated as in 3and 4, where Asso(c,t j ) denotes the association between a character c and a semantic category t j , and P(X) and f(X) denote the probability and frequency of X respectively.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "(3)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": ") ( ) ( ) , ( log ) , ( j j j MI t P c P t c P t c Asso = (4) ) , ( max ) , ( ) , ( 2 k k j j t c t c t c Asso \u03b1 \u03b1 \u03c7 = (5) ) ( ) ( )] , ( [ ) , ( 2 j j j t f c f t c f t c + = \u03b1", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "Once the character-category associations are calculated, the association between a word w and a category t j , Asso(w,t j ), can be calculated as the sum of the weighted associations between each of the word's characters and the category, as in (6), where c i denotes the ith character of w, |w| denotes the length of w, and \u03bb i denotes the weight of Asso(c i ,t j ). The \u03bb's add up to 1. The weights are determined empirically based on the positions of the characters in the word.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "(6) = ) , ( j t w Asso \u2211 = | | 1 ) , ( w i j i i t c Asso \u03bb", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "As in variant 2 of the baseline model, the character-category association model can also be made sensitive to the positions in which the characters occur in the words. To this end, we first need to compute the position-sensitive associations between a category and a character in the word-initial, word-middle, and word-final positions separately. The position-sensitive association between an unknown word and a category can then be computed as the sum of the weighted position-sensitive associations between each of its characters and the category.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "Once the word-category associations are computed, we can propose the highest ranked category or a ranked list of categories for each unknown word.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Character-Category Associations", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "The third knowledge-based model uses linguistic rules to classify unknown words based on the syntactic and semantic categories of their component characters. Rule-based models have not been used for this task before. However, there are some regularities in the relationship between the semantic categories of unknown words and those of their component characters that can be captured in a more direct and effective way by linguistic rules than by statistical models.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Rule-Based Model", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "A separate set of rules are developed for words of different lengths. Rules are initially developed based on knowledge about Chinese word formation, and are then refined by examining the development data. In general, the complete rule set takes a few hours to develop.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Rule-Based Model", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "The rule in 7is developed for bisyllabic unknown words. This rule proposes the common category of a bisyllabic word's two characters as its category. It is especially useful for words with a parallel structure, i.e., words whose two characters have the same meaning and syntactic category, e.g., \u574d\u584c t\u0101nt\u0101 'collapse', where \u574d t\u0101n and \u584c t\u0101 both mean 'collapse' and share the category Id05. The thresholds for f A and f B are determined empirically and are both set to 1 if AB is a noun and to 0 and 3 respectively otherwise.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Rule-Based Model", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "(7) For a bisyllabic word AB, if A and B share a category c 1 , let f A and f B denote the number of times A and B occur in word-initial and word-final positions in c respectively. If f A and f B both surpass the predetermined thresholds, propose c for AB.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Rule-Based Model", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "A number of rules are developed for trisyllabic words. While most rules in the model are general, the first rule in this set is rather specific, as it handles words with three specific prefixes, \u5927 d\u00e0 'big', \u5c0f xi\u0103o 'little', and \u8001 l\u0103o 'old', which usually do not change the category of the root word. The other four rules again utilize the categories of the unknown word's component characters. The rules in (8b) and (8c) are similar to the rule in (7). The ones in (8d) and (8e) search for neighbor words with a similar structure as the target word. Eligible neighbors have a common morpheme with the target word in the same position and a second morpheme that shares a category with the second morpheme of the target word. For example, an eligible neighbor for \u63a8\u9500\u5546 tu\u012bxi\u0101o-sh\u0101ng 'sales-man' is \u9500\u552e\u5546 xi\u0101osh\u00f2u-sh\u0101ng 'distribut-or'. These two words share the morpheme \u5546 sh\u0101ng 'businessman' in the word-final position, and the morphemes \u63a8\u9500 tu\u012bxi\u0101o 'to market' and \u9500\u552e xi\u0101osh\u00f2u 'distribute' share the category He03. The rule in (8d) therefore applies in this case. 8 The rules for four-character words are given in (9). Like the rules in (8d) and (8e), these rules also search for neighbors of the target word.