{ "paper_id": "J98-4002", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T02:53:12.007425Z" }, "title": "Selective Sampling for Example-based Word Sense Disambiguation", "authors": [ { "first": "Atsushi", "middle": [], "last": "Fujii", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Hozumi", "middle": [], "last": "Tanaka", "suffix": "", "affiliation": {}, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This paper proposes an efficient example sampling method for example-based word sense disambiguation systems. To construct a database of practical size, a considerable overhead for manual sense disambiguation (overhead for supervision) is required. In addition, the time complexity of searching a large-sized database poses a considerable problem (overhead for search). To counter these problems, our method selectively samples a smaller-sized effective subset from a given example set for use in word sense disambiguation. Our method is characterized by the reliance on the notion of training utility: the degree to which each example is informative for future example sampling when used for the training of the system. The system progressively collects examples by selecting those with greatest utility. The paper reports the effectiveness of our method through experiments on about one thousand sentences. Compared to experiments with other example sampling methods, our method reduced both the overhead for supervision and the overhead for search, without the degeneration of the performance of the system.", "pdf_parse": { "paper_id": "J98-4002", "_pdf_hash": "", "abstract": [ { "text": "This paper proposes an efficient example sampling method for example-based word sense disambiguation systems. To construct a database of practical size, a considerable overhead for manual sense disambiguation (overhead for supervision) is required. In addition, the time complexity of searching a large-sized database poses a considerable problem (overhead for search). To counter these problems, our method selectively samples a smaller-sized effective subset from a given example set for use in word sense disambiguation. Our method is characterized by the reliance on the notion of training utility: the degree to which each example is informative for future example sampling when used for the training of the system. The system progressively collects examples by selecting those with greatest utility. The paper reports the effectiveness of our method through experiments on about one thousand sentences. Compared to experiments with other example sampling methods, our method reduced both the overhead for supervision and the overhead for search, without the degeneration of the performance of the system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Word sense disambiguation is a potentially crucial task in many NLP applications, such as machine translation (Brown, Della Pietra, and Della Pietra 1991) , parsing (Lytinen 1986; Nagao 1994 ) and text retrieval (Krovets and Croft 1992; Voorhees 1993) . Various corpus-based approaches to word sense disambiguation have been proposed (Bruce and Wiebe 1994; Charniak 1993; Dagan and Itai 1994; Fujii et al. 1996; Hearst 1991; Karov and Edelman 1996; Kurohashi and Nagao 1994; Li, Szpakowicz, and Matwin 1995; Ng and Lee 1996; Niwa and Nitta 1994; Sch~itze 1992; Uramoto 1994b; Yarowsky 1995) . The use of corpus-based approaches has grown with the use of machine-readable text, because unlike conventional rule-based approaches relying on hand-crafted selectional rules (some of which are reviewed, for example, by Hirst [1987] ), corpus-based approaches release us from the task of generalizing observed phenomena through a set of rules. Our verb sense disambiguation system is based on such an approach, that is, an example-based approach. A preliminary experiment showed that our system performs well when compared with systems based on other approaches, and motivated us to further explore the example-based approach (we elaborate on this experiment in Section 2.3). At the same time, we concede that other approaches for word sense disambiguation are worth further exploration, and while we focus on example-based approach in this paper, we do not wish to draw any premature conclusions regarding tlhe relative merits of different generalized approaches.", "cite_spans": [ { "start": 110, "end": 154, "text": "(Brown, Della Pietra, and Della Pietra 1991)", "ref_id": "BIBREF1" }, { "start": 165, "end": 179, "text": "(Lytinen 1986;", "ref_id": "BIBREF27" }, { "start": 180, "end": 190, "text": "Nagao 1994", "ref_id": "BIBREF32" }, { "start": 212, "end": 236, "text": "(Krovets and Croft 1992;", "ref_id": "BIBREF22" }, { "start": 237, "end": 251, "text": "Voorhees 1993)", "ref_id": "BIBREF55" }, { "start": 334, "end": 356, "text": "(Bruce and Wiebe 1994;", "ref_id": "BIBREF2" }, { "start": 357, "end": 371, "text": "Charniak 1993;", "ref_id": "BIBREF4" }, { "start": 372, "end": 392, "text": "Dagan and Itai 1994;", "ref_id": "BIBREF6" }, { "start": 393, "end": 411, "text": "Fujii et al. 1996;", "ref_id": "BIBREF9" }, { "start": 412, "end": 424, "text": "Hearst 1991;", "ref_id": "BIBREF14" }, { "start": 425, "end": 448, "text": "Karov and Edelman 1996;", "ref_id": "BIBREF21" }, { "start": 449, "end": 474, "text": "Kurohashi and Nagao 1994;", "ref_id": "BIBREF23" }, { "start": 475, "end": 507, "text": "Li, Szpakowicz, and Matwin 1995;", "ref_id": "BIBREF26" }, { "start": 508, "end": 524, "text": "Ng and Lee 1996;", "ref_id": "BIBREF37" }, { "start": 525, "end": 545, "text": "Niwa and Nitta 1994;", "ref_id": "BIBREF38" }, { "start": 546, "end": 560, "text": "Sch~itze 1992;", "ref_id": null }, { "start": 561, "end": 575, "text": "Uramoto 1994b;", "ref_id": "BIBREF52" }, { "start": 576, "end": 590, "text": "Yarowsky 1995)", "ref_id": "BIBREF57" }, { "start": 814, "end": 826, "text": "Hirst [1987]", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "As with most example-based systems (Fujii et al. 1996 ; Kurohashi and Nagao 1994; Li, Szpakowicz, and Matwin 1995; Uramoto 1994b) , our system uses an example database (database, hereafter) that contains example sentences associated with each verb sense. Given an input sentence containing a polysemous verb, the system chooses the most plausible verb sense from predefined candidates. In this process, the system computes a scored similarity between the input and examples in the database, and choses the verb sense associated with the example that maximizes the score. To realize this, we have to manually disambiguate polysemous verbs appearing in examples, prior to their use by the system. We shall call these examples supervised examples. A preliminary experiment on eleven polysemous Japanese verbs showed that (a) the more supervised examples we provided to the system, the better it performed, and (b) in order to achieve a reasonable result (say over 80% accuracy), the system needed a hundred-order supervised example set for each verb. Therefore, in order to build an operational system, the following problems have to be taken into account1:", "cite_spans": [ { "start": 35, "end": 53, "text": "(Fujii et al. 1996", "ref_id": "BIBREF9" }, { "start": 56, "end": 81, "text": "Kurohashi and Nagao 1994;", "ref_id": "BIBREF23" }, { "start": 82, "end": 114, "text": "Li, Szpakowicz, and Matwin 1995;", "ref_id": "BIBREF26" }, { "start": 115, "end": 129, "text": "Uramoto 1994b)", "ref_id": "BIBREF52" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "given human resource limitations, it is not reasonable to supervise every example in large corpora (\"overhead for supervision\"),", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "given the fact that example-based systems, including our system, search the database for the examples most similar to the input, the computational cost becomes prohibitive if one works with a very large database size (\"overhead for search\").", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "These problems suggest a different approach, namely to select a small number of optimally informative examples from given corpora. Hereafter we will call these examples samples.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Our example sampling method, based on the utility maximization principle, decides on the preference for including a given example in the database. This decision procedure is usually called selective sampling (Cohn, Atlas, and Ladner 1994) . The overall control flow of selective sampling systems can be depicted as in Figure 1 , where \"system\" refers to our verb sense disambiguation system, and \"examples\" refers to an unsupervised example set. The sampling process basically cycles between the word sense disambiguation (WSD) and training phases. During the WSD phase, the system generates an interpretation for each polysemous verb contained in the input example (\"WSD outputs\" of Figure 1 ). This phase is equivalent to normal word sense disambiguation execution. During the training phase, the system selects samples for training from the previously produced outputs. During this phase, a human expert supervises samples, that is, provides the correct interpretation for the verbs appearing in the samples. Thereafter, samples are simply incorporated into the database without any computational overhead (as would be associated with globally reestimating parameters in statistics-based systems), meaning that the system can be trained on the remaining examples (the \"residue\") for the next iteration. Iterating between these two sampling ~WSD~sD out ut~~ :(~~", "cite_spans": [ { "start": 208, "end": 238, "text": "(Cohn, Atlas, and Ladner 1994)", "ref_id": "BIBREF5" } ], "ref_spans": [ { "start": 318, "end": 326, "text": "Figure 1", "ref_id": null }, { "start": 684, "end": 692, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Flow of control of the example sampling system. phases, the system progressively enhances the database. Note that the selective sampiing procedure gives us an optimally informative database of a given size irrespective of the stage at which processing is terminated. Several researchers have proposed this type of approach for NLP applications. Engelson and Dagan (1996) proposed a committee-based sampling method, which is currently applied to HMM training for part-of-speech tagging. This method sets several models (the committee) taken from a given supervised data set, and selects samples based on the degree of disagreement among the committee members as to the output. This method is implemented for statistics-based models. How to formalize and map the concept of selective sampling into example-based approaches has yet to be explored. Lewis and Gale (1994) proposed an uncertainty sampling method for statisticsbased text classification. In this method, the system always samples outputs with an uncertain level of correctness. In an example-based approach, we should also take into account the training effect a given example has on other unsupervised examples. This is introduced as training utility in our method. We devote Section 4 to further comparison of our approach and other related works.", "cite_spans": [ { "start": 345, "end": 370, "text": "Engelson and Dagan (1996)", "ref_id": "BIBREF7" }, { "start": 845, "end": 866, "text": "Lewis and Gale (1994)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Figure 1", "sec_num": null }, { "text": "With respect to the problem of overhead for search, possible solutions would include the generalization of similar examples (Kaji, Kida, and Morimoto 1992; Nomiyama 1993) or the reconstruction of the database using a small portion of useful instances selected from a given supervised example set (Aha, Kibler, and Albert 1991; Smyth and Keane 1995) . However, such approaches imply a significant overhead for supervision of each example prior to the system's execution. This shortcoming is precisely what our approach aims to avoid: we aim to reduce the overhead for supervision as well as the overhead for search.", "cite_spans": [ { "start": 124, "end": 155, "text": "(Kaji, Kida, and Morimoto 1992;", "ref_id": "BIBREF19" }, { "start": 156, "end": 170, "text": "Nomiyama 1993)", "ref_id": "BIBREF39" }, { "start": 296, "end": 326, "text": "(Aha, Kibler, and Albert 1991;", "ref_id": "BIBREF0" }, { "start": 327, "end": 348, "text": "Smyth and Keane 1995)", "ref_id": "BIBREF48" } ], "ref_spans": [], "eq_spans": [], "section": "Figure 1", "sec_num": null }, { "text": "Section 2 describes the basis of our verb sense disambiguation system and preliminary experiment, in which we compared our method with other disambiguation methods. Section 3 then elaborates on our example sampling method. Section 4 reports on the results of our experiments through comparison with other proposed selective sampling methods, and discusses theoretical differences between those methods.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 1", "sec_num": null }, { "text": "Our verb sense disambiguation system is based on the method proposed by Kurohashi and Nagao (1994) and later enhanced by Fujii et al. (1996) . The system uses a database containing examples of collocations for each verb sense and its associated case frame(s). ", "cite_spans": [ { "start": 72, "end": 98, "text": "Kurohashi and Nagao (1994)", "ref_id": "BIBREF23" }, { "start": 121, "end": 140, "text": "Fujii et al. (1996)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "The Basic Idea", "sec_num": "2.1" }, { "text": "A fragment of the database, and the entry associated with the Japanese verb toru. Figure 2 shows a fragment of the entry associated with the Japanese verb toru. The verb toru has multiple senses, a sample of which are 'to take/steal,' 'to attain,' 'to subscribe,' and 'to reserve.' The database specifies the case frame(s) associated with each verb sense. In Japanese, a complement of a verb consists of a noun phrase (case filler) and its case marker suffix, for example ga (nominative) or wo (accusative). The database lists several case filler examples for each case. The task of the system is to \"interpret\" the verbs occurring in the input text, i.e., to choose one sense from among a set of candidates. 