{ "paper_id": "I11-1023", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:31:55.593184Z" }, "title": "Japanese Predicate Argument Structure Analysis Exploiting Argument Position and Type", "authors": [ { "first": "Yuta", "middle": [], "last": "Hayashibe", "suffix": "", "affiliation": { "laboratory": "", "institution": "Nara Institute of Science and Technology", "location": { "postCode": "8916-5, 630-0192", "settlement": "Takayama", "region": "Ikoma Nara", "country": "Japan" } }, "email": "yuta-h@is.naist.jp" }, { "first": "Mamoru", "middle": [], "last": "Komachi", "suffix": "", "affiliation": { "laboratory": "", "institution": "Nara Institute of Science and Technology", "location": { "postCode": "8916-5, 630-0192", "settlement": "Takayama", "region": "Ikoma Nara", "country": "Japan" } }, "email": "komachi@is.naist.jp" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "", "affiliation": { "laboratory": "", "institution": "Nara Institute of Science and Technology", "location": { "postCode": "8916-5, 630-0192", "settlement": "Takayama", "region": "Ikoma Nara", "country": "Japan" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "We propose an approach to Japanese predicate argument structure analysis exploiting argument position and type. In particular, we propose the following two methods. First, in order to use information in the sentences in preceding context of the predicate more effectively, we propose an improved similarity measure between argument positions which is more robust than a previous co-reference-based measure. Second, we propose a flexible selection-and-classification approach which accounts for the minor types of arguments. Experimental results show that our proposed method achieves state-ofthe-art accuracy for Japanese predicate argument structure analysis.", "pdf_parse": { "paper_id": "I11-1023", "_pdf_hash": "", "abstract": [ { "text": "We propose an approach to Japanese predicate argument structure analysis exploiting argument position and type. In particular, we propose the following two methods. First, in order to use information in the sentences in preceding context of the predicate more effectively, we propose an improved similarity measure between argument positions which is more robust than a previous co-reference-based measure. Second, we propose a flexible selection-and-classification approach which accounts for the minor types of arguments. Experimental results show that our proposed method achieves state-ofthe-art accuracy for Japanese predicate argument structure analysis.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The goal of predicate-argument structure analysis is to extract semantic relations such as \"who did what to whom\" that hold between a predicate and its arguments constituting a semantic unit of a sentence. It is an important step in many Natural Language Processing applications such as machine translation, summarization and information extraction.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Arguments are classified into three categories according to their positions relative to the predicates: intra-sentential arguments (those that have direct syntactic dependency with the predicates), zero intra-sentential arguments (those appearing as zero-pronouns but have their antecedents in the same sentence), and inter-sentential arguments (those appearing as zero-pronouns and their antecedents are not in the same sentence). We call them INTRA D, INTRA Z, and INTER respectively. Furthermore, we call these categories the argument types. While the analysis of INTRA D is comparatively easy, INTRA Z and INTER are more difficult. We consider that there are two reasons for this.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The first reason is the poverty of features for argument identification compared to INTRA D. While for INTRA D we have important clues such as the function word or directly dependency relation, we don't for INTRA Z and INTER.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The second reason is the limited amount of training examples. For example, in a Japanese newswire corpus, INTRA Z and INTER account for 30.5% and 12.4% of all the nominative (ga) cases, and 13.1% and 0.2% of all of the accusative (wo) cases (Iida et al., 2007) .", "cite_spans": [ { "start": 241, "end": 260, "text": "(Iida et al., 2007)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper, in order to solve these problems we propose the following two methods exploiting argument position and type.