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Rule-Based Model", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "(9) For a four-character word ABCD:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Rule-Based Model", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "a. If there is a word XYZD/YZD such that XYZ/YZ and ABC share a category, propose the category of XYZ/YZ for ABCD. b. If there is a word ABCX such that X and D share a category, propose the category of ABCX for ABCD. c. If there is a word XYCD such that XY and AB share a category, propose the category of XYCD for ABCD. d. If there is a word XBCD/BCD, propose the category of XBCD/BCD for ABCD.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Rule-Based Model", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "The knowledge-based models described above classify unknown words using information about the syntactic and semantic categories of their component characters. Another useful source of information is the context in which unknown words occur. While contextual information is the primary source of information used in WSD research and has been used for acquiring semantic lexicons and classifying unknown words in other languages (e.g., Roark and Charniak 1998; Ci-aramita 2003; Curran 2005) , it has been used in only one previous study on semantic classification of Chinese unknown words (Chen and Lin, 2000) . Part of the goal of this study is to investigate whether and how these two different sources of information can be combined to improve performance on semantic classification of Chinese unknown words. To this end, we first use the knowledge-based models to propose a list of five candidate categories for the target word, then extract a generalized context for each category in Cilin from a corpus, and finally compute the similarity between the context of the target word and the generalized context of each of its candidate categories. Comparing the context of the target word with generalized contexts of categories instead of contexts of individual words alleviates the data-sparseness problem, as infrequent words have limited contextual information. Limiting the search space for each target word to the top five candidate categories reduces the computational cost that comes with the full search space.", |
| "cite_spans": [ |
| { |
| "start": 434, |
| "end": 458, |
| "text": "Roark and Charniak 1998;", |
| "ref_id": "BIBREF7" |
| }, |
| { |
| "start": 459, |
| "end": 475, |
| "text": "Ci-aramita 2003;", |
| "ref_id": null |
| }, |
| { |
| "start": 476, |
| "end": 488, |
| "text": "Curran 2005)", |
| "ref_id": "BIBREF4" |
| }, |
| { |
| "start": 587, |
| "end": 607, |
| "text": "(Chen and Lin, 2000)", |
| "ref_id": "BIBREF3" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A Corpus-Based Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "A generalized context for each semantic category is built from the contexts of its member words. This is done based on the assumption that as the words in the same category have the same or similar meaning, they tend to occur in similar contexts. In terms of context extraction and representation, we need to consider four factors.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Context Extraction and Representation", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "The issue here is whether to include the contexts of polysemous member words in building the generalized context of a category. Including these contexts without discrimination introduces noise. To measure the effect of such noise, we build two versions of generalized context for each category, one using contexts of unambiguous member words only, and the other using contexts of all member words.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Member Words", |
| "sec_num": null |
| }, |
| { |
| "text": "Context Words There are two issues in selecting words for context representation. First, words that contribute little information to the discrimination of meaning of other words, including conjunctions, numerals, auxiliaries, and non-Chinese sequences, are excluded. Second, to model the effect of frequency on the context words' contribution to meaning discrimination, we use two sets of context words: one consists of the 1000 most frequent words in the corpus; the other consists of all words in the corpus.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Member Words", |
| "sec_num": null |
| }, |
| { |
| "text": "Window Size For WSD, both topical context and microcontext have been used (Ide and V\u00e9ronis 1998) . Topical context includes substantive words that co-occur with the target word within a larger window, whereas microcontext includes words in a small window around the target word. We experiment with topical context and microcontext with window sizes of 100 and 6 respectively (i.e., 50 and 3 words to the left and right of the target word respectively).", |
