Resolving Direct and Indirect Anaphora for Japanese Definite

Resolving Direct and Indirect Anaphora for Japanese Deﬁnite Noun Phrases

Naoya Inoue Ryu Iida Kentaro Inui Yuji Matsumoto NAIST, Japan Tokyo Institute of Technology, Japan NAIST, Japan NAIST, Japan [email protected] [email protected] [email protected] [email protected]

Abstract An anaphoric relation can be either direct in parallel and classify the anaphora type. How- or indirect. In some cases, the antecedent ever, these issues have received little attention in being referred to lies outside the discourse the literature. to which its anaphor belongs. Therefore, In this paper, we regard the task of nominal an anaphora resolution model needs to con- anaphora resolution as a combination of two dis- sider the following two decisions in paral- tinct but arguably interdependent subtasks: lel: (i) antecedent selection–selecting the • Antecedent selection: the task of identifying the antecedent itself, and (ii) anaphora type antecedent of a given input definite NP, and classification–classifying exophora, direct • Anaphora type classification: the task of judg- anaphora, or indirect anaphora. However, ing whether the anaphor is direct anaphoric, in- the anaphora type classification has re- direct anaphoric or exophoric. ceived little attention in the literature. In Given this task decomposition, the following unex- this paper, we address this issue taking plored issues immediately arise: Japanese as our target language. Our find- • Whether the model for antecedent selection ings were: first, an antecedent selection should be designed and trained separately for model should be trained separately for each direct anaphora and indirect anaphora or if it anaphora type. Second, the best candidate can be trained as a single common model, antecedent selected by an antecedent selec- • How the two subtasks can be best combined tion model provides contextual information (e.g. which subtask should be carried out first), useful for anaphora type classification. In and consequence, antecedent selection should • Which context information is useful for be carried out before anaphora type clas- anaphora type classification. sification. In this paper, we explore these issues taking 1 Introduction Japanese as our target language. Anaphora resolution has been studied inten- Definite NPs in Japanese behave similarly to sively in recent years because of its significance those in English; that is, a definite NP may bear in many NLP applications such as information a direct anaphoric relation but may also bear an in- extraction and machine translation. In nominal direct anaphoric relation to its antecedent as shown anaphora resolution, an anaphor (typically a defi- in examples (4) and (5). nite noun phrase) and its antecedent hold either a (4) 新型車が発売された。このミニバンは燃費が非常に direct anaphoric relation (e.g., coreference) or an よい。 A new car was released. The minivan has good gas indirect relation (e.g., bridging reference (Clark, mileage. 1977)). In example (1), the anaphor the CD refers (5) 家具屋で机を見た。そのデザインは素晴らしかった。 to Her new album directly. I saw a desk in a furniture shop. The design was mar- (1) Her new album was released yesterday. velous. I want to get the CD as soon as possible. In contrast, the CD refers to her new song indi- “このミニバン (the minivan)” refers to “新型車 rectly in example (2). (a new car)” directly in (4), while “そのデザイン机 (2) The artist announced her new song. (the design)” refers to “ (a desk)” indirectly in I want to get the CD as soon as possible. (5). In Japanese, however, the problem can be even In cases such as example (3), an antecedent re- more complex because a definite NP is not always ferred to is not in the same discourse. marked by a definiteness modifier. In an example (3) I want to get the CD as soon as possible. (6), a bare NP 大統領 (literally, President) refers to 韓国大統領 We call the reference in example (1) direct (Korean President). anaphora; in example (2) indirect anaphora; and (6) 今月 4 日、韓国大統領が来日した。大統領は翌日の記者会見で新プランの詳細を語った。 in example (3) exophora. In anaphora resolution, Korean President visited Japan on the 4th this month. besides identifying the antecedent itself, it is there- (The) President talked about the details of his new plan fore necessary to consider these three possibilities at the news conference next day. Anaphora Type Classifier Antecedent Selection Model Anaphor Anaphor Anaphor Anaphor Anaphor Candidates Information Anaphor Direct Indirect Direct Indirect Anaphora Anaphora Anaphora Anaphora Single