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "First, we propose an improved similarity measure between argument positions of two predicates that take semantically similar arguments. For example, someone possibly arrested can also surrender him/herself, that is, objects of \"arrest\" and subjects of \"surrender (oneself)\" are occupied by semantically similar nouns. Gerber and Chai (2010) proposed analysis of English nominal predicates with this similarity to take discourse context into account. However, the similarity measure they used has drawbacks: it requires a co-reference resolver and a large number of documents. We improve their similarity measure alleviating these drawbacks by using argument position. We detail previous work on capturing discourse context in Section 2, and our proposal in Section 3.1.", "cite_spans": [ { "start": 318, "end": 340, "text": "Gerber and Chai (2010)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Second, we propose a selection-andclassification approach. In this approach, in order to compensate for the relative infrequency of examples of INTRA Z and INTER, we select a candidate argument for each argument type independently. After selecting candidates, we use classifiers to choose the correct argument type. This allows us to flexibly design features for each step and we can use pairwise features between the candidate arguments. We detail this in Section 3.2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The experimental results demonstrated that our proposed method achieved the state-of-the-art of Japanese predicate argument structure analysis. Iida et al. (2003) used Salient Reference List (Nariyama, 2002) based on Centering Theory (Grosz et al., 1995) , which explains the structure of discourse and the transition of topics in order to capture discourse context. The list has the following four ordered slots.", "cite_spans": [ { "start": 144, "end": 162, "text": "Iida et al. (2003)", "ref_id": "BIBREF5" }, { "start": 191, "end": 207, "text": "(Nariyama, 2002)", "ref_id": "BIBREF15" }, { "start": 234, "end": 254, "text": "(Grosz et al., 1995)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "TOPIC (marked by wa-particle)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "> SUBJECT (ga) > INDIRECT OBJECT (ni) > DIRECT OBJECT (wo),", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We check whether each candidate corresponds to any slots from the beginning of a document. If the candidate corresponds to a slot, we (over)write the slot with the candidate. We repeat this until we reach the predicate to analyze. We use the ranks of candidates in the list as a feature. Iida et al. (2003) used a feature (CHAIN LENGTH) that stands for how often each candidate is used as an argument of predicates in preceding context. Imamura et al. (2009) used a similar binary feature (USED) that shows if each candidate is ever used as an argument of predicates or not. However, they did not investigate the effect of these features explicitly in their systems. Therefore we also investigate these in this paper.", "cite_spans": [ { "start": 288, "end": 306, "text": "Iida et al. (2003)", "ref_id": "BIBREF5" }, { "start": 437, "end": 458, "text": "Imamura et al. (2009)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In the study of implicit arguments 1 for English nominal predicates, Gerber and Chai (2010) used similarity features between an argument position and a co-reference chain, inspired by Chambers and Jurafsky (2008) , who proposed unsupervised learning of narrative event chains using pointwise mutual information (PMI) between syntactic positions. This method stands on the assumption that similar argument positions tend to have the arguments which belong to a common co-reference chain. For instance, co-referring arguments at such argument positions like plead, ARG 0 , admit, ARG 0 , convict, ARG 1 , tend to take semantically similar nouns as the argument positions like sentence, ARG 1 , parole, ARG 1 .", "cite_spans": [ { "start": 69, "end": 91, "text": "Gerber and Chai (2010)", "ref_id": "BIBREF2" }, { "start": 184, "end": 212, "text": "Chambers and Jurafsky (2008)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Similarity between an Argument Position and a co-Reference Chain", "sec_num": "2.3" }, { "text": "They first automatically label a subset of the Gigaword corpus (Graff, 2003) with verbal and nominal semantic role labeling. They then identify co-references between arguments using a coreference resolver. They compute PMI as follows.", "cite_spans": [ { "start": 63, "end": 76, "text": "(Graff, 2003)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Similarity between an Argument Position and a co-Reference Chain", "sec_num": "2.3" }, { "text": "Suppose the resulting data has N co-referential pairs of argument positions and M of these pairs comprising E a = P a , A a , E b = P b , A b , and E c = P c , A c . P a , P b , and P c are predicates, and A a , A b , and A c are labels such as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Similarity between an Argument Position and a co-Reference Chain", "sec_num": "2.3" }, { "text": "ARG 0 or ARG 1 . pmi(E a , E b ) = log G(E a , E b ) G(E a , * )G(E b , * ) G(E a , E b ) =", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Similarity between an Argument Position and a co-Reference Chain", "sec_num": "2.3" }, { "text": "M N With this similarity between argument positions, they defined scores between an argument position and a co-reference chain.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Similarity between an Argument Position and a co-Reference Chain", "sec_num": "2.3" }, { "text": "Analysis Exploiting Argument Position and Type", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicate Argument Structure", "sec_num": "3" }, { "text": "Suppose we want to identify the argument of (surrendered) in Example (1). The argument is an antecedent of zero-pronoun \u03c6 of the predicate.