| "cite_spans": [ |
| { |
| "start": 74, |
| "end": 96, |
| "text": "(Ide and V\u00e9ronis 1998)", |
| "ref_id": "BIBREF5" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Member Words", |
| "sec_num": null |
| }, |
| { |
| "text": "We represent the context of a category as a vector <w 1 , w 2 , ..., w n >, where n is the total number of context words, and w i is the weight of the ith context word. To arrive at this representation, we first record the number of times each context word occurs within a specified window of each member word of a category in the corpus as a vector <f 1 , f 2 , ..., f n >, where f i is the number of times the ith context word co-occurs with a member word of the category. We then compute the weight of a context word w in context c, W(w, c), using mutual information and t-test, which were reported by Weeds and Weir (2005) to perform the best on a pseudo-disambiguation task. These weight functions are computed as in (10) and 11, where N denotes the size of the corpus. ", |
| "cite_spans": [ |
| { |
| "start": 605, |
| "end": 626, |
| "text": "Weeds and Weir (2005)", |
| "ref_id": "BIBREF9" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Context Representation", |
| "sec_num": null |
| }, |
| { |
| "text": "We compute the similarity between the context vectors of the unknown word and its candidate categories using cosine. The cosine of two ndimensional vectors x r and y r , cos( x r , y r ), is computed as in (12), where x i and y i denote the weight of the ith context word in x r and y r .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Contextual Similarity Measurement", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "\u2211 \u2211", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Contextual Similarity Measurement", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "\u2211 = = = = n 1 i 2 n 1 i 2 n 1 i ) y , x cos( i i i i y x y x r r", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Contextual Similarity Measurement", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "The models are developed and tested using the Contemporary Chinese Corpus from Peking University (Yu et al. 2002) We classify words into the third-level categories in the extended Cilin, which are equivalent to the small categories in the original Cilin. The development and test sets consist of 3,000 words each, which are randomly selected from the subset of words in Cilin that are two to four characters long, that have occurred in the Contemporary Chinese Corpus, and that are tagged as nouns, verbs, or adjectives in the corpus. The words in the development and test sets are also controlled for frequency, with 1/3 of them occurring 1-3 times, 3-6 times, and 7 or more times in the corpus respectively.", |
| "cite_spans": [ |
| { |
| "start": 97, |
| "end": 113, |
| "text": "(Yu et al. 2002)", |
| "ref_id": "BIBREF10" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experiment Setup", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "As Chen (2004) noted, excluding all the words in the development and test data in the testing stage worsens the data-sparseness problem for knowledge-based models, as some categories have few member words, and some characters appear in few words in some categories. To alleviate this problem, the remove-one method is used for testing the knowledge-based models. In other words, the models are re-trained for each test word using information about all the words in Cilin except the test word. The corpus-based model is trained once using the training data only, as the data-sparseness problem is alleviated by using generalized contexts of categories. Finally, if a word is polysemous, it is considered to have been correctly classified if the proposed category is one of its categories.", |
| "cite_spans": [ |
| { |
| "start": 3, |
| "end": 14, |
| "text": "Chen (2004)", |
| "ref_id": "BIBREF0" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experiment Setup", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "Tables 2 and 3 summarize the results of the baseline model in terms of the accuracy of its best guess and best five guesses respectively.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results of the Baseline Model", |
| "sec_num": "4.2" |
| }, |
| { |
| "text": "The columns labeled \"Non-filtered\" report results where all categories are considered for each unknown word, and the ones labeled \"POS-filtered\" report results where only the categories that agree with the POS category of the unknown word are considered. In the latter case, if the target word is a noun, only the small categories under major categories A-D are considered; otherwise, only those under major categories E-L are considered. The results show that using POS information about the unknown word to filter categories improves performance. Variant 2 performs better when only the best guess is considered, indicating that it is useful to model the effect of position on a character's contribution to word meaning in this case. However, it is not helpful to be sensitive to character position when the best five guesses are considered. ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results of the Baseline Model", |