Direct Indirect Direct Indirect Anaphora Anaphora Anaphora Anaphora Direct Indirect Direct Indirect Direct Indirect Direct Indirect Anaphora Anaphora Anaphora Anaphora Anaphora Anaphora Direct Indirect Direct Indirect Anaphora Anaphora Anaphora Anaphora Anaphora Anaphora Indirect Direct Exophora Exophora Anaphora Anaphora Exophora Exophora Exophora Exophora a-Classify-then-Select c-Classify-then-Select s-Select-then-Classify d-Select-then-Classify i-Select-then-Classify p-Select-then-Classify (aC/S) (cC/S) (sS/C) (dS/C) (iS/C) (pS/C) Figure 1: Anaphora resolution models Furthermore, it is sometimes difficult even for hu- and the antecedent to resolve direct anaphora. In man annotators to determine the definiteness of a example (1), the resolution model has to know bare NP. Therefore, our study for now deals with that CD and album are synonymous. For indirect the resolution of noun phrases modified by a defi- anaphora, on the other hand, the model is required niteness modifier, i.e., この (this) + NP, その (the) to recognize such semantic relations as part-whole + NP and あの (that) + NP. This problem alone is and attribute-of as in example (2), where knowing large enough, because noun phrases modified by a that CD is an object related to song is crucial. This definiteness modifier account for a large proportion expectation is supported empirically by our exper- of occurrences of nominal anaphora in Japanese iments as reported in section 5. texts. 2.2 Anaphora type classification This paper is organized as follows. In the next For anaphora type classification, an interesting section, we give a detailed explanation of our in- question is whether this subtask should be carried vestigations. In section 3, we introduce related out before antecedent selection or after. Accord- work. In section 4, the experimental setup is de- ingly, we consider two kinds of configurations: scribed. In section 5, the results of our investiga- Classify-then-Select and Select-then-Classify as tions are shown and discussed. Finally, the conclu- summarized in Figure 1. sion is given and our future work is described. 2.2.1 Classify-then-Select (C/S) model 2 Anaphora Resolution Models Given an anaphor, this model first determines To explore how our two subtasks can be best whether the anaphor bears direct anaphora, indirect configured, we consider several possible models anaphora or exophora. If the anaphora type is clas- for each subtask. sified as direct, then the model calls for the direct 2.1 Antecedent selection antecedent selection model. If classified as indi- For antecedent selection, we can employ a va- rect, on the other hand, then the model calls for the riety of existing machine learning-based methods indirect antecedent selection model. No antecedent designed for coreference resolution ranging from selection model is called for if the anaphora type classification-based models (Soon et al., 2001, is classified as exophora. The model for anaphora etc.) and preference-based models (Ng and Cardie, type classification is trained in a supervised fash- 2001, etc.) to comparison-based models (Iida ion. et al., 2005; Yang et al., 2003, etc.). In this pa- According to the information used for the classi- per, we adopt Iida et al.’s tournament model, a fication, we can consider the following two config- state-of-the-art method for coreference resolution urations: in Japanese (Iida et al., 2005). • a-Classify-then-Select (aC/S) Model: Classify One issue to explore for this subtask is whether the anaphora type by using the anaphor and its a single common model should be trained for both properties before selecting the antecedent. direct and indirect anaphora or whether a separate • c-Classify-then-Select (cC/S) Model: Classify model should be trained for each. So we consider the anaphora type using the anaphor, its prop- the following two approaches: erties and the information from all potential an- • Single model: Train a single model with labeled tecedents before selecting the antecedent. examples of both direct and indirect anaphora. To verify the effects of using additional informa- • Separate model: Train two distinct models sep- tion (i.e., contextual information) we can choose arately for direct and indirect cases; i.e., a direct the aC/S model as a baseline. We give a detail of antecedent selection model is trained only with the feature set for these models in Section 4.2. labeled examples of direct