2 All verb senses we use are defined in IPAL (information-technology Promotion Agency, 1987), a machine-readable dictionary. IPAL also contains example case fillers as shown in Figure 2 . Given an input, which is currently limited to a simple sentence, the system identifies the verb sense on the basis of the scored similarity between the input and the examples given for each verb sense. Let us take the sentence below as an example input: hisho ga shindaisha wo toru. (secretary-NOM) (sleeping car-ACC) (?) In this example, one may consider hisho ('secretary') and shindaisha ('sleeping car') to be semantically similar to joshu ('assistant') and hikouki ('airplane') respectively, and since both collocate with the 'to reserve' sense of toru, one could infer that toru should be interpreted as 'to reserve.' This resolution originates from the analogy principle (Nagao 1984) , and can be called nearest neighbor resolution because the verb in the input is disambiguated by superimposing the sense of the verb appearing in the example of highest similarity. 3 The similarity between an input and an example is estimated based on the similarity between case \u2022lers marked with the same case.", "cite_spans": [ { "start": 1215, "end": 1218, "text": "(?)", "ref_id": null }, { "start": 1575, "end": 1587, "text": "(Nagao 1984)", "ref_id": "BIBREF33" } ], "ref_spans": [ { "start": 82, "end": 90, "text": "Figure 2", "ref_id": null }, { "start": 886, "end": 894, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Figure 2", "sec_num": null }, { "text": "Furthermore, since the restrictions imposed by the case fillers in choosing the verb sense are not equally selective, Fujii et al. (1996) proposed a weighted case contribution to disambiguation (CCD) of the verb senses. This CCD factor is taken into account \u00a9 nominative accusative", "cite_spans": [ { "start": 118, "end": 137, "text": "Fujii et al. (1996)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Figure 2", "sec_num": null }, { "text": "The semantic ranges of the nominative and accusative for the verb toru.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 3", "sec_num": null }, { "text": "database nq-mc~nc nc3-rnllc3 v (?) Cs~,c~ G,c2 &,;3 -- v (s~)l -- ~s3,c2 ~s3,C3 - -- V (S3) J", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 3", "sec_num": null }, { "text": "An input and the database.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 4", "sec_num": null }, { "text": "when computing the score for each sense of the verb in question. Consider again the case of toru in Figure 2 . Since the semantic range of nouns collocating with the verb in the nominative does not seem to have a strong delinearization in a semantic sense (in Figure 2 , the nominative of each verb sense displays the same general concept, i.e., HUMAN), it would be difficult, or even risky, to properly interpret the verb sense based on similarity in the nominative. In contrast, since the semantic ranges are disparate in tile accusative, it would be feasible to rely more strongly on similarity here. This argument can be illustrated as in Figure 3 , in which the symbols el and e2 denote example case fillers of different case frames, and an input sentence includes two case fillers denoted by x and y. The figure shows the distribution of example case fillers for the respective case frames, denoted in a semantic space. The semantic similarity between two given case fillers is represented by the physical distance between the two symbols. In the nominative, since x happens to be much closer to an e2 than any el, x may be estimated to belong to the range of e2's, although x actually belongs to both sets of el's and e2's. In the accusative, however, y would be properly estimated to belong to the set of el's due to the disjunction of the two accusative case filler sets, even though examples do not fully cover each of the ranges of el's and e2's. Note that this difference would be critical if example data were sparse. We will explain the method used to compute CCD in Section 2.2.", "cite_spans": [], "ref_spans": [ { "start": 100, "end": 108, "text": "Figure 2", "ref_id": null }, { "start": 260, "end": 268, "text": "Figure 2", "ref_id": null }, { "start": 643, "end": 651, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Figure 4", "sec_num": null }, { "text": "To illustrate the overall algorithm, we will consider an abstract specification of both an input and the database (Figure 4 ). Let the input be {no1 -reel, nc2 -mc2, nc3 -rnc3, v}, where nci denotes the case filler for the case ci, and mci denotes the case marker for ci, and assume that the interpretation candidates for v are derived from the database as sl, s2 and s3. The database also contains a set Gi,cj of case filler examples for each case cj of each sense si (\"--\" indicates that the corresponding case is not allowed).", "cite_spans": [], "ref_spans": [ { "start": 114, "end": 123, "text": "(Figure 4", "ref_id": null } ], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "During the verb sense disambiguation process, the system first discards those candidates whose case frame does not fit the input. In the case of Figure 4 , s3 is discarded because the case frame of v (s3) does not subcategorize for the case cl. The relation between the length of the path between two nouns nl and n2 in the Bunruigoihyo thesaurus (len(nl, n2)), and their relative similarity (sire(n1, n2)). fen(n1, n2) 0 2 4 6 8 10 12 sire(n1, n2) 11 10 9 8 7 5 0", "cite_spans": [], "ref_spans": [ { "start": 145, "end": 153, "text": "Figure 4", "ref_id": null } ], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "In the next step the system computes the score of the remaining candidates and chooses as the most plausible interpretation the one with the highest score. The score of an interpretation is computed by considering the weighted average of the similarity degrees of the input case fillers with respect to each of the example case \u2022lers (in the corresponding case) listed in the database for the sense under evaluation. Formally, this is expressed by Equation 1, where Score(s) is the score of sense s of the input verb, and SIM (nc, G,c) is the maximum similarity degree between the input case filler nc and the corresponding case fillers in the database example set ~s,c (calculated through Equation (2)). CCD(c) is the weight factor of case c, which we will explain later in this section.", "cite_spans": [ { "start": 526, "end": 535, "text": "(nc, G,c)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "Score(s) = ~c SIM(n\u00a2, &,c)' CCD(c) (1) CCD(c) SIM(nc, &,c) = max sim(nc, e) (2) eC G,c", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "With regard to the computation of the similarity between two different case fillers (sim(n~, e) in Equation 1), we experimentally used two alternative approaches. The first approach uses semantic resources, that is, hand-crafted thesauri (such as the Roger's thesaurus [Chapman 1984] or WordNet [Miller et al. 1993] in the case of English, and Bunruigoihyo [National Language Research Institute 1964] or EDR [Japan Electronic Dictionary Research Institute 1995] in the case of Japanese), based on the intuitively feasible assumption that words located near each other within the structure of a thesaurus have similar meaning. Therefore, the similarity between two given words is :represented by the length of the path between them in the thesaurus structure (Fujii et al. 1996; Kurohashi and Nagao 1994; Li, Szpakowicz, and Matwin 1995; Uramoto 1994b) . 4 We used the similarity function empirically identified by Kurohashi and Nagao in which the relation between the length of the path in the Bunruigoihyo thesaurus and the similarity between words is defined as shown in Table 1 . In this thesaurus, each entry is assigned a seven-digit class code. In other words, this thesaurus can be considered as a tree, seven levels in depth, with each leaf as a set of words. Figure 5 shows a fragment of the Bunruigoihyo thesaurus including some of the nouns in both Figure 2 and the input sentence above.", "cite_spans": [ { "start": 269, "end": 283, "text": "[Chapman 1984]", "ref_id": "BIBREF3" }, { "start": 295, "end": 315, "text": "[Miller et al. 1993]", "ref_id": "BIBREF30" }, { "start": 357, "end": 400, "text": "[National Language Research Institute 1964]", "ref_id": null }, { "start": 408, "end": 461, "text": "[Japan Electronic Dictionary Research Institute 1995]", "ref_id": null }, { "start": 758, "end": 777, "text": "(Fujii et al. 1996;", "ref_id": "BIBREF9" }, { "start": 778, "end": 803, "text": "Kurohashi and Nagao 1994;", "ref_id": "BIBREF23" }, { "start": 804, "end": 836, "text": "Li, Szpakowicz, and Matwin 1995;", "ref_id": "BIBREF26" }, { "start": 837, "end": 851, "text": "Uramoto 1994b)", "ref_id": "BIBREF52" }, { "start": 914, "end": 933, "text": "Kurohashi and Nagao", "ref_id": null } ], "ref_spans": [ { "start": 1073, "end": 1080, "text": "Table 1", "ref_id": "TABREF1" }, { "start": 1268, "end": 1276, "text": "Figure 5", "ref_id": null }, { "start": 1360, "end": 1368, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "The second approach is based on statistical modeling. We adopted one typical implementation called the \"vector space model\" (VSM) (Frakes and Baeza-Yates 1992; Leacock, Towell, and Voorhees 1993; Salton and McGill 1983; Sch/itze 1992) , which has a long history of application in information retrieval (IR) and text categorization (TC) tasks. In the case of IR/TC, VSM is used to compute the similarity between documents, which is represented by a vector comprising statistical factors of content words in a document. Similarly, in our case, each noun is represented by a vector comprising A fragment of the Bunruigoihyo thesaurus. statistical factors, although statistical factors are calculated in terms of the predicate argument structure in which each noun appears. Predicate argument structures, which consist of complements (case filler nouns and case markers) and verbs, have also been used in the task of noun classification (Hindle 1990 ). This can be expressed by Equation 3, where ff is the vector for the noun in question, and items ti represent the statistics for predicate argument structures including n. ff = (h, t2,..., ti .... ) ( 3)In regard to ti, we used the notion of TF. IDF (Salton and McGill 1983) . TF (term frequency) gives each context (a case marker/verb pair) importance proportional to the number of times it occurs with a given noun. The rationale behind IDF (inverse document frequency) is that contexts that rarely occur over collections of nouns are valuable, and that therefore the IDF of a context is inversely proportional to the number of noun types that appear in that context. This notion is expressed by Equation 4, wheref ((n, c, v) ) is the frequency of the tuple (n, c, v), nf ((c, v) ) is the number of noun types which collocate with verb v in the case c, and N is the number of noun types within the overall co-occurrence data.", "cite_spans": [ { "start": 130, "end": 159, "text": "(Frakes and Baeza-Yates 1992;", "ref_id": "BIBREF8" }, { "start": 160, "end": 195, "text": "Leacock, Towell, and Voorhees 1993;", "ref_id": "BIBREF24" }, { "start": 196, "end": 219, "text": "Salton and McGill 1983;", "ref_id": "BIBREF45" }, { "start": 220, "end": 234, "text": "Sch/itze 1992)", "ref_id": null }, { "start": 933, "end": 945, "text": "(Hindle 1990", "ref_id": "BIBREF15" }, { "start": 1125, "end": 1146, "text": "(h, t2,..., ti .... )", "ref_id": null }, { "start": 1198, "end": 1222, "text": "(Salton and McGill 1983)", "ref_id": "BIBREF45" }, { "start": 1665, "end": 1675, "text": "((n, c, v)", "ref_id": null }, { "start": 1722, "end": 1729, "text": "((c, v)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "N ti =d((n, c, v)). log nf((c, v))", "eq_num": "(4)" } ], "section": "Methodology", "sec_num": "2.2" }, { "text": "We compute the similarity between nouns nt and n2 by the cosine of the angle between the two vectors t~ and n2. This is realized by Equation (5).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "/I/1 \" n2 sire(n1, n2) -~ ]]~21", "eq_num": "(5)" } ], "section": "Methodology", "sec_num": "2.2" }, { "text": "We extracted co-occurrence data from the RWC text base RWC-DB-TEXT-95-1 (Real World Computing Partnership 1995) . This text base consists of four years worth of Mainichi Shimbun newspaper articles (Mainichi Shimbun 1991 -1994 , which have been automatically annotated with morphological tags. The total morpheme content is about one hundred million. Since full parsing is usually expensive, a simple heuristic rule was used to obtain collocations of nouns, case markers, and verbs in the form of tuples (n, c, v) . This rule systematically associates each sequence of noun and case marker to the verb of highest proximity, and produced 419,132 tuples. This co-occurrence data was used in the preliminary experiment described in Section 2.3. s", "cite_spans": [ { "start": 72, "end": 111, "text": "(Real World Computing Partnership 1995)", "ref_id": null }, { "start": 197, "end": 219, "text": "(Mainichi Shimbun 1991", "ref_id": null }, { "start": 220, "end": 225, "text": "-1994", "ref_id": "BIBREF32" }, { "start": 503, "end": 512, "text": "(n, c, v)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "In Equation 1, CCD(c) expresses the weight factor of the contribution of case c to (current) verb sense disambiguation. Intuitively, preference should be given to cases displaying case fillers that are classified in semantic categories of greater disjunction.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "Thus, c's contribution to the sense disambiguation of a given verb, CCD(c), is likely to be higher if the example case filler sets {gsi,c I i = 1,..., n} share fewer elements, as in Equation (6 ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "Here, o~ is a constant for pararneterizing the extent to which CCD influences verb sense disambiguation. The larger oe is, the stronger is CCD's influence on the system output. To avoid data sparseness, we smooth each element (noun example) in gsi,c. In practice, this involves generalizing each example noun into a five-digit class based on the Bunruigoihyo thesaurus, as has been commonly used for smoothing.