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Similarity between Argument Positions using Distribution Similarity", "sec_num": "3.1" }, { "text": "(1) police wa-particle hanako wo-particle arrested Police arrested Hanako.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Similarity between Argument Positions using Distribution Similarity", "sec_num": "3.1" }, { "text": "had surrendered that heard I heard that \u03c6 had surrendered.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I (\u03c6 )", "sec_num": null }, { "text": "With Salient Reference List for \" (surrendered)\", the rank of \" (police)\" is higher than that of \" (Hanako)\" and it is noisy information for analysis. We also cannot distinguish them with argument frequency information, because frequencies of both \" (Hanako)\" and \" (police)\" are 1. Though it is reasonable to use the similarity between an argument position and a co-reference chain, the similarity measure described in Section 2.3 has two problems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I (\u03c6 )", "sec_num": null }, { "text": "One is the strong dependency on the accuracy of co-reference resolver system. In fact, the accuracy of Japanese co-reference resolvers is not accurate enough to create co-reference chains in good quality. 2 The other problem is the problem that it needs a lot of documents, because the method does not use any non co-referring nouns.", "cite_spans": [ { "start": 205, "end": 206, "text": "2", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "I (\u03c6 )", "sec_num": null }, { "text": "To avoid using an unreliable co-reference resolver, we can suppose the same noun lemmas without pronouns in the same document are coreferences. Pekar 2006called the noun lemmas anchors and they supposed the similarity measure between syntactic positions. For example, there are two anchors: \"Mary\" and \"house\" in the sentences \"Mary bought a house. The house belongs to Mary.\" They extract two groups: { buy(obj:X), belong(subj:X) } and {buy(subj:X), belong(to:X). } Nevertheless, this method also requires many documents because noun lemmas without anchors are not used for the calculation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I (\u03c6 )", "sec_num": null }, { "text": "In this paper, we propose a more robust similarity measure between argument positions which does not depend on unreliable co-reference annotations by the resolver. (arrest)\" following wo look alike. We can expect that an arrested person is more likely to be a person who has surrendered than an arrestee. We define a novel similarity of two argument positions 2 We implemented the method proposed by Iida et al. (2005a) , and the F-measure was 66%.", "cite_spans": [ { "start": 360, "end": 361, "text": "2", "ref_id": null }, { "start": 400, "end": 419, "text": "Iida et al. (2005a)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "I (\u03c6 )", "sec_num": null }, { "text": "E2 E3 and the accusative ( wo ) case of \" (arrest)\" (E 3 ). We will use these values as features of predicate-argument analysis in the experiments.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "E1", "sec_num": null }, { "text": "Considering Argument Type", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Selection-and-Classification Approach", "sec_num": "3.2" }, { "text": "In previous work, argument analysis was performed with common features regardless its argument type. However, these methods have difficulty in distinguishing the marginal cases where two candidates have different argument types because of the difference of quantity by argument types. Thus we propose the selection-and-classification approach for Japanese predicate argument structure analysis. This approach consists of two steps: the selection step and the classification step. This approach is inspired by two models. The first is the selection-and-classification model (Iida et al., 2005b) for noun phrase anaphora resolution. The model first selects a likely antecedent of the target (possibly) anaphoric expression. Second, the model classifies the target anaphoric ex- The second is the tournament model (Iida et al., 2003) for zero-anaphora resolution. For all the candidate antecedents (virtually all noun phrases appearing in preceding context), the model repeats two-class classification: which candidate in the pair of candidates is likely to be the antecedent for the zero-anaphora. The advantage of the tournament model is that the model can use pairwise features of candidates. Similarly, in the classification step of our approach we select an argument comparing most likely candidates of arguments of each argument type.", "cite_spans": [ { "start": 573, "end": 593, "text": "(Iida et al., 2005b)", "ref_id": "BIBREF7" }, { "start": 811, "end": 830, "text": "(Iida et al., 2003)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Selection-and-Classification Approach", "sec_num": "3.2" }, { "text": "Selection: At the first step, we select three most likely arguments of INTRA D, INTRA Z, and IN-TER for each predicate using any argument identification model. We may use different features for models of different argument types.