anaphora and an in- 2.2.2 Select-then-Classify (S/C) model direct antecedent selection model only with la- Given an anaphor, this model first selects the beled examples of indirect anaphora. most likely antecedent and then calls for the The separate model approach is expected to be anaphora type classifier to determine the anaphora advantageous because the semantic relation be- type referring to both the anaphor and the selected tween an anaphor and its antecedent tends to be candidate antecedent(s). This way of configura- different in direct and indirect cases. It is crucial to tion has an advantage over the Classify-then-Select identify synonymous relations between an anaphor models in that it determines the anaphora type of a given anaphor taking into account the information The case of the dS/C configuration is slightly of its most likely candidate antecedent. The can- more complex. Analogous to the sS/C model, for didate antecedent selected in the first step can be (7), the classifier takes the CDi paired with its expected to provide contextual information useful direct antecedent albumi as an instance of direct for anaphora type classification: for example, if her anaphora. For (8), however, since the true indirect new song is selected as the candidate antecedent in song is unlikely to be selected as the best candi- example (2) in section 1, the anaphora type will date by the direct anaphora model, we run the di- be easily identified based on the lexical knowledge rect anaphora model to select the pseudo-best can- that CD is potentially an object related to song. didate. Suppose artist is selected; then, we create a Since we have two options for antecedent selec- training instance of indirect anaphora from the the tion, the single and separate models, as described CD j paired with artist. in 2.1, we can consider at least the following four An analogous method applies also to the iS/C configurations (see Figure 1): configuration. Assuming that yesterday is selected • s-Select-then-Classify (sS/C) Model: Select as the pseudo-best candidate indirect antecedent, the best candidate antecedent with the Single we label the CDi paired with yesterday as direct model and then classify the anaphora type. anaphora and the CD j paired with artist as indirect • d-Select-then-Classify (dS/C) Model: Select anaphora. the best candidate antecedent by using the di- Finally, for the pS/C configuration, we create a training instance of direct anaphora from hthe CDi, rect anaphora model and then determine the i anaphora type. If the candidate is classified albumi, yesterday and a training instance of indi- h 0 i as indirect anaphora, search for the antecedent rect anaphora from the CD j, artist, song j . with the indirect anaphora model. 3 Related Work • i-Select-then-Classify (iS/C) Model: Reverse Anaphora resolution is an important process the steps in the d-Select-then-Classify (dS/C) for various NLP applications. In contrast with model. rule-based approaches, such as (Brennan et al., • p-Select-then-Classify (pS/C) Model: Call for 1987; Lappin and Leass, 1994; Baldwin, 1995; the direct anaphora and indirect anaphora mod- Nakaiwa and Shirai, 1996; Okumura and Tamura, els in parallel to select the best candidate an- 1996; Mitkov, 1997, etc.), empirical, or machine tecedent for each case and then determine the learning-based approaches to this problem have anaphora type referring to both candidates. shown to be a cost-efficient solution achieving per- The pS/C configuration provides the anaphora type formance that is comparable to the best performing classifier with richer contextual information than rule-based systems (McCarthy and Lehnert, 1995; any other configuration while it requires more com- N.Ge et al., 1998; Soon et al., 2001; Ng and Cardie, putational cost because it always searches for the 2001; Strube and Muller, 2003; Iida et al., 2005; best candidate, simultaneously searching for direct Yang et al., 2003, etc.). Most of these studies are anaphora and indirect anaphora. focused only on the coreference resolution task, The training of the classifier depends on which particularly in the context of evaluation-oriented configuration is chosen. Assume that we have the research programs such as Message Understanding following examples in our training set: Conference (MUC)1 and Automatic Content Ex- 2 (7) Her new albumi was released yesterday. traction (ACE) , leaving the issues regarding indi- I want to get the CDi as soon as