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "2.2" }, { "text": "We estimated the performance of our verb sense disambiguation method through an experiment, in which we compared the following five methods:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminary Experimentation", "sec_num": "2.3" }, { "text": "\u2022 lower bound (LB), in which the system systematically chooses the most frequently appearing verb sense in the database (Gale, Church, and Yarowsky 1992 ),", "cite_spans": [ { "start": 120, "end": 152, "text": "(Gale, Church, and Yarowsky 1992", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Preliminary Experimentation", "sec_num": "2.3" }, { "text": "\u2022 rule-based method (RB), in which the system uses a thesaurus to (automatically) identify appropriate semantic classes as selectional restrictions for each verb complement,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminary Experimentation", "sec_num": "2.3" }, { "text": "\u2022 Naive-Bayes method (NB), in which the system interprets a given verb based on the probability that it takes each verb sense,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminary Experimentation", "sec_num": "2.3" }, { "text": "\u2022 example-based method using the vector space model (VSM), in which the system uses the above mentioned co-occurrence data extracted from the RWC text base,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Preliminary Experimentation", "sec_num": "2.3" }, { "text": "\u2022 example-based method using the Bunruigoihyo thesaurus (BGH), in which the system uses Table 1 for the similarity computation.", "cite_spans": [], "ref_spans": [ { "start": 88, "end": 95, "text": "Table 1", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Preliminary Experimentation", "sec_num": "2.3" }, { "text": "In the rule-based method, selectional restrictions are represented by thesaurus classes, and allow only those nouns dominated by the given class in the thesaurus structure as verb complements. In order to identify appropriate thesaurus classes, we used the association measure proposed by Resnik (1993) , which computes the information-theoretic association degree between case fillers and thesaurus classes, for each verb sense (Equation (7)). 6", "cite_spans": [ { "start": 289, "end": 302, "text": "Resnik (1993)", "ref_id": "BIBREF43" } ], "ref_spans": [], "eq_spans": [], "section": "Preliminary Experimentation", "sec_num": "2.3" }, { "text": "A (s,c,r) ", "cite_spans": [ { "start": 2, "end": 9, "text": "(s,c,r)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "P(rls, c)", "sec_num": null }, { "text": "= P(rls, c ) \u2022 log p(rlc) (7)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "P(rls, c)", "sec_num": null }, { "text": "Here, A(s, c, r) is the association degree between verb sense s and class r (selectional restriction candidate) with respect to case c. P(rls, c) is the conditional probability that a case filler example associated with case c of sense s is dominated by class r in the thesaurus. P(rlc ) is the conditional probability that a case filler example for case c (disregarding verb sense) is dominated by class r. Each probability is estimated based on training data. We used the semantic classes defined in the Bunruigoihyo thesaurus.", "cite_spans": [ { "start": 6, "end": 16, "text": "A(s, c, r)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "P(rls, c)", "sec_num": null }, { "text": "In practice, every r whose association degree is above a certain threshold is chosen as a selectional restriction (Resnik 1993; Ribas 1995) . By decreasing the value of the threshold, system coverage can be broadened, but this opens the way for irrelevant (noisy) selectional rules.", "cite_spans": [ { "start": 114, "end": 127, "text": "(Resnik 1993;", "ref_id": "BIBREF43" }, { "start": 128, "end": 139, "text": "Ribas 1995)", "ref_id": "BIBREF44" } ], "ref_spans": [], "eq_spans": [], "section": "P(rls, c)", "sec_num": null }, { "text": "The Naive-Bayes method assumes that each case filler included in a given input is conditionally independent of other case fillers: the system approximates the probability that an input x takes a verb sense s (P(slx)), simply by computing the product of the probability that each verb sense s takes nc as a case filler for case c. The verb sense with maximal probability is then selected as the interpretation (Equation 8 Here, P(ncls) is the probability that a case filler associated with sense s for case c in the training data is nc. We estimated P(s) based on the distribution of the verb senses in the training data. In practice, data sparseness leads to not all case fillers nc appearing in the database, so we generalize each nc into a semantic class defined in the Bunruigoihyo thesaurus.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "P(rls, c)", "sec_num": null }, { "text": "All methods except the lower bound method involve a parametric constant: the threshold value for the association degree (RB), a generalization level for case filler nouns (NB), and a in Equation (6) (VSM and BGH). For these parameters, we conducted several trials prior to the actual comparative experiment, to determine the optimal parameter values over a range of data sets. For our method, we set a extremely large, which is equivalent to relying almost solely on the SIM of the case with the greatest CCD. However, note that when the SIM of the case with the greatest CCD is equal for multiple verb senses, the system computes the SIM of the case with the second highest CCD. This process is repeated until only one verb sense remains. When more than one verb sense is selected for any given method (or none of them remains, for the rule-based method), the system simply selects the verb sense that appears most frequently in the database, s", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "P(rls, c)", "sec_num": null }, { "text": "In the experiment, we conducted sixfold cross-validation, that is, we divided the training/test data into six equal parts, and conducted six trials in which a different 7 A number of experimental results have shown the effectiveness of the Naive-Bayes method for word sense disambiguation (Gale, Church, and Yarowsky 1993; Leacock, Towell, and Voorhees 1993; Mooney 1996; Ng 1997; Pedersen, Bruce, and Wiebe 1997) . 8 One may argue that this goes against the basis of the rule-based method, in that, given a proper threshold value for the association degree, the system could improve on accuracy (potentially sacrificing coverage), and that the trade-off between coverage and accuracy is therefore a more appropriate evaluation criterion. However, our trials on the rule-based method with different threshold values did not show significant correlation between the improvement of accuracy and the degeneration of coverage.", "cite_spans": [ { "start": 289, "end": 322, "text": "(Gale, Church, and Yarowsky 1993;", "ref_id": "BIBREF12" }, { "start": 323, "end": 358, "text": "Leacock, Towell, and Voorhees 1993;", "ref_id": "BIBREF24" }, { "start": 359, "end": 371, "text": "Mooney 1996;", "ref_id": "BIBREF31" }, { "start": 372, "end": 380, "text": "Ng 1997;", "ref_id": "BIBREF36" }, { "start": 381, "end": 413, "text": "Pedersen, Bruce, and Wiebe 1997)", "ref_id": "BIBREF40" } ], "ref_spans": [], "eq_spans": [], "section": "P(rls, c)", "sec_num": null }, { "text": "The verbs contained in the corpus used, and the accuracy of the different verb sense disambiguation methods (LB: lower bound, RB: rule-based method, NB: Naive-Bayes method, VSM: vector space model, BGH: the Bunruigoihyo thesaurus). part was used as test data each time, and the rest as training data (the database). 9 We evaluated the performance of each method according to its accuracy, that is, the ratio of the number of correct outputs compared to the total number of inputs. The training/test data used in the experiment contained about one thousand simple Japanese sentences collected from news articles. Each sentence in the training/test data contained one or more complement(s) followed by one of the eleven verbs described in Table 2 . In Table 2 , the column \"English Gloss\" describes typical English translations of the Japanese verbs. The column \"# of Sentences\" denotes the number of sentences in the corpus, and \"# of Senses\" denotes the number of verb senses contained in IPAL.", "cite_spans": [], "ref_spans": [ { "start": 737, "end": 744, "text": "Table 2", "ref_id": null }, { "start": 750, "end": 757, "text": "Table 2", "ref_id": null } ], "eq_spans": [], "section": "Table 2", "sec_num": null }, { "text": "The column \"accuracy\" shows the accuracy of each method. Looking at Table 2 , one can see that our example-based method performed better than the other methods (irrespective of the similarity computation), although the Naive-Bayes method is relatively comparable in performance. Surprisingly, despite the relatively ad hoc similarity definition used (see Table 1 ), the Bunruigoihyo thesaurus led to a greater accuracy gain than the vector space model. In order to estimate the upper bound (limitation) of the disambiguation task, that is, to what extent a human expert makes errors in disambiguation (Gale, Church, and Yarowsky 1992) , we analyzed incorrect outputs and found that roughly 30% of the system errors using the Bunruigoihyo thesaurus fell into this category. It should be noted that while the vector space model requires computational cost (time/memory) of an order proportional to the size of the vector, determination of paths in the Bunruigoihyo thesaurus comprises a trivial cost.", "cite_spans": [ { "start": 601, "end": 634, "text": "(Gale, Church, and Yarowsky 1992)", "ref_id": "BIBREF11" } ], "ref_spans": [ { "start": 68, "end": 75, "text": "Table 2", "ref_id": null }, { "start": 355, "end": 362, "text": "Table 1", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Table 2", "sec_num": null }, { "text": "We also investigated errors made by the rule-based method to find a rational explanation for its inferiority. We found that the association measure in Equation (7) tends to give a greater value to less frequently appearing verb senses and lower level (more specified) classes, and therefore chosen rules are generally overspecified. 1\u00b0 Consequently, frequently appearing verb senses are likely to be rejected. On the other hand, when attempting to enhance the rule set by setting a smaller threshold value for the association score, overgeneralization can be a problem. We also note that one of the theoretical differences between the rule-based and example-based methods is that the former statically generalizes examples (prior to system usage), while the latter does so dynamically. Static generalization would appear to be relatively risky for sparse training data.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Table 2", "sec_num": null }, { "text": "Although comparison of different approaches to word sense disambiguation should be further investigated, this experimental result gives us good motivation to explore example-based verb sense disambiguation approaches, i.e., to introduce the notion of selective sampling into them.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Table 2", "sec_num": null }, { "text": "Let us discuss how further enhancements to our example-based verb sense disambiguation system could be made. First, since inputs are simple sentences, information for word sense disambiguation is inadequate in some cases. External information such as the discourse or domain dependency of each word sense (Guthrie et al. 1991; Nasukawa 1993; Yarowsky 1995 ) is expected to lead to system improvement. Second, some idiomatic expressions represent highly restricted collocations, and overgeneralizing them semantically through the use of a thesaurus can cause further errors. Possible solutions would include one proposed by Uramoto, in which idiomatic expressions are described separately in the database so that the system can control their overgeneralization (Uramoto 1994b ). Third, a number of existing NLP tools such as JUMAN (a morphological analyzer) (Matsumoto et al. 1993 ) and QJP (a morphological and syntactic analyzer) (Kameda 1996) could broaden the coverage of our system, as inputs are currently limited to simple, morphologically analyzed sentences. Finally, it should be noted that in Japanese, case markers can be omitted or topicalized (for example, marked with postposition wa), an issue which our framework does not currently consider.", "cite_spans": [ { "start": 305, "end": 326, "text": "(Guthrie et al. 1991;", "ref_id": "BIBREF13" }, { "start": 327, "end": 341, "text": "Nasukawa 1993;", "ref_id": "BIBREF34" }, { "start": 342, "end": 355, "text": "Yarowsky 1995", "ref_id": "BIBREF57" }, { "start": 760, "end": 774, "text": "(Uramoto 1994b", "ref_id": "BIBREF52" }, { "start": 857, "end": 879, "text": "(Matsumoto et al. 1993", "ref_id": "BIBREF29" }, { "start": 931, "end": 944, "text": "(Kameda 1996)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Enhancement of Verb Sense Disambiguation", "sec_num": "2.4" }, { "text": "Let us look again at Figure 1 in Section 1. In this figure, \"WSD outputs\" refers to a corpus in which each sentence is assigned an expected verb interpretation during the WSD phase. In the training phase, the system stores supervised samples (with each interpretation simply checked or appropriately corrected by a human) in the database, to be used in a later WSD phase. In this section, we turn to the problem of which examples should be selected as samples. Lewis and Gale (1994) proposed the notion of uncertainty sampling for the training of statistics-based text classifiers. Their method selects those examples that the system classifies with minimum certainty, based on the assumption that there is no need for teaching the system the correct answer when it has answered with sufficiently high certainty. However, we should take into account the training effect a given example has on other remaining (unsupervised) examples. In other words, we would like to select samples so as to be able to correctly disambiguate as many examples as possible in the next iteration. If this is successfully done, the number of examples to be supervised will ", "cite_spans": [ { "start": 461, "end": 482, "text": "Lewis and Gale (1994)", "ref_id": "BIBREF25" } ], "ref_spans": [ { "start": 21, "end": 29, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Overview", "sec_num": "3.1" }, { "text": "Example of a given corpus associated with the verb yameru.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 6", "sec_num": null }, { "text": "decrease. We consider maximization of this effect by means of a training utility function aimed at ensuring that the most useful example at a given point in time is the example with the greatest training utility factor. Intuitively speaking, the training utility of an example is greater when we can expect greater increase in the interpretation certainty of the remaining examples after training using that example.