| "sec_num": "4.2" |
| }, |
| { |
| "text": "In this model, only categories that agree with the POS category of the unknown word and that share at least one character with the unknown word are considered. These filtering steps significantly reduce the search space for this model. We discussed three parameters of the model in Section 2.2, including the statistical measure, the sensitivity to character position in computing character-category associations, and the weights of the associations between categories and characters in different positions. In addition, the computation of the character-category associations can be sensitive or insensitive to the POS categories of the words containing the characters. In the POS-sensitive way, associations are computed among nouns (words in categories A-D) and non-nouns (words in categories E-L) separately, whereas in the POS-insensitive way, they are computed using all the words.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results of the Character-Category Association Model", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "Tables 4 and 5 summarize the results of the character-category association model in terms of the accuracy of its best guess and best five guesses respectively. In all cases, the weights assigned to word-initial, word-middle, and wordfinal characters are 0.49, 0, and 0.51 respectively.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results of the Character-Category Association Model", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "In terms of the best guess, the model achieves a best accuracy of 58.2%, a 6.5% improvement over the baseline result. The results show that \u03c7 2 consistently performs better than mutual information, and computing position-sensitive character-category associations consistently improves performance. However, computing POSsensitive associations gives mixed results.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results of the Character-Category Association Model", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "In terms of the best five guesses, the model achieves a best accuracy of 83.8% on the test data, a 2.1% improvement over the best baseline result. Using \u03c7 2 again achieves better results. However, in this case, the best results are achieved when the character-category associations are insensitive to both character position and the POS categories of words. Table 6 summarizes the results of the rulebased model in terms of recall, precision and Fscore. The model returns multiple categories for some words, and it is considered to have correctly classified a word only when it returns a single, correct category for the word. Precision of the model is computed over all the cases where the model returns a single guess, and recall is computed over all cases. The model achieves an overall precision of 80.3% on the test data, much higher than the accuracy of the other two knowledge-based models. However, recall of the model is only 21.6%. The comparable results on the development and test sets indicate that the encoded rules are general. The model generally performs better on longer words than on shorter words.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 358, |
| "end": 365, |
| "text": "Table 6", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Results of the Character-Category Association Model", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "Given that the rule-based model achieves a higher precision but a lower recall than the character-category association model, the two models can be combined to improve the overall performance. In general, if the rule-based model returns one or more categories, these categories are first ranked among themselves by their associations with the unknown word. They are then followed by the other categories returned by the character-category association model. Tables 7 and 8 In terms of the best guess, the combined model achieves an accuracy of 61.6%, a 3.4% improvement over the best result of the character-category association model alone. This is achieved using \u03c7 2 with POS-sensitive and position-sensitive computation of character-category associations. In terms of the best five guesses, the model achieves an accuracy of 85.6%, a 1.8% improvement over the best result of the character-category association model alone.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 458, |
| "end": 473, |
| "text": "Tables 7 and 8", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Combining the Character-Category Association and Rule-Based Models", |
| "sec_num": "4.5" |
| }, |
| { |
| "text": "To facilitate comparison with previous studies, the results of the combined model in terms of its best guess in classifying unknown words into major and medium categories are summarized in Table 9 . As \u03c7 2 consistently outperforms mutual information, results are reported for \u03c7 2 only. With POS-sensitive and position-sensitive com-putation of character-category associations, the combined model achieves an accuracy of 83.0% and 69.9% for classifying unknown words into major and medium categories respectively. Table 9 : Results of the combined model for classifying unknown words into major and medium categories: best guess", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 189, |