possible. rect anaphora relatively unexplored. - anaphor: the CDi As mentioned in Section 1, there has been little - antecedent: albumi - anaphora type: direct anaphora attention paid to the issue of anaphora type classi- (8) The artist announced her new song j0 . fication. An exception can be seen in Vieira and I want to get the CD j as soon as possible. Poesio (2000)’s work, which proposes a method - anaphor: the CD j of anaphora resolution for definite noun phrases. - antecedent: song j Their model has two folds. In the first phase, it - anaphora type: indirect anaphora classifies a given input definite NP into one of three (9) I want to get the CD j as soon as possible. categories, which (very loosely) correspond to our - anaphor: the CD j - antecedent: None anaphora types: direct anaphora (reference to an - anaphora type: exophora antecedent with the same head noun), bridging de- Given these annotated examples, sS/C configura- scription (reference to the same discourse entity already introduced in the discourse using a different tion simply takes the CDi paired with albumi from (7) as an instance of direct anaphora and the CD head noun or reference to an antecedent which is j a distinct but related object of the anaphor), and paired with song j0 from (8) as an instance of indirect anaphora. For (9), all the configurations discourse-new (reference to a discourse entity not including sS/C take the CD and run its related already introduced in the discourse). This classi- antecedent selection model to select pseudo an- fication is done by a set of hand-coded rules. In tecedent. The antecedent is paired with the CD as 1http://www-nlpir.nist.gov/related projects/muc/index.html an instance of exophora. 2http://www.nist.gov/speech/tests/ace/ the second phase, the antecedent is selected from a Table 1: Distribution of anaphoric relations in the set of potential candidates appearing in the preced- annotated corpus ing discourse in a fashion depending on the cate- Anaphora Type Noun Predicate Overall gory determined in the first phase. So Vieira and Direct anaphora 530 70 600 Poesio’s model carries out anaphora type classifi- Indirect anaphora 466 435 901 Ambiguous 0 8 8 cation before antecedent selection. In this paper, Exophora - - 248 however, we show through empirical evaluations Overall 996 513 1,757 that this way of configuring the two subtasks is not ‘Noun’ and ‘Predicate’ denote the syntactic category of an necessarily the best choice. antecedent. ‘Ambiguous’ was annotated to an anaphor which holds both direct and indirect anaphoric relations. In our 4 Evaluation evaluations, we discarded ‘Ambiguous’ instances. Our evaluation consists of three steps. First, in and 248 instances of exophora. order to find out whether an antecedent selection 4.2 Feature Set model should be designed and trained separately The feature set for antecedent selection was de- for direct anaphora and indirect anaphora, we eval- signed based on the literature on coreference reso- uate two antecedent selection models, the single lution (Iida et al., 2005; Ng and Cardie, 2001; Soon model and separate model described in Section 2.1. et al., 2001; Denis and Baldridge, 2008; Yang et al., Second, the anaphora type classification models 2003, etc). Our feature set captures lexical charac- described in Section 2.2 are evaluated to explore ters of compared candidates and an anaphor, se- what information helps the anaphora type classifi- mantic agreements between them, and so on. See cation. Finally, we evaluate the overall accuracy of (Iida et al., 2005) in further detail. the entire anaphora resolution task to explore how • the models can best be configured. The evaluation SYNONYM OF and HYPONYM OF: The is carried out by 10-fold cross-validation.3 synonymous and hyponymous relationships be- For the three-way classification for anaphora tween an anaphor and a candidate antecedent type classification, we adopt one-versus-rest were identified by the Japanese WordNet (Isa- method in our experiments. For creating bi- hara et al., 2008), the Bunrui Goi Hyo thesaurus nary classifiers used in antecedent selection and (NIJL, 2004), and a very large hypernymy hi- anaphora type classification, we adopt Support erarchy (about three million hypernymy rela- 4 tions) automatically created from Web texts Vector Machines (Vapnik, 1995) , with a polyno- and Wikipedia (Sumida et al., 2008). mial kernel of degree 2 and its default parameters. • NOUN-NOUN SIMILARITY: The distribu- 