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 6", "sec_num": null }, { "text": "To explain this notion intuitively, let us take Figure 6 as an example corpus. In this corpus, all sentences contain the verb yameru, which has two senses according to IPAL, sl ('to stop (something)') and s2 ('to quit (occupation)'). In this figure, sentences el and e2 are supervised examples associated with senses Sl and s2, respectively, and xi's are unsupervised examples. For the sake of enhanced readability, the examples xi's are partitioned according to their verb senses, that is, xl to x5 correspond to sense Sl, and x6 to x9 correspond to sense s2. In addition, note that examples in the corpus can be readily categorized based on case similarity, that is, into clusters {Xl, X2, X3, X4} ('someone/something stops service'), {Ca, X6, X7} ('someone leaves organization'), {Xs, X9} ('someone quits occupation'), {el}, and {Xs}. Let us simulate the sampling procedure with this example corpus. In the initial stage with {el, e2} in the database, x6 and x7 can be interpreted as s2 with greater certainty than for the other xi's, because these two examples are similar to e2. Therefore, uncertainty sampling selects any example except x6 and x7 as the sample. However, any one of examples Xl to x4 is more desirable because by way of incorporating one of these examples, we can obtain more xi's with greater certainty. Assuming that Xl is selected as the sample and incorporated into the database with sense Sl, either of x8 and x9 will be more highly desirable than other unsupervised xi's in the next stage.", "cite_spans": [], "ref_spans": [ { "start": 48, "end": 56, "text": "Figure 6", "ref_id": null } ], "eq_spans": [], "section": "Figure 6", "sec_num": null }, { "text": "Let S be a set of sentences, i.e., a given corpus, and D be the subset of supervised examples stored in the database. Further, let X be the set of unsupervised examples, realizing Equation (9).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 6", "sec_num": null }, { "text": "The example sampling procedure can be illustrated as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "S = D u X (9)", "sec_num": null }, { "text": "1. WSD(D, X)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "S = D u X (9)", "sec_num": null }, { "text": "2. e ~ arg maxx~x TU(x) 3.", "cite_spans": [ { "start": 18, "end": 23, "text": "TU(x)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "S = D u X (9)", "sec_num": null }, { "text": "D ~--D U {e}, X ~--X n {e} 4. goto 1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "S = D u X (9)", "sec_num": null }, { "text": "where WSD(D, X) is the verb sense disambiguation process on input X using D as the database. In this disambiguation process, the system outputs the following for ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "S = D u X (9)", "sec_num": null }, { "text": "The concept of interpretation certainty. The case where the interpretation certainty of the enclosed x's is great is shown in (a). The case where the interpretation certainty of the x's contained in the intersection of senses 1 and 2 is small is shown in (b).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 7", "sec_num": null }, { "text": "each input: (a) a set of verb sense candidates with interpretation scores, and (b) an interpretation certainty. These factors are used for the computation of TU(x), newly introduced in our method. TU(x) computes the training utility factor for an example", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 7", "sec_num": null }, { "text": "x. The sampling algorithm gives preference to examples of maximum utility.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 7", "sec_num": null }, { "text": "We will explain in the following sections how TU(x) is estimated, based on the estimation of the interpretation certainty.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 7", "sec_num": null }, { "text": "Lewis and Gale (1994) estimate certainty of an interpretation as the ratio between the probability of the most plausible text category and the probability of any other text category, excluding the most probable one. Similarly, in our verb sense disambiguation system, we introduce the notion of interpretation certainty of examples based on the following preference conditions: .", "cite_spans": [ { "start": 10, "end": 21, "text": "Gale (1994)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Interpretation Certainty", "sec_num": "3.2" }, { "text": "the highest interpretation score is greater, the difference between the highest and second highest interpretation scores is greater.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "2.", "sec_num": null }, { "text": "The rationale for these conditions is given below. Consider Figure 7 , where each symbol denotes an example in a given corpus, with symbols x as unsupervised examples and symbols e as supervised examples. The curved lines delimit the semantic vicinities (extents) of the two verb senses 1 and 2, respectively. 11 The semantic similarity between two examples is graphically portrayed by the physical distance between the two symbols representing them. In Figure 7 (a), x's located inside a semantic vicinity are expected to be interpreted as being similar to the appropriate example e with high certainty, a fact which is in line with condition 1 above. However, in Figure 7 (b), the degree of certainty for the interpretation of any x located inside the intersection of the two semantic vicinities cannot be great. This occurs when the case fillers associ- ", "cite_spans": [], "ref_spans": [ { "start": 60, "end": 68, "text": "Figure 7", "ref_id": null }, { "start": 454, "end": 462, "text": "Figure 7", "ref_id": null }, { "start": 665, "end": 673, "text": "Figure 7", "ref_id": null } ], "eq_spans": [], "section": "2.", "sec_num": null }, { "text": "The relation between coverage and accuracy with different A's.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 8", "sec_num": null }, { "text": "ated with two or more verb senses are not selective enough to allow for a clear-cut delineation between them. This situation is explicitly rejected by condition 2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 8", "sec_num": null }, { "text": "Based on the above two conditions, we compute interpretation certainties using Equation (10), where C(x) is the interpretation certainty of an example x, Scorel(x) and Score2(x) are the highest and second highest scores for x, respectively, and ,~, which ranges from 0 to 1, is a parametric constant used to control the degree to which each condition affects the computation of C(x).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 8", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "C(x)=A. Scorel(x)+(1 -~).(Scorel(x) -Score2(x))", "eq_num": "(10)" } ], "section": "Figure 8", "sec_num": null }, { "text": "Through a preliminary experiment, we estimated the validity of the notion of interpretation certainty, by the trade-off between accuracy and coverage of the system. Note that in this experiment, accuracy is the ratio of the number of correct outputs and the number of cases where the interpretation certainty of the output is above a certain threshold. Coverage is the ratio of the number of cases where the interpretation certainty of the output is above a certain threshold and the number of inputs. By raising the value of the threshold, accuracy also increases (at least theoretically), while coverage decreases.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 8", "sec_num": null }, { "text": "The system used the Bunruigoihyo thesaurus for the similarity computation, and was evaluated by way of sixfold cross-validation using the same corpus as that used for the experiment described in Section 2.3. Figure 8 shows the result of the experiment with several values of ,~, from which the optimal )~ value seems to be in the range around 0.5. It can be seen that, as we assumed, both of the above conditions are essential for the estimation of interpretation certainty. The concept of training utility. The case where the training utility of a is greater than that of b because a has more unsupervised neighbors is shown in (a); (b) shows the case where the training utility of a is greater than that of b because b closely neighbors e, contained in the database.", "cite_spans": [], "ref_spans": [ { "start": 208, "end": 216, "text": "Figure 8", "ref_id": null } ], "eq_spans": [], "section": "Figure 8", "sec_num": null }, { "text": "The training utility of an example a is greater than that of another example b when the total interpretation certainty of unsupervised examples increases more after training with example a than with example b. Let us consider Figure 9 , in which the x-axis mono-dimensionally denotes the semantic similarity between two unsupervised examples, and the y-axis denotes the interpretation certainty of each example. Let us compare the training utility of the examples a and b in Figure 9 (a). Note that in this figure, whichever example we use for training, the interpretation certainty for each unsupervised example (x) neighboring the chosen example increases based on its similarity to the supervised example. Since the increase in the interpretation certainty of a given x becomes smaller as the similarity to a or b diminishes, the training utility of the two examples can be represented by the shaded areas. The training utility of a is greater as it has more neighbors than b. On the other hand, in Figure 9 (b), b has more neighbors than a. However, since b is semantically similar to e, which is already contained in the database, the total increase in interpretation certainty of its neighbors, i.e. the training utility of b, is smaller than that of a. Let AC(x = s,y) be the difference in the interpretation certainty of y c X after training with x c X, taken with the sense s. TU(x = s), which is the training utility function for x taken with sense s, can be computed by Equation (11).", "cite_spans": [], "ref_spans": [ { "start": 226, "end": 234, "text": "Figure 9", "ref_id": null }, { "start": 475, "end": 483, "text": "Figure 9", "ref_id": null }, { "start": 1002, "end": 1010, "text": "Figure 9", "ref_id": null } ], "eq_spans": [], "section": "Training Utility", "sec_num": "3.3" }, { "text": "yEX It should be noted that in Equation 11, we can replace X with a subset of X that consists of neighbors of x. However, in order to facilitate this, an efficient algorithm to search for neighbors of an example is required. We will discuss this problem in Section 3.5. Since there is no guarantee that x will be supervised with any given sense s, it can be risky to rely solely on TU(x = s) for the computation of TU(x). We estimate TU(x), by the expected value of x, calculating the average of each TU(x = s), weighted by the probability that x takes sense s. This can be realized by Equation (12), where P(slx ) is the probability that x takes the sense s.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TU(x = s) = E AC(x = s,y)", "sec_num": null }, { "text": "s Given the fact that (a) P(sIx ) is difficult to estimate in the current formulation, and (b) the cost of computation for each TU(x = s) is not trivial, we temporarily approximate TU(x) as in Equation 13, where K is a set of the k-best verb sense(s) of x with respect to the interpretation score in the current state.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TU(x) = E P(slx) \" TU(x = s)", "sec_num": null }, { "text": "1 E TU(x = s)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TU(x) = E P(slx) \" TU(x = s)", "sec_num": null }, { "text": "(13) TU(x) ; sEK", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TU(x) = E P(slx) \" TU(x = s)", "sec_num": null }, { "text": "In this section, we discuss how to enhance the computation associated with our example sampling algorithm.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Enhancement of Computation", "sec_num": "3.4" }, { "text": "First, we note that computation of TU(x = s) in Equation 11above becomes time consuming because the system is required to search the whole set of unsupervised examples for examples whose interpretation certainty will increase after x is used for training. To avoid this problem, we could apply a method used in efficient database search techniques, by which the system can search for neighbor examples of x with optimal time complexity (Utsuro et al. 1994) . However, in this section, we will explain another efficient algorithm to identify neighbors of x, in which neighbors of case fillers are considered to be given directly by the thesaurus structure. 12 The basic idea is the following: the system searches for neighbors of each case filler of x instead of x as a whole, and merges them as a set of neighbors of x. Note that by dividing examples along the lines of each case filler, we can retrieve neighbors based on the structure of the Bunruigoihyo thesaurus (instead of the conceptual semantic space as in Figure 7 ).", "cite_spans": [ { "start": 436, "end": 456, "text": "(Utsuro et al. 1994)", "ref_id": "BIBREF54" } ], "ref_spans": [ { "start": 1015, "end": 1023, "text": "Figure 7", "ref_id": null } ], "eq_spans": [], "section": "Enhancement of Computation", "sec_num": "3.4" }, { "text": "Let Nx=s,c be a subset of unsupervised neighbors of x whose interpretation certainty will increase after x is used for training, considering only case c of sense s. The actual neighbor set of x with sense s (Nx=s) is then defined as in Equation 14.