| "end": 196, |
| "text": "Table 9", |
| "ref_id": null |
| }, |
| { |
| "start": 513, |
| "end": 520, |
| "text": "Table 9", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Combining the Character-Category Association and Rule-Based Models", |
| "sec_num": "4.5" |
| }, |
| { |
| "text": "The corpus-based model re-ranks the five highest ranked categories proposed by the combined knowledge-based model. Table 10 enumerates the parameters of the model and lists the labels used to denote the various settings in Table 10 : Parameter settings of the corpus-based model Table 11 summarizes the results of 16 runs of the model with different parameter settings. The best accuracy on the test data is 37.1%, achieved in run 5 with the following parameter settings: using unambiguous member words for building contexts of categories, using all words in the corpus for context representation, using a window size of 100, and using mutual information as the weight function. As the combined knowledge-based model gives an accuracy of 85.6% for its best five guesses, the expected accuracy of a naive model that randomly picks a candidate for each word as its best guess is 17.1%. Compared with this baseline, the corpus-based model achieves a 13.0% improvement, but it performs much worse than the knowledge-based models. Table 12 summarizes the accuracy of the top three runs of the model on words with different frequency in the corpus. Each of the three groups consists of 1,000 words that have occurred 1-2, 3-6, and 7 or more times in the corpus respectively. The model consistently performs better on words with higher frequency, suggesting that it may benefit from a larger corpus. similarity between their modifiers, using the concept of information load (IC) of the least common ancestor (LCA) of the modifiers' semantic categories. They reported an accuracy of 81% for classifying 200 unknown nouns. Given the small test set of their study, it is hard to directly compare their results with ours.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 115, |
| "end": 123, |
| "text": "Table 10", |
| "ref_id": null |
| }, |
| { |
| "start": 223, |
| "end": 231, |
| "text": "Table 10", |
| "ref_id": null |
| }, |
| { |
| "start": 279, |
| "end": 287, |
| "text": "Table 11", |
| "ref_id": "TABREF6" |
| }, |
| { |
| "start": 1026, |
| "end": 1034, |
| "text": "Table 12", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Results of the Corpus-Based Model", |
| "sec_num": "4.6" |
| }, |
| { |
| "text": "Tseng used a morphological analyzer in the same way, but she also derived the morphosyntactic relationship between the morphemes. She retrieved examples that share a morpheme with the target word in the same position and filtered those with a different morpho-syntactic relationship. Finally, she computed the similarity between two words as the similarity between their non-shared morphemes, using a similar concept of IC of the LCA of two categories. She classified unknown words into the 12 major categories only, and reported accuracies 65.8% on adjectives, 71.4% on nouns, and 52.8% on verbs. These results are not as good as the 83.0% overall accuracy our combined knowledge-based model achieved for classifying unknown words into major categories. Chen and Lin (2000) is the only study that used contextual information for the same task. To generate candidate categories for a word, they looked up its translations in a Chinese-English dictionary and the synsets of the translations in WordNet, and mapped the synsets to the categories in Cilin. They used a corpus-based model similar to ours to rank the candidates. They reported an accuracy of 34.4%, which is close to the 37.1% accuracy of our corpus-based model, but lower than the 61.6% accuracy of our combined knowledge-based model. In addition, they could only classify the unknown words listed in the Chinese-English dictionary.", |
| "cite_spans": [ |
| { |
| "start": 755, |
| "end": 774, |
| "text": "Chen and Lin (2000)", |
| "ref_id": "BIBREF3" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results of the Corpus-Based Model", |
| "sec_num": "4.6" |
| }, |
| { |
| "text": "We presented three knowledge-based models and a corpus-based model for classifying Chinese unknown words into fine-grained categories in the Chinese thesaurus Cilin, a task important for lexical acquisition and NLP applications that require semantic annotation. The knowledgebased models use information about the categories of the unknown words' component characters, while the corpus-based model uses contextual information. By combining the charactercategory association and rule-based models, we achieved an accuracy of 61.6%. The corpusbased model did not improve performance.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions", |
| "sec_num": "6" |
| }, |
| { |