4.1 Data tional similarity between an anaphor and its For training and testing our models, we created candidate antecedent was calculated from a the annotated corpus using the NAIST Text Cor- cooccurrence matrix of (n,hc,vi), where n is pus (Iida et al., 2007), which is publicly available a noun phrase appearing in an argument posi- from the Web site of the NAIST NLP group. The tion of a verb v marked by a case particle c. NAIST Text Corpus also contains anaphoric rela- The cooccurrences were counted using twenty tions of noun phrases, but they are strictly lim- years of newspaper articles and their distribu- ited to coreference relations (i.e. two NP must re- tion P(n,hc,vi) was estimated by pLSI (Hof- fer to the same entity in the world). For this rea- mann, 1999) with 1,000 hidden topic classes. son, we re-annotated (i) direct anaphoric relations, • ASSOCIATIVENESS: To resolve indirect (ii) indirect anaphoric relations and (iii) exophoric anaphora, the degree of associativeness be- NPs marked by three definiteness modifiers, i.e., tween an anaphor ANA and candidate CND was このそのあの calculated differently depending on whether (this), (the) and (that). In the CND is a noun or predicate. In the case of specification of our corpus, not only noun phrases a noun, the associativeness is calculated from but verb phrases too are chosen as antecedents. the cooccurrences of ANA and CND in the pat- Consequently, we obtained 600 instances of direct tern “CND の ANA (ANA of CND)”, follow- anaphora, 901 instances of indirect anaphora, and ing the literature of bridging reference resolu- 248 instances of exophora. The detailed distribution (Poesio et al., 2004, etc.). Cooccurrence tion is shown in Table 1. counts were obtained from Web Japanese N- In our experiments, we used anaphors whose an- gram Version 1 (Kudo and Kazawa, 2007). In tecedent is the head of an NP which appears in the the case of a predicate, on the other hand, the preceding context of the anaphor (i.e., cataphora is associativeness is calculated from the cooccur- ignored). Therefore, we used 572 instances of di- rences of ANA and CND in the pattern where rect anaphora, 878 instances of indirect anaphora CND syntactically depends on (i.e., modifies) ANA (in English, the pattern like “ANA that 3In our evaluation of antecedent selection, we also manu- (subj) CND”). If we find many occurrences of, ally checked for cases in which a selected antecedent is not for example, “闘う (to fight)” modifying “夢 labeled in our corpus but can be evaluated as correct (e.g., as when a discourse entity located in the same anaphoric chain (a dream)” in a corpus, then “夢 (a dream)” is as the labeled antecedent is selected instead). likely to refer to an event referred to by “闘う 4SVMlight http://svmlight.joachims.org/ (to fight)” as in (10). (10) チャンピオンと闘いたい。その夢は実現すると信 Table 2: Results of antecedent selection じている。 Anaphora Type Single Model Separate Model I want to fighti the champion. I’m sure that Direct anaphora 63.3% (362/572) 65.4% (374/572) that dreami will come true. Indirect anaphora 50.5% (443/878) 53.2% (467/878) Overall 55.2% (801/1,450) 58.0% (841/1,450) For anaphora type classification, we use a different feature set depending on the configuration de- 70% Direct anaphora model scribed in 2.2. For the Classify-then-Select config- Indirect anaphora model urations, we designed our feature set based on the 65% literature (Poesio et al., 2004, etc.). Our feature set 60% captures contextual information and lexical syntactic properties (i.e., NP head, POS, case particle Accuracy 55% and type of definiteness modifier) of an anaphor. 50% The contextual information is encoded by using (i) part-of-speech of the candidate antecedents that the 45% 3.125% 6.25% 12.50% 25% 50% 100% context holds, (ii) the information on whether the Training data size context has candidate antecedent(s) with the same Figure 2: Learning curve for Separate Models head as a head of the anaphor or whether the candidate antecedent(s) is a hyponym or synonym of an of “Frankenstein”i. I think the moviei is indeed a great anaphor, and (iii) maximum noun-noun similarity masterpiece. and associativeness (see above) for an anaphor and where “映画 j (movies j)” was wrongly selected as 6 each candidate in the context. the antecedent of “この映画 i (the moviei)”. As For the Select-then-Classify configurations, on can be imagined from this example, there is still the other hand, the anaphora type classifier uses the room for improvement