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Enhancement of Computation", "sec_num": "3.4" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "Nx=s=UNx:s,c", "eq_num": "(14)" } ], "section": "Enhancement of Computation", "sec_num": "3.4" }, { "text": "\u00a3 Figure 10 shows a fragment of the thesaurus, in which the x and the y's are unsupervised case filler examples. Symbols el and e2 are case filler examples stored in the database taken as senses sl and s2, respectively. The triangles represent subtrees of the structure, and the labels ni represent nodes. In this figure, it can be seen that the interpretation score of Sl never changes for examples other than the children of n4, after x is used for training with sense Sl. In addition, incorporating x into the database with sense sl never changes the score of examples y for other sense candidates. Therefore, Nx=sl,c includes only examples dominated by n4, in other words, examples that are closer to x than el in the thesaurus structure. Since, during the WSD phase, the system determines el as the supervised neighbor of x for sense Sl, identifying Nx=sl,c does not require any extra computational overhead. We should point out that the technique presented here is not applicable when the vector space model (see Section 2.2) is used for the similarity computation. However, automatic clustering algorithms, which assign a hierarchy to a set of words based on the similarity between them (such as the one proposed by Tokunaga, Iwayama, and Tanaka [1995] ), could potentially facilitate the application of this retrieval method to the vector space model. Second, sample size at each iteration should ideally be one, so as to avoid the supervision of similar examples. On the other hand, a small sampling size generates a considerable computation overhead for each iteration of the sampling procedure. This can be a critical problem for statistics-based approaches, as the reconstruction", "cite_spans": [ { "start": 1223, "end": 1259, "text": "Tokunaga, Iwayama, and Tanaka [1995]", "ref_id": "BIBREF50" } ], "ref_spans": [ { "start": 2, "end": 11, "text": "Figure 10", "ref_id": null } ], "eq_spans": [], "section": "Enhancement of Computation", "sec_num": "3.4" }, { "text": "J ?7,2 n5 L 7~3 7~4", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Enhancement of Computation", "sec_num": "3.4" }, { "text": "el y y y x y e2 y y", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Enhancement of Computation", "sec_num": "3.4" }, { "text": "A fragment of the thesaurus including neighbors of x associated with case c.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 10", "sec_num": null }, { "text": "of statistic classifiers is expensive. However, example-based systems fortunately do not require reconstruction, and examples simply have to be stored in the database. Furthermore, in each disambiguation phase, our example-based system needs only to compute the similarity between each newly stored example and its unsupervised neighbors, rather than between every example in the database and every unsupervised example. Let us reconsider Figure 10 . As mentioned above, when x is stored in the database with sense sl, only the interpretation score of y's dominated by n4, i.e., Nx=sl,c, will be changed with respect to sense sl. This algorithm reduces the time complexity of each iteration from O(N 2) to O(N), given that N is the total number of examples in a given corpus.", "cite_spans": [], "ref_spans": [ { "start": 439, "end": 448, "text": "Figure 10", "ref_id": null } ], "eq_spans": [], "section": "Figure 10", "sec_num": null }, { "text": "3.5.1 Sense Ambiguity of Case Fillers in Selective Sampling. The semantic ambiguity of case fillers (nouns) should be taken into account during selective sampling. Figure 11 , which uses the same basic notation as Figure 7 , illustrates one possible problem caused by case filler ambiguity. Let xl be a sense of a case filler x, and Yl and y2 be different senses of a case filler y. On the basis of Equation (10), the interpretation certainty of x and y is small in Figures 11(a) and 11(b), respectively. However, in the situation shown in Figure 11 (b), since (a) the task of distinguishing between the verb senses 1 and 2 is easier, and (b) instances where the sense ambiguity of case fillers corresponds to distinct verb senses will be rare, training using either yl or y2 will be less effective than using a case filler of the type of x. It should also be noted that since Bunruigoihyo is a relatively small-sized thesaurus with limited word sense coverage, this problem is not critical in our case. However, given other existing thesauri like the EDR electronic dictionary (Japan Electronic Dictionary Research Institute 1995) or WordNet (Miller et al. 1993 ), these two situations should be strictly differentiated. Figure 12 , where the basic notation is the same as in Figure 7 , exemplifies a limitation of our sampling method. In this figure, the only supervised examples contained in the database are el and e2, and x represents an unsupervised example belonging to sense 2. Given this scenario, x is informative because (a) it clearly evidences the semantic vicinity of sense 2, and (b) without x as sense 2 in the database, the system may misinterpret other examples neighboring x. However, in our current implementation, the training utility of x would be small because it would be mistakenly interpreted as sense I with great certainty due to its relatively close semantic proximity to el. Even if x has a number of unsupervised neighbors, the total increment of their interpretation certainty cannot be expected to be large. This shortcoming often presents itself when the semantic vicinities of different sense 1 / (.~sense 2 sense 1 sense 2 (a) (b)", "cite_spans": [ { "start": 1143, "end": 1162, "text": "(Miller et al. 1993", "ref_id": "BIBREF30" } ], "ref_spans": [ { "start": 164, "end": 173, "text": "Figure 11", "ref_id": null }, { "start": 214, "end": 222, "text": "Figure 7", "ref_id": null }, { "start": 466, "end": 479, "text": "Figures 11(a)", "ref_id": null }, { "start": 540, "end": 549, "text": "Figure 11", "ref_id": null }, { "start": 1222, "end": 1231, "text": "Figure 12", "ref_id": null }, { "start": 1277, "end": 1285, "text": "Figure 7", "ref_id": null } ], "eq_spans": [], "section": "Discussion", "sec_num": "3.5" }, { "text": "Two separate scenarios in which the interpretation certainty of x is small. In (a), interpretation certainty of x is small because x lies in the intersection of distinct verb senses; in (b), interpretation certainty of y is small because y is semantically ambiguous.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 11", "sec_num": null }, { "text": "sense ~ sense 2", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 11", "sec_num": null }, { "text": "The case where informative example x is not selected.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 12", "sec_num": null }, { "text": "verb senses are closely aligned or their semantic ranges are not disjunctive. Here, let us consider Figure 3 again, in which the nominative case would parallel the semantic space shown in Figure 12 more closely than the accusative. Relying more on the similarity in the accusative (the case with greater CCD) as is done in our system, we aim to map the semantic space in such a way as to achieve higher semantic disparity and minimize this shortcoming.", "cite_spans": [], "ref_spans": [ { "start": 100, "end": 108, "text": "Figure 3", "ref_id": null }, { "start": 188, "end": 197, "text": "Figure 12", "ref_id": null } ], "eq_spans": [], "section": "Figure 12", "sec_num": null }, { "text": "~n order to investigate the effectiveness of our example sampling method, we conducted an experiment in which we compared the following four sampling methods:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparative Experimentation", "sec_num": "4.1" }, { "text": "\u2022 a control (random), in which a certain proportion of a given corpus is randomly selected for training,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparative Experimentation", "sec_num": "4.1" }, { "text": "\u2022 uncertainty sampling (US), in which examples with minimum interpretation certainty are selected (Lewis and Gale 1994 ),", "cite_spans": [ { "start": 98, "end": 118, "text": "(Lewis and Gale 1994", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Comparative Experimentation", "sec_num": "4.1" }, { "text": "\u2022 committee-based sampling (CBS) (Engelson and Dagan 1996) ,", "cite_spans": [ { "start": 33, "end": 58, "text": "(Engelson and Dagan 1996)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Comparative Experimentation", "sec_num": "4.1" }, { "text": "\u2022 our method based on the notion of training utility (TU).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparative Experimentation", "sec_num": "4.1" }, { "text": "We elaborate on uncertainty sampling and committee-based sampling in Section 4.2. We compared these sampling methods by evaluating the relation between the number of training examples sampled and the performance of the system. We conducted sixfold cross-validation and carried out sampling on the training set. With regard to the training/test data set, we used the same corpus as that used for the experiment described in Section 2.3. Each sampling method uses examples from IPAL to initialize the system, with the number of example case fillers for each case being an average of about 3.7. For each sampling method, the system uses the Bunruigoihyo thesaurus for the similarity computation. In Table 2 (in Section 2. 3), the column of \"accuracy\" for \"BGH\" denotes the accuracy of the system with the entire set of training data contained in the database. Each of the four sampling methods achieved this figure at the conclusion of training. We evaluated each system performance according to its accuracy, that is the ratio of the number of correct outputs, compared to the total number of inputs. For the purpose of this experiment, we set the sample size to 1 for each iteration, A = 0.5 for Equation 10, and k = 1 for Equation (13). Based on a preliminary experiment, increasing the value of k either did not improve the performance over that for k = 1, or lowered the overall performance. Figure 13 shows the relation between the number of training data sampled and the accuracy of the system. In Figure 13 , zero on the x-axis represents the system using only the examples provided by 1PAL. Looking at Figure 13 one can see that compared with random sampling and committee-based sampling, our sampling method reduced the number of the training data required to achieve any given accuracy. For example, to achieve an accuracy of 80%, the number of training data required for our method was roughly one-third of that for random sampling. Although the accuracy of our method was surpassed by that of uncertainty sampling for larger sizes of training data, this minimal difference for larger data sizes is overshadowed by the considerable performance gain attained by our method for smaller data sizes.", "cite_spans": [], "ref_spans": [ { "start": 696, "end": 718, "text": "Table 2 (in Section 2.", "ref_id": null }, { "start": 1394, "end": 1403, "text": "Figure 13", "ref_id": null }, { "start": 1502, "end": 1511, "text": "Figure 13", "ref_id": null }, { "start": 1608, "end": 1617, "text": "Figure 13", "ref_id": null } ], "eq_spans": [], "section": "Comparative Experimentation", "sec_num": "4.1" }, { "text": "Since IPAL has, in a sense, been manually selectively sampled in an attempt to model the maximum verb sense coverage, the performance of each method is biased by the initial contents of the database. To counter this effect, we also conducted an experiment involving the construction of the database from scratch, without using examples from IPAL. During the initial phase, the system randomly selected one example for each verb sense from the training set, and a human expert provided the correct interpretation to initialize the system. Figure 14 shows the performance of the various methods, from which the same general tendency as seen in Figure 13 is observable. However, in this case, our method was generally superior to other methods. Through these comparative experiments, we can conclude that our example sampling method is able to decrease the number of training data, i.e., the overhead for both supervision and searching, without degrading the system performance. ", "cite_spans": [], "ref_spans": [ { "start": 538, "end": 547, "text": "Figure 14", "ref_id": null }, { "start": 642, "end": 651, "text": "Figure 13", "ref_id": null } ], "eq_spans": [], "section": "Comparative Experimentation", "sec_num": "4.1" }, { "text": "The relation between the number of training data sampled and the accuracy of the system without using examples from IPAL. Sompolinsky 1992), the system selects samples based on the degree of disagreement between models randomly taken from a given training set (these models are called \"committee members\"). This is achieved by iteratively repeating the steps given below, in which the number of committee members is given as two without loss of generality: \u00b0 2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 14", "sec_num": null }, { "text": ". draw two models randomly, classify unsupervised example x according to each model, producing classifications C1 and C2, if C1 # C2 (the committee members disagree), select x for the training of the system. Figure 15 shows a typical disparity evident between committee-based sampling and our sampling method. The basic notation in this figure is the same as in Figure 7 , and both x and y denote unsupervised examples, or more formally D = {el, e2}, and X = {x, y}. Assume a pair of committee members {el} and {e2 } have been selected from the database D. In this case, the committee members disagree as to the interpretations of both x and y, and consequently, either example can potentially be selected as a sample for the next iteration. In fact, committee-based sampling tends to require a number of similar examples (similar to el and y) in the database, otherwise committee members taken from the database will never agree. This is in contrast to our method, in which similar examples are less informative. In our method, therefore, x is preferred to y as a sample. This contrast can also correlate to the fact that committee-based sampling is currently applied to statistics-based language models (HMM classifiers), in other words, statistical models generally require that the distribution of the training data _sense 1 s sense 2", "cite_spans": [], "ref_spans": [ { "start": 208, "end": 217, "text": "Figure 15", "ref_id": null }, { "start": 362, "end": 370, "text": "Figure 7", "ref_id": null } ], "eq_spans": [], "section": "Figure 14", "sec_num": null }, { "text": "A case where either x or y can be selected in committee-based sampling.