| "text": "Several avenues can be taken for further research. First, additional resources, such as bilingual dictionaries, morphological analyzers, parallel corpora, and larger corpora with richer linguistic annotation may prove useful for improving both the knowledge-based and corpus-based models. Second, we only explored one way to combine the knowledge-based and corpus-based models. Future work may explore alternative ways to combine these models to make better use of contextual information.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions", |
| "sec_num": "6" |
| }, |
| { |
| "text": "A and B may each belong to more than one category.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "The few previous studies on semantic classification of Chinese unknown word have primarily adopted knowledge-based models. Chen (2004) proposed a model that retrieves the word with the greatest association with the target word. This model is computationally more expensive than our character-category association model, as it entails computing associations between every character-category, category-character, character-character, and word-word pair. He reported an accuracy of 61.6% on bisyllabic V-V compounds. However, he included all the test words in training the model. If we also include the test words in computing character-category associations, the computationally cheaper model achieves an overall accuracy of 75.6%, with an accuracy of 75.1% on verbs. adopted similar exemplar-based models. Chen and Chen used a morphological analyzer to identify the head of the target word and the semantic categories of its modifier. They then retrieved examples with the same head as the target word. Finally, they computed the similarity between two words as the", |
| "cite_spans": [ |
| { |
| "start": 123, |
| "end": 134, |
| "text": "Chen (2004)", |
| "ref_id": "BIBREF0" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "5" |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Character-sense association and compounding template similarity: Automatic semantic classification of Chinese compounds", |
| "authors": [ |
| { |
| "first": "C.-J", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proceedings of the 3rd SIGHAN Workshop on Chinese Language Processing", |
| "volume": "", |
| "issue": "", |
| "pages": "33--40", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "C.-J. Chen. 2004. Character-sense association and compounding template similarity: Automatic se- mantic classification of Chinese compounds. In Proceedings of the 3rd SIGHAN Workshop on Chi- nese Language Processing, pages 33-40.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Supersense tagging of unknown nouns in WordNet", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Ciaramita", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Johnson", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proceedings of EMNLP-2003", |
| "volume": "", |
| "issue": "", |
| "pages": "594--602", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "M. Ciaramita and M. Johnson. 2003. Supersense tag- ging of unknown nouns in WordNet. In Proceed- ings of EMNLP-2003, pages 594-602.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Automatic semantic classification for Chinese unknown compound nouns", |
| "authors": [ |
| { |
| "first": "K.-J", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| }, |
| { |
| "first": "C.-J", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Proceedings of COLING-2000", |
| "volume": "", |
| "issue": "", |
| "pages": "173--179", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "K.-J. Chen and C.-J. Chen. 2000. Automatic semantic classification for Chinese unknown compound nouns. In Proceedings of COLING-2000, pages 173-179.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Sense-tagging Chinese corpus", |
| "authors": [ |
| { |
| "first": "H.-H", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| }, |
| { |
| "first": "C.-C", |
| "middle": [], |
| "last": "Lin", |
| "suffix": "" |
| } |
| ], |
| "year": 2000, |
| "venue": "Proceedings of the 2nd Chinese Language Processing Workshop", |
| "volume": "", |
| "issue": "", |
| "pages": "7--14", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H.-H. Chen and C.-C. Lin. 2000. Sense-tagging Chi- nese corpus. In Proceedings of the 2nd Chinese Language Processing Workshop, pages 7-14.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Supersense tagging of unknown nouns using semantic similarity", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Curran", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of ACL-2006", |
| "volume": "", |
| "issue": "", |
| "pages": "26--33", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. Curran. 2005. Supersense tagging of unknown nouns using semantic similarity. In Proceedings of ACL-2006, pages 26-33.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Introduction on the special issue on word sense disambiguation: The state of the art", |
| "authors": [ |
| { |
| "first": "N", |
| "middle": [], |
| "last": "Ide", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "V\u00e9ronis", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Computational Linguistics", |
| "volume": "24", |
| "issue": "1", |