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Method of Argument Analysis", "sec_num": null }, { "text": "Classification: At the second step, we determine which INTRA D, INTRA Z, and INTER is the correct argument or if there is no explicit argument appearing in the context. This step is composed of three binary classification models illustrated in Figure 1 .", "cite_spans": [], "ref_spans": [ { "start": 244, "end": 252, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "A Method of Argument Analysis", "sec_num": null }, { "text": "(a) Judge which of INTRA D or INTRA Z is more likely to be an argument of the predicate. (b) Judge which of INTER or the candidate selected at (a) is more likely to be an argument of the predicate. (c) Judge whether the candidate selected at (b) qualifies as an argument of the predicate or not.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Method of Argument Analysis", "sec_num": null }, { "text": "We show the example of analysis of \u03c6 in Example (1). We first select argument candidates of IN-TRA D, INTRA Z, and INTER. Suppose the most likely argument of INTRA D is not selected and \" (I)\" and \" (Hanako)\" are selected as ones of INTRA Z and INTER respectively in the 'selection' step. Because INTRA D is not selected, the classifier selects INTRA Z at (a). Suppose IN-TER is selected at (b) comparing \" \" selected at (a) and \" \". Finally, \" \" is selected as the argument by the classifier of (c).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Method of Argument Analysis", "sec_num": null }, { "text": "Furthermore, though we tried different orders for 'Classification' step in the preliminary experiment, this order was the best.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Method of Argument Analysis", "sec_num": null }, { "text": "We train each binary classifier in the order of (a), (b), and (c). We create training examples of classifiers with two argument candidates and a predicate as shown in Tables 3 and 4. The following arguments are used for training: (a) the correct argument and the most likely argument selected at the 'Selection' step (b) the correct argument and the most likely argument selected by (a) at the 'Classification' step (c) the correct argument and the most likely argument selected by (b) in the 'Classification' step For instance, \u03c6 in Example (1) is \" (Hanako)\" whose argument type is INTER. Hence CHAIN LENGTH A frequency of being arguments in previous sentences (Iida et al., 2003 ) USED Whether being arguments in former sentences or not (Imamura et al., 2009) ", "cite_spans": [ { "start": 663, "end": 681, "text": "(Iida et al., 2003", "ref_id": "BIBREF5" }, { "start": 740, "end": 762, "text": "(Imamura et al., 2009)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Training Method of Classifiers for the 'Classification' Step", "sec_num": null }, { "text": "Scores between an argument position and co-reference chain calculated with the similarity which Gerber and Chai (2010) used (described in Section 2.3) SIM CS Scores between an argument position and co-reference chain calculated with our proposed similarity Table 5 : Discourse context features used in the experiment we generate two training examples: One is an example of (b) with the label INTER, \" \", and the most likely argument selected by (a) at 'Classification' step. The other one is an example of (c) with the label HAVE-ARG and \"", "cite_spans": [], "ref_spans": [ { "start": 257, "end": 264, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "SIM COREF", "sec_num": null }, { "text": "\".", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "SIM COREF", "sec_num": null }, { "text": "We evaluate our proposed selection-andclassification approach by comparing it with other models and the discourse context features shown in Table 5 by adding them to the baseline features at Japanese predicate argument structure analysis of nominative case. In the experiment, systems refer only nouns in co-reference chains which are intra-sentential arguments. In addition, we used human annotated data of co-reference and predicate-argument structure to make discourse context features. For SIM COREF and SIM CS, we used maximum, minimum and average scores of similarities.", "cite_spans": [], "ref_spans": [ { "start": 140, "end": 147, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "Evaluation Setting of Predicate Argument Structure Analysis Exploiting Argument Position and Type", "sec_num": "4" }, { "text": "We used two datasets for the calculation of similarities: the Newspapers (NEWS) and the Web texts (WEB).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Dataset for Similarity Calculation", "sec_num": "4.1" }, { "text": "We used about 21,000,000 sentences in Mainichi newspapers published from 1991 to 2003 (excluded 1995). We part-of-speech tagged the data with MeCab 0.98 3 and dependency structure parsed with CaboCha 0.60pre4 4 . Both taggers used the NAIST Japanese Dictionary 0.6.3 5 . We extracted 27,282,277 pairs of a predicate and an argument. 