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 15", "sec_num": null }, { "text": "reflects that of the overall text. Through this argument, one can assume that committeebased sampling is better suited to statistics-based systems, while our method is more suitable for example-based systems. Engelson and Dagan (1996) criticized uncertainty sampling (Lewis and Gale 1994) , which they call a \"single model\" approach, as distinct from their \"multiple model\" approach: sufficient statistics may yield an accurate 0.51 probability estimate for a class c in a given example, making it certain that c is the appropriate classificationJ 3 However, the certainty that c is the correct classification is low, since there is a 0.49 chance that c is the wrong class for the example. A single model can be used to estimate only the second type of uncertainty, which does not correlate directly with the utility of additional training. (p. 325) We note that this criticism cannot be applied to our sampling method, despite the fact that our method falls into the category of a single model approach. In our sampling method, given sufficient statistics, the increment of the certainty degree for unsupervised examples, i.e., the training utility of additional supervised examples, becomes small (theoretically, for both example-based and statistics-based systems). Thus, the utility factor can be considered to correlate directly with additional training, for our method.", "cite_spans": [ { "start": 209, "end": 234, "text": "Engelson and Dagan (1996)", "ref_id": "BIBREF7" }, { "start": 267, "end": 288, "text": "(Lewis and Gale 1994)", "ref_id": "BIBREF25" }, { "start": 841, "end": 849, "text": "(p. 325)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Figure 15", "sec_num": null }, { "text": "Corpus-based approaches have recently pointed the way to a promising trend in word sense disambiguation. However, these approaches tend to require a considerable overhead for supervision in constructing a large-sized database, additionally resulting in a computational overhead to search the database. To overcome these problems, our method, which is currently applied to an example-based verb sense disambiguation system, selectively samples a smaller-sized subset from a given example set. This method is expected to be applicable to other example-based systems. Applicability for other types of systems needs to be further explored.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5." }, { "text": "The process basically iterates through two phases: (normal) word sense disambiguation and a training phase. During the disambiguation phase, the system is provided with sentences containing a polysemous verb, and searches the database for the 13 By appropriate classification, Engelson and Dagan mean the classification given by a perfectly trained model. most semantically similar example to the input (nearest neighbor resolution). Thereafter, the verb is disambiguated by superimposing the sense of the verb appearing in the supervised example. The similarity between the input and an example, or more precisely the similarity between the case fillers included in them, is computed based on an existing thesaurus. In the training phase, a sample is then selected from the system outputs and provided with the correct interpretation by a human expert. Through these two phases, the system iteratively accumulates supervised examples into the database. The critical issue in this process is to decide which example should be selected as a sample in each iteration. To resolve this problem, we considered the following properties: (a) an example that neighbors more unsupervised examples is more influential for subsequent training, and therefore more informative, and (b) since our verb sense disambiguation is based on nearest neighbor resolution, an example similar to one already existing in the database is redundant. Motivated by these properties, we introduced and formalized the concept of training utility as the criterion for example selection. Our sampling method always gives preference to that example which maximizes training utility. We reported on the performance of our sampling method by way of experiments in which we compared our method with random sampling, uncertainty sampling (Lewis and Gale 1994) , and committee-based sampling (Engelson and Dagan 1996) . The result of the experiments showed that our method reduced both the overhead for supervision and the overhead for searching the database to a larger degree than any of the above three methods, without degrading the performance of verb sense disambiguation. Through the experiment and discussion, we claim that uncertainty sampling considers property (b) mentioned above, but lacks property (a). We also claim that committee-based sampling differs from our sampling method in terms of its suitability to statistics-based systems as compared to example-based systems.", "cite_spans": [ { "start": 1800, "end": 1821, "text": "(Lewis and Gale 1994)", "ref_id": "BIBREF25" }, { "start": 1853, "end": 1878, "text": "(Engelson and Dagan 1996)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5." }, { "text": "Note that these problems are associated with corpus-based approaches in general, and have been identified by a number of researchers(Engelson and Dagan 1996;Lewis and Gale 1994;Uramoto 1994a;Yarowsky 1995).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that unlike the automatic acquisition of word sense definitions(Fukumoto and Tsujii 1994;Pustejovsky and Boguraev 1993;Utsuro 1996;Zernik 1989), the task of the system is to identify the best matched category with a given input, from predefined candidates.3 In this paper, we use \"example-based systems\" to refer to systems based on nearest neighbor resolution.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Different types of application of hand-crafted thesauri to word sense disambiguation have been proposed, for example, byYarowsky (1992).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that each verb in co-occurrence data should ideally be annotated with its verb sense. However, there is no existing Japanese text base with sufficient volume of word sense tags.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that previous research has applied this technique to tasks other than verb sense disambiguation, such as syntactic disambiguation(Resnik 1993) and disambiguation of case filler noun senses(Ribas 1995).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Ideally speaking, training and test data should be drawn from different sources, to simulate a real application. However, the sentences were already scrambled when provided to us, and therefore we could not identify the original source corresponding to each sentence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "This problem has also been identified byCharniak (1993).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that this method can easily be extended for a verb with more than two senses. In Section 4, we describe an experiment using multiply polysemous verbs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Utsuro's method requires the constiuction of large-scale similarity templates prior to similarity computation(Utsuro et al. 1994), and this is what we would like to avoid.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "The authors would like to thank Manabu Okumura (JAIST, Japan), Timothy Baldwin (TITECH, Japan), Michael Zock (LIMSI, France), Dan Tufts (Romanian Academy, Romania) and anonymous reviewers for their comments on an earlier version of this paper. This research is partially supported by a Research Fellowship of the Japan Society for the Promotion of Science for Young Scientists.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null }, { "text": " I I I I I 80 {i 65 ~__._ ~---? .... ~----Z;:;; = S .... ~ ..... ~. ..... ~_ ..... ~ .... 2I~.:.:.: .:''''' ", "cite_spans": [ { "start": 36, "end": 117, "text": "~---? .... ~----Z;:;; = S .... ~ ..... ~. ..... ~_ ..... ~ .... 2I~.:.:.: .:'''''", "ref_id": null } ], "ref_spans": [ { "start": 1, "end": 29, "text": "I I I I I 80 {i 65", "ref_id": null } ], "eq_spans": [], "section": "85", "sec_num": null }, { "text": "The relation between the number of training data sampled and the accuracy of the system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 13", "sec_num": null }, { "text": "D ~ DU {e}, X +-XN {e}", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "3.", "sec_num": null }, { "text": "Let us discuss the theoretical difference between this and our method. Considering Figure 9 again, one can see that the concept of training utility is supported by the following properties: .", "cite_spans": [], "ref_spans": [ { "start": 83, "end": 91, "text": "Figure 9", "ref_id": null } ], "eq_spans": [], "section": "goto i", "sec_num": "4." }, { "text": "an example that neighbors more unsupervised examples is more informative (Figure 9(a) ), an example less similar to one already existing in the database is more informative (Figure 9(b) ).Uncertainty sampling directly addresses the second property but ignores the first. It differs from our method more crucially when more unsupervised examples remain, because these unsupervised examples have a greater influence on the computation of training utility. This can be seen in the comparative experiments in Section 4, in which our method outperformed uncertainty sampling to the highest degree in early stages.", "cite_spans": [], "ref_spans": [ { "start": 73, "end": 85, "text": "(Figure 9(a)", "ref_id": null }, { "start": 173, "end": 185, "text": "(Figure 9(b)", "ref_id": null } ], "eq_spans": [], "section": "2.", "sec_num": null }, { "text": "In committee-based sampling (Engelson and Dagan 1996) , which follows the \"query by committee\" principle (Seung, Opper, and ", "cite_spans": [ { "start": 28, "end": 53, "text": "(Engelson and Dagan 1996)", "ref_id": "BIBREF7" }, { "start": 105, "end": 123, "text": "(Seung, Opper, and", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Committee-based Sampling.", "sec_num": "4.2.2" } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Instance-based learning algorithms", "authors": [ { "first": "David", "middle": [ "W" ], "last": "Aha", "suffix": "" }, { "first": "Dennis", "middle": [], "last": "Kibler", "suffix": "" }, { "first": "Marc", "middle": [ "K" ], "last": "Albert", "suffix": "" } ], "year": 1991, "venue": "Machine Learning", "volume": "6", "issue": "", "pages": "37--66", "other_ids": {}, "num": null, "urls": [], "raw_text": "Aha, David W., Dennis Kibler, and Marc K. Albert. 1991. Instance-based learning algorithms. Machine Learning, 6(1):37-66.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Word-sense disambiguation using statistical methods", "authors": [ { "first": "Peter", "middle": [ "E" ], "last": "Brown", "suffix": "" }, { "first": "Stephen", "middle": [ "A" ], "last": "Della Pietra", "suffix": "" }, { "first": "Vincent", "middle": [ "J Della" ], "last": "Pietra", "suffix": "" } ], "year": 1991, "venue": "Proceedings of the 29th Annual Meeting", "volume": "", "issue": "", "pages": "264--270", "other_ids": {}, "num": null, "urls": [], "raw_text": "Brown, Peter E, Stephen A. Della Pietra, and Vincent J. Della Pietra. 1991. Word-sense disambiguation using statistical methods. In Proceedings of the 29th Annual Meeting, pages 264-270, Association for Computational Linguistics.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Word-sense disambiguation using decomposable models", "authors": [ { "first": "Rebecca", "middle": [], "last": "Bruce", "suffix": "" }, { "first": "Janyce", "middle": [], "last": "Wiebe", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 32nd Annual Meeting", "volume": "", "issue": "", "pages": "139--146", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bruce, Rebecca and Janyce Wiebe. 1994. Word-sense disambiguation using decomposable models. In Proceedings of the 32nd Annual Meeting, pages 139-146, Association for Computational Linguistics.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Roget's International Thesaurus. Fourth Edition", "authors": [ { "first": "Robert", "middle": [ "L" ], "last": "Chapman", "suffix": "" } ], "year": 1984, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chapman, Robert L. 1984. Roget's International Thesaurus. Fourth Edition. Harper and Row.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Statistical Language Learning", "authors": [ { "first": "Eugene", "middle": [], "last": "Charniak", "suffix": "" } ], "year": 1993, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Charniak, Eugene. 1993. Statistical Language Learning. MIT Press, Cambridge, MA.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Improving generalization with active learning", "authors": [ { "first": "David", "middle": [], "last": "Cohn", "suffix": "" }, { "first": "Les", "middle": [], "last": "Atlas", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Ladner", "suffix": "" } ], "year": 1994, "venue": "Machine Learning", "volume": "15", "issue": "", "pages": "201--221", "other_ids": {}, "num": null, "urls": [], "raw_text": "Cohn, David, Les Atlas, and Richard Ladner. 1994. Improving generalization with active learning. Machine Learning, 15(2):201-221.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Word sense disambiguation using a second language monolingual corpus", "authors": [ { "first": "Ido", "middle": [], "last": "Dagan", "suffix": "" }, { "first": "Alon", "middle": [], "last": "Itai", "suffix": "" } ], "year": 1994, "venue": "Computational Linguistics", "volume": "20", "issue": "4", "pages": "563--596", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dagan, Ido and Alon Itai. 1994. Word sense disambiguation using a second language monolingual corpus. Computational Linguistics, 20(4):563-596.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Minimizing manual annotation cost in supervised training from corpora", "authors": [ { "first": "Sean", "middle": [ "P" ], "last": "Engelson", "suffix": "" }, { "first": "Ido", "middle": [], "last": "Dagan", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 34th Annual Meeting", "volume": "", "issue": "", "pages": "319--326", "other_ids": {}, "num": null, "urls": [], "raw_text": "Engelson, Sean P. and Ido Dagan. 1996. Minimizing manual annotation cost in supervised training from corpora. In Proceedings of the 34th Annual Meeting, pages 319-326, Association for Computational Linguistics.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Information Retrieval: Data Structure & Algorithms", "authors": [ { "first": "William", "middle": [ "B" ], "last": "Frakes", "suffix": "" }, { "first": "Ricardo", "middle": [], "last": "Baeza-Yates", "suffix": "" } ], "year": 1992, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Frakes, William B. and Ricardo Baeza-Yates. 1992. Information Retrieval: Data Structure & Algorithms. PTR Prentice-Hall.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "To what extent does case contribute to verb sense disambiguation?", "authors": [ { "first": "Atsushi", "middle": [], "last": "Fujii", "suffix": "" }, { "first": "Kentaro", "middle": [], "last": "Inui", "suffix": "" }, { "first": "Takenobu", "middle": [], "last": "Tokunaga", "suffix": "" }, { "first": "Hozumi", "middle": [], "last": "Tanaka", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 16th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "59--64", "other_ids": {}, "num": null, "urls": [], "raw_text": "Fujii, Atsushi, Kentaro Inui, Takenobu Tokunaga, and Hozumi Tanaka. 