| "pages": "2--40", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "N. Ide and J. V\u00e9ronis. 1998. Introduction on the spe- cial issue on word sense disambiguation: The state of the art. Computational Linguistics 24(1):2-40.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Noun-phrase cooccurrence statistics for semi-automatic semantic lexicon construction", |
| "authors": [ |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Roark", |
| "suffix": "" |
| }, |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Charniak", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Proceedings of COL-ING/ACL-1998", |
| "volume": "", |
| "issue": "", |
| "pages": "1110--1116", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "B. Roark and E. Charniak. 1998. Noun-phrase co- occurrence statistics for semi-automatic semantic lexicon construction. In Proceedings of COL- ING/ACL-1998, pages 1110-1116.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Semantic classification of Chinese unknown words", |
| "authors": [ |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Tseng", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proceedings of ACL-2003 Student Research Workshop", |
| "volume": "", |
| "issue": "", |
| "pages": "72--79", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "H. Tseng. 2003. Semantic classification of Chinese unknown words. In Proceedings of ACL-2003 Stu- dent Research Workshop, pages 72-79.", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Co-occurrence retrieval: A flexible framework for lexical distributional similarity", |
| "authors": [ |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Weeds", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [], |
| "last": "Weir", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Computational Linguistics", |
| "volume": "31", |
| "issue": "4", |
| "pages": "439--475", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "J. Weeds and D. Weir. 2005. Co-occurrence retrieval: A flexible framework for lexical distributional similarity. Computational Linguistics 31(4):439- 475.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "The basic processing of Contemporary Chinese Corpus at Peking University", |
| "authors": [ |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Duan", |
| "suffix": "" |
| }, |
| { |
| "first": "X", |
| "middle": [], |
| "last": "Zhu", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Sun", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Journal of Chinese Information Processing", |
| "volume": "16", |
| "issue": "5", |
| "pages": "49--64", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "S. Yu, H. Duan, X. Zhu, and B. Sun. 2002. The basic processing of Contemporary Chinese Corpus at Peking University. Journal of Chinese Information Processing 16(5):49-64.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "TABREF0": { |
| "num": null, |
| "html": null, |
| "content": "<table/>", |
| "type_str": "table", |
| "text": "For a trisyllabic word ABC: a. If A equals \u5927 d\u00e0 'big', \u5c0f xi\u0103o 'little', or \u8001 l\u0103o 'old', propose the category of AB for ABC if C is the diminutive suffix \u513f er or the category of BC for ABC otherwise. b. If A and BC share a category c, propose c for ABC. c. If AB and C share a category c, propose c for ABC. d. If there is a word XYC such that XY and AB share a category, propose the category of XYC for ABC. e. If there is a word XBC such that X and A share a category, propose the category of XBC for ABC." |
| }, |
| "TABREF1": { |
| "num": null, |
| "html": null, |
| "content": "<table><tr><td>Length</td><td>Unambiguous</td><td>Polysemous</td><td>Total</td></tr><tr><td>1</td><td>2,674</td><td>2,068</td><td>4,742</td></tr><tr><td>2</td><td>39,057</td><td>5,403</td><td>44,460</td></tr><tr><td>3</td><td>15,112</td><td>752</td><td>15,864</td></tr><tr><td>4</td><td>9,397</td><td>942</td><td>10,338</td></tr><tr><td>\u22655</td><td>590</td><td>34</td><td>624</td></tr><tr><td>Total</td><td>66,830</td><td>9,199</td><td>76,029</td></tr><tr><td colspan=\"4\">Table 1: Word distribution in the extended Cilin</td></tr></table>", |
| "type_str": "table", |
| "text": "and the extended Cilin released by the Information Retrieval Lab at Harbin Institute of Technology. The corpus contains all the articles published in January, 1999 in People's Daily, a major newspaper in China. It contains over 1.12 million tokens and is word-segmented and POS-tagged. Table 1 summarizes the distribution of words in Cilin. Of the 76,029 words in Cilin, 35,151 are found in the Contemporary Chinese Corpus." |
| }, |
| "TABREF6": { |
| "num": null, |
| "html": null, |
| "content": "<table><tr><td colspan=\"3\">Parameter Label Setting</td><td>Label</td></tr><tr><td>Member</td><td>MW</td><td>All members words</td><td>all</td></tr><tr><td>words</td><td/><td>Unambiguous members</td><td>un</td></tr><tr><td>Context</td><td>CW</td><td>All words</td><td>all</td></tr><tr><td>words</td><td/><td>1000 most frequent</td><td>1000</td></tr><tr><td>Window</td><td>WS</td><td>100</td><td>100</td></tr><tr><td>size</td><td/><td>6</td><td>6</td></tr><tr><td>Weight</td><td>WF</td><td>Mutual information</td><td>mi</td></tr><tr><td>function</td><td/><td>t-test</td><td>t</td></tr></table>", |
| "type_str": "table", |
| "text": "" |
| } |
| } |
| } |
| } |