by carefully taking into ac- information of the best candidate(s) selected in the count this kind of error using other clues such as first step. This sort of information is encoded as information from salience. For indirect anaphora, a feature set analogous to the feature set for an- we conducted an evaluation where the ASSOCIA- tecedent selection. TIVENESS feature set was disabled in order to evaluate our lexical resource as described in sec- 5 Results tion 4.2. The results of this additional evaluation 5.1 Antecedent selection showed that the model obtained 51.4% (451/878) The results are shown in Table 2, which indi- accuracy, which is no significant difference com- cates that the selection models should be designed pared with the original accuracy. We need to find for each anaphora type separately.5 We therefore more, and more useful, clues to capture the asso- discarded the Single Model for the subsequent ex- ciativeness between an anaphor and the related ob- periments. ject in indirect anaphora. We also illustrate the learning curves of each model in Figure 2. Reducing the training data to 5.2 Anaphora type classification 50%, 25%, 12.5%, 6.25% and 3.13%, we con- The results of anaphora type classification are ducted the evaluation over three random trials shown in Table 3. The aC/S model obtained an for each size and averaged the accuracies. The accuracy of 75.4%, better than the cC/S model curves indicate that in the direct antecedent selec- (73.6%), which indicates that contextual information model the accuracy improves as the training tion proposed in the literature (Poesio et al., 2004, data increase, whereas the increase in the accuracy etc.) was not actually informative. The dS/C model of the indirect antecedent selection model is min- achieved the best accuracy (78.7%) and the im- imal, even though our data set included more in- provement indicates that the selected best candi- stances for indirect than direct anaphora. These re- date antecedent provides useful contextual infor- sults support the findings in research showing that mation for anaphora type classification. indirect anaphora is harder to resolve than direct As shown in Table 3, identifying exophoric anaphora and suggesting that we need a more so- cases tends to be more difficult than identifying phisticated antecedent selection model for indirect anaphoric cases. This difference largely arises anaphora. from poor performance of our baseline model Our error analysis revealed that a majority (about (aC/S). While our aC/S for indirect anaphora 60%) of errors in direct anaphora were caused by achieved an F-measure of 83.7%, the aC/S model a confusion between candidates belonging to the for exophora achieved only an F-measure of same semantic category. Here is a typical example 48.9%. Indirect anaphors can sometimes be iden- of this sort of error: tified even only by their head words, such as 結 (11) 私は映画 j の知識がないが、『フランケンシュタイン』i 果 (result) and 他 (other). However, our anal- ぐらいは知っている。この映画i は、本当に名作だ。 ysis could not find such good indicators for ex- I don’t have good knowledge of movies j but still know ophora. Nevertheless, the pS/C model successfully 5The difference between the Single Model and the Sep- arate Model have statistical significant in the McNemar test 6In Japanese, in general there is no singular/plural distinc- (p < 0.01). tion in morphological form. Table 3: Results of anaphora type classification Direct Anaphora Indirect Anaphora Exophora Accuracy Model P R F P R F P R F ATC Overall aC/S 67.7% 74.5% 70.9% 80.6% 87.1% 83.7% 75.0% 36.3% 48.9% 75.4% 47.3% cC/S 69.4% 73.4% 71.4% 74.9% 87.5% 80.7% 92.5% 25.0% 39.4% 73.6% 46.3% dS/C 70.9% 84.6% 77.1% 83.2% 85.6% 84.4% 90.1% 40.3% 55.7% 78.7% 50.6% iS/C 67.7% 74.8% 71.1% 78.1% 88.3% 82.9% 93.2% 27.8% 42.9% 74.9% 46.3% pS/C 71.2% 82.0% 76.1% 82.1% 86.7% 84.3% 91.9% 41.1% 57.2% 78.4% 50.4% ATC in the accuracy column indicates the accuracy of anaphora type classification. improved its performance by using the information H. H. Clark. 1977. Bridging. In Thinking: Readings in Cog- of selected candidate antecedents. nitive Science. P. Denis and J. Baldridge. 2008. Specialized models and rank- 5.3 Overall anaphora resolution ing for coreference resolution. In Proc. of the 2008 Con- Finally, the overall accuracy of the entire ference on EMNLP, pages 660–669. anaphora resolution task was evaluated by: National Institute for Japanese Language, editor. 2004. Bunrui Goi Hyo (in Japanese). Shuei Shuppan. |C ∩C | T. Hofmann. 