1996. To what extent does case contribute to verb sense disambiguation? In Proceedings of the 16th International Conference on Computational Linguistics, pages 59-64.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Automatic recognition of verbal polysemy", "authors": [ { "first": "Fumiyo", "middle": [], "last": "Fukumoto", "suffix": "" }, { "first": "Jun'ichi", "middle": [], "last": "Tsujii", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 15th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "764--768", "other_ids": {}, "num": null, "urls": [], "raw_text": "Fukumoto, Fumiyo and Jun'ichi Tsujii. 1994. Automatic recognition of verbal polysemy. In Proceedings of the 15th International Conference on Computational Linguistics, pages 764-768.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Estimating upper and lower bounds on the performance of word-sense disambiguation programs", "authors": [ { "first": "William", "middle": [], "last": "Gale", "suffix": "" }, { "first": "Kenneth", "middle": [ "Ward" ], "last": "Church", "suffix": "" }, { "first": "David", "middle": [], "last": "Yarowsky", "suffix": "" } ], "year": 1992, "venue": "Proceedings of the 30th Annual Meeting", "volume": "", "issue": "", "pages": "249--256", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gale, William, Kenneth Ward Church, and David Yarowsky. 1992. Estimating upper and lower bounds on the performance of word-sense disambiguation programs. In Proceedings of the 30th Annual Meeting, pages 249-256, Association for Computational Linguistics.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "A method for disambiguating word senses in a large corpus", "authors": [ { "first": "William", "middle": [], "last": "Gale", "suffix": "" }, { "first": "Kenneth", "middle": [ "Ward" ], "last": "Church", "suffix": "" }, { "first": "David", "middle": [], "last": "Yarowsky", "suffix": "" } ], "year": 1993, "venue": "Computers and the Humanities", "volume": "26", "issue": "", "pages": "415--439", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gale, William, Kenneth Ward Church, and David Yarowsky. 1993. A method for disambiguating word senses in a large corpus. Computers and the Humanities, 26:415-439.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Subject-dependent co-occurrence and word sense disambiguation", "authors": [ { "first": "Joe", "middle": [ "A" ], "last": "Guthrie", "suffix": "" }, { "first": "Louise", "middle": [], "last": "Guthrie", "suffix": "" }, { "first": "Yorick", "middle": [], "last": "Wilks", "suffix": "" }, { "first": "Homa", "middle": [], "last": "Aidinejad", "suffix": "" } ], "year": 1991, "venue": "Proceedings of the 29th Annual Meeting", "volume": "", "issue": "", "pages": "146--152", "other_ids": {}, "num": null, "urls": [], "raw_text": "Guthrie, Joe A., Louise Guthrie, Yorick Wilks, and Homa Aidinejad. 1991. Subject-dependent co-occurrence and word sense disambiguation. In Proceedings of the 29th Annual Meeting, pages 146-152, Association for Computational Linguistics.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Noun homograph disambiguation using local context in large text corpora", "authors": [ { "first": "Marti", "middle": [ "A" ], "last": "Hearst", "suffix": "" } ], "year": 1991, "venue": "Proceedings of the 7th Annual Conference of the University of Waterloo Centre for the New OED and Text Research", "volume": "", "issue": "", "pages": "1--22", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hearst, Marti A. 1991. Noun homograph disambiguation using local context in large text corpora. In Proceedings of the 7th Annual Conference of the University of Waterloo Centre for the New OED and Text Research, pages 1-22.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Noun classification from predicate-argument structures", "authors": [ { "first": "Donald", "middle": [], "last": "Hindle", "suffix": "" } ], "year": 1990, "venue": "Proceedings of the 28th Annual Meeting", "volume": "", "issue": "", "pages": "268--275", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hindle, Donald. 1990. Noun classification from predicate-argument structures. In Proceedings of the 28th Annual Meeting, pages 268-275, Association for Computational Linguistics.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Semantic Interpretation and the Resolution of Ambiguity", "authors": [ { "first": "Graeme", "middle": [], "last": "Hirst", "suffix": "" } ], "year": 1987, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hirst, Graeme. 1987. Semantic Interpretation and the Resolution of Ambiguity. Cambridge University Press.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "IPAL Japanese dictionary for computers (basic verbs", "authors": [], "year": 1987, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Information-technology Promotion Agency. 1987. IPAL Japanese dictionary for computers (basic verbs) (in Japanese).", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Japan Electronic Dictionary Research Institute. 1995. EDR electronic dictionary technical guide", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Japan Electronic Dictionary Research Institute. 1995. EDR electronic dictionary technical guide (in Japanese).", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Learning translation templates from bilingual text", "authors": [ { "first": "Hiroyuki", "middle": [], "last": "Kaji", "suffix": "" }, { "first": "Yuuko", "middle": [], "last": "Kida", "suffix": "" }, { "first": "Yasutsugu", "middle": [], "last": "Morimoto", "suffix": "" } ], "year": 1992, "venue": "Proceedings of the 14th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "672--678", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kaji, Hiroyuki, Yuuko Kida, and Yasutsugu Morimoto. 1992. Learning translation templates from bilingual text. In Proceedings of the 14th International Conference on Computational Linguistics, pages 672-678.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "A portable & quick Japanese parser: QJP", "authors": [ { "first": "Masayuki", "middle": [], "last": "Kameda", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 16th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "616--621", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kameda, Masayuki. 1996. A portable & quick Japanese parser: QJP. In Proceedings of the 16th International Conference on Computational Linguistics, pages 616-621.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Learning similarity-based word sense disambiguation", "authors": [ { "first": "Yael", "middle": [], "last": "Karov", "suffix": "" }, { "first": "Shimon", "middle": [], "last": "Edelman", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 4th Workshop on Very Large Corpora", "volume": "", "issue": "", "pages": "42--55", "other_ids": {}, "num": null, "urls": [], "raw_text": "Karov, Yael and Shimon Edelman. 1996. Learning similarity-based word sense disambiguation. In Proceedings of the 4th Workshop on Very Large Corpora, pages 42-55.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Lexical ambiguity and information retrieval", "authors": [ { "first": "Robert", "middle": [], "last": "Krovets", "suffix": "" }, { "first": "W. Bruce", "middle": [], "last": "Croft", "suffix": "" } ], "year": 1992, "venue": "ACM Transactions on Information Systems", "volume": "10", "issue": "2", "pages": "115--141", "other_ids": {}, "num": null, "urls": [], "raw_text": "Krovets, Robert and W. Bruce Croft. 1992. Lexical ambiguity and information retrieval. ACM Transactions on Information Systems, 10(2):115-141.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "A method of case structure analysis for Japanese sentences based on examples in case frame dictionary", "authors": [ { "first": "Sadao", "middle": [], "last": "Kurohashi", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1994, "venue": "IEICE Transactions on Information and Systems", "volume": "", "issue": "2", "pages": "227--239", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kurohashi, Sadao and Makoto Nagao. 1994. A method of case structure analysis for Japanese sentences based on examples in case frame dictionary. IEICE Transactions on Information and Systems, E77-D(2):227-239.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Corpus-based statistical sense resolution", "authors": [ { "first": "Claudia", "middle": [], "last": "Leacock", "suffix": "" }, { "first": "Geoffrey", "middle": [], "last": "Towell", "suffix": "" }, { "first": "Ellen", "middle": [], "last": "Voorhees", "suffix": "" } ], "year": 1993, "venue": "Proceedings of ARPA Human Language Technology Workshop", "volume": "", "issue": "", "pages": "260--265", "other_ids": {}, "num": null, "urls": [], "raw_text": "Leacock, Claudia, Geoffrey Towell, and Ellen Voorhees. 1993. Corpus-based statistical sense resolution. In Proceedings of ARPA Human Language Technology Workshop, pages 260-265.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "A sequential algorithm for training text classifiers", "authors": [ { "first": "David", "middle": [ "D" ], "last": "Lewis", "suffix": "" }, { "first": "William", "middle": [], "last": "Gale", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "3--12", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lewis, David D. and William Gale. 1994. A sequential algorithm for training text classifiers. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3-12.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "A WordNet-based algorithm for word sense disambiguation", "authors": [ { "first": "Xiaobin", "middle": [], "last": "Li", "suffix": "" }, { "first": "Stan", "middle": [], "last": "Szpakowicz", "suffix": "" }, { "first": "Stan", "middle": [], "last": "Matwin", "suffix": "" } ], "year": 1995, "venue": "Proceedings of the 14th International Joint Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "1368--1374", "other_ids": {}, "num": null, "urls": [], "raw_text": "Li, Xiaobin, Stan Szpakowicz, and Stan Matwin. 1995. A WordNet-based algorithm for word sense disambiguation. In Proceedings of the 14th International Joint Conference on Artificial Intelligence, pages 1368-1374.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Dynamically combining syntax and semantics in natural language processing", "authors": [ { "first": "Steven", "middle": [ "L" ], "last": "Lytinen", "suffix": "" } ], "year": 1986, "venue": "Proceedings of AAAI-86", "volume": "", "issue": "", "pages": "574--578", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lytinen, Steven L. 1986. Dynamically combining syntax and semantics in natural language processing. In Proceedings of AAAI-86, pages 574-578.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "Mainichi shimbun CD-ROM '91-'94", "authors": [ { "first": "Mainichi", "middle": [], "last": "Shimbun", "suffix": "" } ], "year": 1991, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mainichi Shimbun. 1991-1994. Mainichi shimbun CD-ROM '91-'94 (in Japanese).", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "JUMAN Users Manual", "authors": [ { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" }, { "first": "Sadao", "middle": [], "last": "Kurohashi", "suffix": "" }, { "first": "Takehito", "middle": [], "last": "Utsuro", "suffix": "" }, { "first": "Yutaka", "middle": [], "last": "Myoki", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1993, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Matsumoto, Yuji, Sadao Kurohashi, Takehito Utsuro, Yutaka Myoki, and Makoto Nagao, 1993. JUMAN Users Manual (in Japanese). Kyoto University and Nara Institute of Science and Technology.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "Five papers on WordNet", "authors": [ { "first": "George", "middle": [ "A" ], "last": "Miller", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Beckwith", "suffix": "" }, { "first": "Christiane", "middle": [], "last": "Fellbaum", "suffix": "" }, { "first": "Derek", "middle": [], "last": "Gross", "suffix": "" }, { "first": "Katherine", "middle": [], "last": "Miller", "suffix": "" }, { "first": "Randee", "middle": [], "last": "Tengi", "suffix": "" } ], "year": 1993, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Miller, George A., Richard Beckwith, Christiane Fellbaum, Derek Gross, Katherine Miller, and Randee Tengi. 1993. Five papers on WordNet. Technical Report CLS-Rep-43, Cognitive Science Laboratory, Princeton University.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "Comparative experiments on disambiguating word senses: An illustration of the role of bias in machine learning", "authors": [ { "first": "Raymond", "middle": [ "J" ], "last": "Mooney", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "82--91", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mooney, Raymond J. 1996. Comparative experiments on disambiguating word senses: An illustration of the role of bias in machine learning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 82-91.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "A preferential constraint satisfaction technique for natural language analysis", "authors": [ { "first": "Katashi", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1994, "venue": "IEICE Transactions on Information and Systems", "volume": "", "issue": "2", "pages": "161--170", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nagao, Katashi. 1994. A preferential constraint satisfaction technique for natural language analysis. IEICE Transactions on Information and Systems, E77-D(2):161-170.", "links": null }, "BIBREF33": { "ref_id": "b33", "title": "A framework of a mechanical translation between Japanese and English by analogy principle. ArtiJicial and Human Intelligence", "authors": [ { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1984, "venue": "", "volume": "", "issue": "", "pages": "173--180", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nagao, Makoto. 1984. A framework of a mechanical translation between Japanese and English by analogy principle. ArtiJicial and Human Intelligence, pages 173-180.", "links": null }, "BIBREF34": { "ref_id": "b34", "title": "Discourse constraint in computer manuals", "authors": [ { "first": "Tetsuya", "middle": [], "last": "Nasukawa", "suffix": "" } ], "year": 1993, "venue": "Proceedings of the 5th International Conference on Theoretical and Methodological Issues in Machine Translation", "volume": "", "issue": "", "pages": "183--194", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nasukawa, Tetsuya. 