6 We also extracted 111,173,873,092 coreference chains to calculate SIM COREF with the anaphora resolver which is our reimplementation of (Iida et al., 2005a) . These chains include 2,280,417,516,455 nouns. We used 173,778,624 pairs of a predicate and an argument with the case maker ga ,", "cite_spans": [ { "start": 333, "end": 334, "text": "6", "ref_id": null }, { "start": 471, "end": 491, "text": "(Iida et al., 2005a)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "NEWS:", "sec_num": null }, { "text": "wo and ni .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "NEWS:", "sec_num": null }, { "text": "We used about 500,000,000 sentences which Kawahara and Kurohashi (2006) collected from the web. They are part-of-speech tagged with JUMAN 7 and dependency structure parsed with KNP 8 . We extracted 1,101,472,855 pairs of a predicate and an argument. 9", "cite_spans": [ { "start": 42, "end": 71, "text": "Kawahara and Kurohashi (2006)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "WEB:", "sec_num": null }, { "text": "We used NAIST Text Corpus 1.4\u03b2 (Iida et al., 2007) for training and evaluation. It is based on Kyoto Text Corpus 3.0 10 and annotated with predicate-argument structure, event noun structure, and co-reference of nouns about 40,000 sentences of Japanese newspaper text. We excluded 11 articles due to annotation error. We conducted five-fold cross-validation. In the experiments, base phrases and dependency relations are acquired from the Kyoto Text Corpus 3.0 in the same way of related work.", "cite_spans": [ { "start": 31, "end": 50, "text": "(Iida et al., 2007)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Training and Evaluation Dataset", "sec_num": "4.2" }, { "text": "In order to identify the most likely argument candidate of each INTRA D, INTRA Z, and INTER, we used the tournament model. We emphasize that our proposed approach can use any argument identification model to identify the most likely candidate of an argument.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A Model in the 'Selection' Step", "sec_num": "4.3" }, { "text": "As baseline features, we employed features proposed by Iida et al. (2005a Iida et al. ( , 2007a and Imamura et al. (2009) in addition to a novel one 'PRED DEP POS' shown in Table 6 .", "cite_spans": [ { "start": 55, "end": 73, "text": "Iida et al. (2005a", "ref_id": "BIBREF6" }, { "start": 74, "end": 95, "text": "Iida et al. ( , 2007a", "ref_id": null }, { "start": 100, "end": 121, "text": "Imamura et al. (2009)", "ref_id": "BIBREF10" } ], "ref_spans": [ { "start": 173, "end": 180, "text": "Table 6", "ref_id": "TABREF8" } ], "eq_spans": [], "section": "Baseline Features and Classifier", "sec_num": "4.4" }, { "text": "We used Support Vector Machine (Cortes and Vapnik, 1995) a linear kernel. We used the implementation of LIBLINEAR 1.7 11 with its default parameters.", "cite_spans": [ { "start": 31, "end": 56, "text": "(Cortes and Vapnik, 1995)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Baseline Features and Classifier", "sec_num": "4.4" }, { "text": "We evaluate our selection-and-classification approach by comparing our baseline model with two previous approaches TA and IM.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Targets for Comparison of Predicate Argument Analysis Model", "sec_num": "4.5" }, { "text": "TA: Taira et al. (2008) used decision lists where features were sorted by their weights learned from Support Vector Machine. They simultaneously solved the argument of event nouns in the same lists.", "cite_spans": [ { "start": 4, "end": 23, "text": "Taira et al. (2008)", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Targets for Comparison of Predicate Argument Analysis Model", "sec_num": "4.5" }, { "text": "IM: Imamura et al. (2009) used discriminative models based on maximum entropy. They added the special noun phrase NULL, which expresses that the predicate does not have any argument.", "cite_spans": [ { "start": 4, "end": 25, "text": "Imamura et al. (2009)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Targets for Comparison of Predicate Argument Analysis Model", "sec_num": "4.5" }, { "text": "Because previous work use different features and machine learning methods and experiment on different setting from ours, we also compare with a baseline model BL in order to analyze the effect 11 http://www.csie.ntu.edu.tw/ \u223c cjlin/liblinear/ of dividing a model considering argument type. BL: This model has a single step in 'classification' step. In other words, 'selection' step in this model selects the most likely argument from all noun phrases preceding the predicate. Table 7 presents the result of the experiments. According to the bottom row in Table 7 , we achieved the state-of-the-art of Japanese predicate argument structure analysis by combining all discourse context features (+A+B+C+D+E).", "cite_spans": [], "ref_spans": [ { "start": 476, "end": 483, "text": "Table 7", "ref_id": null }, { "start": 555, "end": 562, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Targets for Comparison of Predicate Argument Analysis Model", "sec_num": "4.5" }, { "text": "We investigate our result from five different standpoints.