1999. Probabilistic latent semantic indexing. In Accuracy = antecedent anaphora type , SIGIR, pages 50–57. # of all instances R. Iida, K. Inui, Y. Matsumoto, and S. Sekine. 2005. Noun phrase coreference resolution in Japanese based on most where C is the set of instances whose an- likely candidate antecedents noun phrase coreference res- antecedent olution in Japanese based on most likely antecedent can- tecedent is correctly selected and Canaphora type is didates. In Journal of Information Processing Society of the set of instances whose anaphora type is cor- Japan, pages 831–844. rectly identified. The results are shown in Table 3. R. Iida, M. Komachi, K. Inui, and Y. Matsumoto. 2007. Anno- Again, the dS/C model achieved the best accu- tating a Japanese text corpus with predicate-argument and racy (50.6%), which is significantly better than the coreference relations. In ACL Workshop on ‘Linguistic An- notation Workshop’, pages 132–139. Classify-then-Select models. H. Isahara, F. Bond, K. Uchimoto, M. Utiyama, and K. Kan- 6 Conclusion zaki. 2008. Development of the Japanese wordnet. In LREC. We have addressed the three issues of nominal T. Kudo and H. Kazawa. 2007. Web Japanese N-gram Version anaphora resolution for Japanese NPs marked by 1. Gengo Shigen Kyokai. a definiteness modifier under two subtasks, i.e., S. Lappin and H. J. Leass. 1994. An algorithm for pronom- antecedent selection and anaphora type classifica- inal anaphora resolution. Computational Linguistics., tion. The issues we addressed were: (i) how the 20(4):535–561. J. F. McCarthy and W. G. Lehnert. 1995. Using decision trees antecedent selection model should be designed, (ii) for coreference resolution. In IJCAI, pages 1050–1555. how the antecedent selection and anaphora type R. Mitkov. 1997. Factors in anaphora resolution: they are classification should be carried out, (iii) what infor- not the only things that matter. a case study based on two mation helps anaphora type classification. Our em- different approaches. In ACL and EACL Workshop. pirical evaluations showed that the separate model H. Nakaiwa and S. Shirai. 1996. Anaphora resolution of achieved better accuracy than the single model Japanese zero pronouns with deictic reference. In COL- and that the d-Select-then-Classify model gave the ING, pages 812–817. best results. We have made two findings through V. Ng and C. Cardie. 2001. Improving machine learning approaches to coreference resolution. In ACL, pages 104– the evaluations: (i) an antecedent selection model 111. should be designed separately for each anaphora N.Ge, J.Hale, and E.Charniak. 1998. A statistical approach to type using the information useful for identify- anaphora resolution. In The 6th Workshop on Very Large ing its antecedent separately, and (ii) the candi- Corpora, pages 161–170. date antecedent selected by an antecedent selection M. Okumura and K. Tamura. 1996. Zero pronoun resolution in Japmaanese discourse based on centering theory. In COL- model provides contextual information useful for ING, pages 871–876. anaphora type classification. In consequence, we M. Poesio, R. Mehta, A. Maroudas, and J. Hitzeman. 2004. concluded that antecedent selection should be car- Learning to resolve bridging references. In ACL, pages ried out before anaphora type classification. 144–151. However, there is still considerable room for im- W. M. Soon, H. T. Ng, and C. Y. Lim. 2001. A machine learning approach to coreference resolution of noun phrases. provement in each of two subtasks. Our error anal- Computational Linguistics, 27(4):521–544. ysis has suggested that we need saliency informa- M. Strube and C. Muller. 2003. A machine learning approach tion and various noun-noun relatedness informa- to pronoun resolution in spoken dialogue. In ACL, pages tion in antecedent selection. To this aim, we need 168–175. more feature engineering as well as more extensive A. Sumida, N. Yoshinaga, and K. Torisawa. 2008. Boosting empirical evaluations. precision and recall of hyponymy relation acquisition from hierarchical layouts in wikipedia. In LREC. References V. N. Vapnik. 1995. The Nature of Statistical Learning The- F. B. Baldwin. 1995. Cogniac: a discourse processing engine. ory. Ph.D. thesis. X. Yang, G. Zhou, J. Su, and C. L. Tan. 2003. Coreference S. E. Brennan, M. W. Friedman, and C. J. Pollard. 1987. A resolution using competition learning approach. In ACL, centering approach to pronouns. In ACL, pages 155–162. pages 176–183.