1993. Discourse constraint in computer manuals. In Proceedings of the 5th International Conference on Theoretical and Methodological Issues in Machine Translation, pages 183-194.", "links": null }, "BIBREF35": { "ref_id": "b35", "title": "Shuei publisher", "authors": [], "year": 1964, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "National Language Research Institute. 1964. Bunruigoihyo (in Japanese). Shuei publisher.", "links": null }, "BIBREF36": { "ref_id": "b36", "title": "Exemplar-based word sense disambiguation: Some recent improvements", "authors": [ { "first": "Hwee", "middle": [], "last": "Ng", "suffix": "" }, { "first": "", "middle": [], "last": "Tou", "suffix": "" } ], "year": 1997, "venue": "Proceedings of the 2nd Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "208--213", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ng, Hwee Tou. 1997. Exemplar-based word sense disambiguation: Some recent improvements. In Proceedings of the 2nd Conference on Empirical Methods in Natural Language Processing, pages 208-213.", "links": null }, "BIBREF37": { "ref_id": "b37", "title": "Integrating multiple knowledge sources to disambiguate word sense: An exemplar-based approach", "authors": [ { "first": "Hwee", "middle": [], "last": "Ng", "suffix": "" }, { "first": "Hian Beng", "middle": [], "last": "Tou", "suffix": "" }, { "first": "", "middle": [], "last": "Lee", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 34th Annual Meeting", "volume": "", "issue": "", "pages": "40--47", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ng, Hwee Tou and Hian Beng Lee. 1996. Integrating multiple knowledge sources to disambiguate word sense: An exemplar-based approach. In Proceedings of the 34th Annual Meeting, pages 40-47, Association for Computational Linguistics.", "links": null }, "BIBREF38": { "ref_id": "b38", "title": "Co-occurrence vectors from corpora vs. distance vectors from dictionaries", "authors": [ { "first": "Yoshiki", "middle": [], "last": "Niwa", "suffix": "" }, { "first": "Yoshihiko", "middle": [], "last": "Nitta", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 15th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "304--309", "other_ids": {}, "num": null, "urls": [], "raw_text": "Niwa, Yoshiki and Yoshihiko Nitta. 1994. Co-occurrence vectors from corpora vs. distance vectors from dictionaries. In Proceedings of the 15th International Conference on Computational Linguistics, pages 304-309.", "links": null }, "BIBREF39": { "ref_id": "b39", "title": "Machine translation by case generalization", "authors": [ { "first": "Hiroshi", "middle": [], "last": "Nomiyama", "suffix": "" } ], "year": 1993, "venue": "", "volume": "34", "issue": "", "pages": "905--912", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nomiyama, Hiroshi. 1993. Machine translation by case generalization (in Japanese). Transactions of Information Processing Society of Japan, 34(5):905-912.", "links": null }, "BIBREF40": { "ref_id": "b40", "title": "Sequential model selection for word sense disambiguation", "authors": [ { "first": "Ted", "middle": [], "last": "Pedersen", "suffix": "" }, { "first": "Rebecca", "middle": [], "last": "Bruce", "suffix": "" }, { "first": "Janyce", "middle": [], "last": "Wiebe", "suffix": "" } ], "year": 1997, "venue": "Proceedings of the 5th Conference on Applied Natural Language Processing", "volume": "", "issue": "", "pages": "388--395", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pedersen, Ted, Rebecca Bruce, and Janyce Wiebe. 1997. Sequential model selection for word sense disambiguation. In Proceedings of the 5th Conference on Applied Natural Language Processing, pages 388-395.", "links": null }, "BIBREF41": { "ref_id": "b41", "title": "Lexical knowledge representation and natural language processing", "authors": [ { "first": "James", "middle": [], "last": "Pustejovsky", "suffix": "" }, { "first": "Branimir", "middle": [], "last": "Boguraev", "suffix": "" } ], "year": 1993, "venue": "Artificial Intelligence", "volume": "63", "issue": "1-2", "pages": "193--223", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pustejovsky, James and Branimir Boguraev. 1993. Lexical knowledge representation and natural language processing. Artificial Intelligence, 63(1-2):193-223.", "links": null }, "BIBREF42": { "ref_id": "b42", "title": "Real World Computing Partnership", "authors": [], "year": 1995, "venue": "RWC text database", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Real World Computing Partnership. 1995. RWC text database (in Japanese).", "links": null }, "BIBREF43": { "ref_id": "b43", "title": "Selection and Information: A Class-Based Approach to Lexical Relationships", "authors": [ { "first": "Philip", "middle": [], "last": "Resnik", "suffix": "" } ], "year": 1993, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Resnik, Philip. 1993. Selection and Information: A Class-Based Approach to Lexical Relationships. Ph.D. thesis, University of Pennsylvania.", "links": null }, "BIBREF44": { "ref_id": "b44", "title": "On learning more appropriate selectional restrictions", "authors": [ { "first": "Francesc", "middle": [], "last": "Ribas", "suffix": "" } ], "year": 1995, "venue": "Proceedings of the 7th Conference of the European Chapter of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "112--118", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ribas, Francesc. 1995. On learning more appropriate selectional restrictions. In Proceedings of the 7th Conference of the European Chapter of the Association for Computational Linguistics, pages 112-118.", "links": null }, "BIBREF45": { "ref_id": "b45", "title": "Introduction to Modern Information Retrieval", "authors": [ { "first": "Gerard", "middle": [], "last": "Salton", "suffix": "" }, { "first": "Michael", "middle": [ "J" ], "last": "Mcgill", "suffix": "" } ], "year": 1983, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Salton, Gerard and Michael J. McGill. 1983. Introduction to Modern Information Retrieval. McGraw-Hill.", "links": null }, "BIBREF46": { "ref_id": "b46", "title": "Dimensions of meaning", "authors": [ { "first": "Hinrich", "middle": [], "last": "Schiitze", "suffix": "" } ], "year": 1992, "venue": "Proceedings of Supercomputing", "volume": "", "issue": "", "pages": "787--796", "other_ids": {}, "num": null, "urls": [], "raw_text": "Schiitze, Hinrich. 1992. Dimensions of meaning. In Proceedings of Supercomputing, pages 787-796.", "links": null }, "BIBREF47": { "ref_id": "b47", "title": "Query by committee", "authors": [ { "first": "H", "middle": [ "S" ], "last": "Seung", "suffix": "" }, { "first": "M", "middle": [], "last": "Opper", "suffix": "" }, { "first": "H", "middle": [], "last": "Sompolinsky", "suffix": "" } ], "year": 1992, "venue": "Proceedings of the 5th Annual ACM Workshop on Computational Learning Theory", "volume": "", "issue": "", "pages": "287--294", "other_ids": {}, "num": null, "urls": [], "raw_text": "Seung, H. S., M. Opper, and H. Sompolinsky. 1992. Query by committee. In Proceedings of the 5th Annual ACM Workshop on Computational Learning Theory, pages 287-294.", "links": null }, "BIBREF48": { "ref_id": "b48", "title": "Remembering to forget: A competence-preserving case deletion policy for case-based reasoning systems", "authors": [ { "first": "Barry", "middle": [], "last": "Smyth", "suffix": "" }, { "first": "Mark", "middle": [ "T" ], "last": "Keane", "suffix": "" } ], "year": 1995, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Smyth, Barry and Mark T. Keane. 1995. Remembering to forget: A competence-preserving case deletion policy for case-based reasoning systems.", "links": null }, "BIBREF49": { "ref_id": "b49", "title": "Proceedings of the 14th International Joint Conference on Artificial Intelligence", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "377--382", "other_ids": {}, "num": null, "urls": [], "raw_text": "In Proceedings of the 14th International Joint Conference on Artificial Intelligence, pages 377-382.", "links": null }, "BIBREF50": { "ref_id": "b50", "title": "Automatic thesaurus construction based on grammatical relations", "authors": [ { "first": "Takenobu", "middle": [], "last": "Tokunaga", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Iwayama", "suffix": "" }, { "first": "Hozumi", "middle": [], "last": "Tanaka", "suffix": "" } ], "year": 1995, "venue": "Proceedings of the 14th International Joint Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "1308--1313", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tokunaga, Takenobu, Makoto Iwayama, and Hozumi Tanaka. 1995. Automatic thesaurus construction based on grammatical relations. In Proceedings of the 14th International Joint Conference on Artificial Intelligence, pages 1308-1313.", "links": null }, "BIBREF51": { "ref_id": "b51", "title": "A best-match algorithm for broad-coverage example-based disambiguation", "authors": [ { "first": "Naohiko", "middle": [], "last": "Uramoto", "suffix": "" } ], "year": 1994, "venue": "hi Proceedings of the 15th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "717--721", "other_ids": {}, "num": null, "urls": [], "raw_text": "Uramoto, Naohiko. 1994a. A best-match algorithm for broad-coverage example-based disambiguation, hi Proceedings of the 15th International Conference on Computational Linguistics, pages 717-721.", "links": null }, "BIBREF52": { "ref_id": "b52", "title": "Example-based word-sense disambiguation", "authors": [ { "first": "Naohiko", "middle": [], "last": "Uramoto", "suffix": "" } ], "year": 1994, "venue": "IEICE Transactions on Information and Systems", "volume": "", "issue": "2", "pages": "240--246", "other_ids": {}, "num": null, "urls": [], "raw_text": "Uramoto, Naohiko. 1994b. Example-based word-sense disambiguation. IEICE Transactions on Information and Systems, E77-D(2):240-246.", "links": null }, "BIBREF53": { "ref_id": "b53", "title": "Sense classification of verbal polysemy based on bilingual class/class association", "authors": [ { "first": "Takehito", "middle": [], "last": "Utsuro", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 16th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "968--973", "other_ids": {}, "num": null, "urls": [], "raw_text": "Utsuro, Takehito. 1996. Sense classification of verbal polysemy based on bilingual class/class association. In Proceedings of the 16th International Conference on Computational Linguistics, pages 968-973.", "links": null }, "BIBREF54": { "ref_id": "b54", "title": "Thesaurus-based efficient example retrieval by generating retrieval queries from similarities", "authors": [ { "first": "Takehito", "middle": [], "last": "Utsuro", "suffix": "" }, { "first": "Kiyotaka", "middle": [], "last": "Uchimoto", "suffix": "" }, { "first": "Mitsutaka", "middle": [], "last": "Matsumoto", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Nagao", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 15th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "1044--1048", "other_ids": {}, "num": null, "urls": [], "raw_text": "Utsuro, Takehito, Kiyotaka Uchimoto, Mitsutaka Matsumoto, and Makoto Nagao. 1994. Thesaurus-based efficient example retrieval by generating retrieval queries from similarities. In Proceedings of the 15th International Conference on Computational Linguistics, pages 1044-1048.", "links": null }, "BIBREF55": { "ref_id": "b55", "title": "Using WordNet to disambiguate word senses for text retrieval", "authors": [ { "first": "Ellen", "middle": [ "M" ], "last": "Voorhees", "suffix": "" } ], "year": 1993, "venue": "Proceedings of the 16th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "171--180", "other_ids": {}, "num": null, "urls": [], "raw_text": "Voorhees, Ellen M. 1993. Using WordNet to disambiguate word senses for text retrieval. In Proceedings of the 16th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 171-180.", "links": null }, "BIBREF56": { "ref_id": "b56", "title": "Word-sense disambiguation using statistical models of Roget's categories trained on large corpora", "authors": [ { "first": "David", "middle": [], "last": "Yarowsky", "suffix": "" } ], "year": 1992, "venue": "Proceedings of the 14th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "454--460", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yarowsky, David. 1992. Word-sense disambiguation using statistical models of Roget's categories trained on large corpora. In Proceedings of the 14th International Conference on Computational Linguistics, pages 454-460.", "links": null }, "BIBREF57": { "ref_id": "b57", "title": "Unsupervised word sense disambiguation rivaling supervised methods", "authors": [ { "first": "David", "middle": [], "last": "Yarowsky", "suffix": "" } ], "year": 1995, "venue": "Proceedings of the 33rd Annual Meeting", "volume": "", "issue": "", "pages": "189--196", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yarowsky, David. 1995. Unsupervised word sense disambiguation rivaling supervised methods. In Proceedings of the 33rd Annual Meeting, pages 189-196, Association for Computational Linguistics.", "links": null }, "BIBREF58": { "ref_id": "b58", "title": "Lexicon acquisition: Learning from corpus by capitalizing on lexical categories", "authors": [ { "first": "Uri", "middle": [], "last": "Zernik", "suffix": "" } ], "year": 1989, "venue": "Proceedings of the 11th International Joint Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "1556--1562", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zernik, Uri. 1989. Lexicon acquisition: Learning from corpus by capitalizing on lexical categories. In Proceedings of the 11th International Joint Conference on Artificial Intelligence, pages 1556-1562.", "links": null } }, "ref_entries": { "FIGREF0": { "num": null, "type_str": "figure", "uris": null, "text": ") = ~,7~ ~ ~ Igs''\u00a2l + ]\u00a3s,,cl 2[\u00a3s~,c r\"l \u00a3s,,~l j=i+l I&,,I 7 I'G*I )" }, "FIGREF1": { "num": null, "type_str": "figure", "uris": null, "text": "). 7 arg msax P(slx) P(s) . P(xls) = arg msax P(x) = argn~axP(s). P(xls) argmaxP(s) II P(ncls) c (8)" }, "FIGREF4": { "num": null, "type_str": "figure", "uris": null, "text": "Figure 9" }, "FIGREF5": { "num": null, "type_str": "figure", "uris": null, "text": "Sampling. The procedure for uncertainty sampling(Lewis and Gale 1994) is as follows, where C(x) represents the interpretation certainty for an examplex (see our sampling procedure in Section 3.1 for comparison):" }, "TABREF1": { "html": null, "content": "", "text": "", "num": null, "type_str": "table" } } } }