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5" }, { "text": "We analyze the effect of our proposed selectionand-classification approach by comparing the first row of Table 7 . SC is superior to BL in all types. This shows that dividing a model considering argument type improves the performance. Table 7 : Comparison of predicate argument structure analysis of nominative case: P , R, and F 1 indicate Precision, Recall, and F-measure(\u03b2 = 1), respectively.", "cite_spans": [], "ref_spans": [ { "start": 105, "end": 112, "text": "Table 7", "ref_id": null }, { "start": 235, "end": 242, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Effect of the Selection-and-Classification Approach", "sec_num": "5.1" }, { "text": "By comparing SC and TA, and SC+USED and IM 12 , the result of our proposed method is competitive or superior to others. Additionally, recall is higher in any type; therefore we consider there is still much room for improvement by replacing the argument identification model in the selectional step with other models.", "cite_spans": [ { "start": 13, "end": 23, "text": "SC and TA,", "ref_id": null }, { "start": 24, "end": 45, "text": "and SC+USED and IM 12", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Comparison between previous work", "sec_num": "5.2" }, { "text": "On comparing +A (CHAIN LENGTH), +B (USED), and +C (SIM COREF NEWS) or +D (SIM CS NEWS) in Table 7 , similarity-based features are superior or competitive to frequencybased feature.", "cite_spans": [], "ref_spans": [ { "start": 90, "end": 97, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Effect of Similarity Metrics", "sec_num": "5.3" }, { "text": "(2) . . . The number of marriages increases 10,000 to 40,000 couples annually . . . . . .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Effect of Similarity Metrics", "sec_num": "5.3" }, { "text": ". . .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Effect of Similarity Metrics", "sec_num": "5.3" }, { "text": "The flu that has been going around and triggered . . . ,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Effect of Similarity Metrics", "sec_num": "5.3" }, { "text": "For instance, the argument of \" (be going around)\" in Example (2) is \" (flu)\" of INTER and is not an argument of previous arguments. Though the topic changes between two sentences, A and B cannot take it into account this and output \" (Marriages)\" which is an argument of \" (increase)\" because the frequencybased feature is active. In contrast, C and D handle this because the similarity between the nominative case of \" \" and \" \" is low. On comparing +C (SIM COREF NEWS) and +D (SIM CS NEWS) in Table 7 , our proposed similarity metrics work better than the coreference-based metrics in INTRA D or INTRA Z by a large margin. This result shows the robustness of our metrics compared to the co-reference based similarity between argument positions.", "cite_spans": [], "ref_spans": [ { "start": 496, "end": 503, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Effect of Similarity Metrics", "sec_num": "5.3" }, { "text": "On comparing +D (SIM CS NEWS) and +E (SIM CS WEB) in Table 7 respectively, the similarity measure using the newswire texts works better for INTRA D and one using the web texts works better for INTRA Z and INTER.", "cite_spans": [], "ref_spans": [ { "start": 53, "end": 60, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Effect of In and Out-of-domain Data", "sec_num": "5.4" }, { "text": "Additionally, the result of +D+E shows that combining proposed similarities calculated from different sources work complementary.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Effect of In and Out-of-domain Data", "sec_num": "5.4" }, { "text": "Removing features one by one from ALL (Adding all of A to E), we inquire about features which have strong effect on ALL. Table 7 shows that the F-measures of INTRA D and INTRA Z fall by a large margin, by removing D and E respectively. Though the F-measure of INTER degrades by removing C, it makes little difference to other argument types. This shows it is our proposed similarity that mainly contributes to the improvement of the F-measure of the overall system.", "cite_spans": [], "ref_spans": [ { "start": 121, "end": 128, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Ablation Features", "sec_num": "5.5" }, { "text": "We analyze errors where our proposed similarity does not work well.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Error Analysis", "sec_num": "6" }, { "text": "In short, this is equivalent to INTER.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://mecab.sourceforge.net/ 4 http://chasen.org/ \u223c taku/software/cabocha/ 5 http://sourceforge.jp/projects/naist-jdic/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We compare SC+USED and IM, because IM used the USED feature.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "We thank Daisuke Kawahara and Sadao Kurohashi for providing the web texts and Joseph Irwin for his comments on the earlier draft.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null }, { "text": "In NAIST Text Corpus, the copula \" \"(In English, \"be\") is annotated as a predicate.(3). . . . . . The price of apartments is going down.. . . \u03c6 \u03c6 of last year was 5.8 times higher than that of this year.However, the behavior of copula is different from other predicates, thus it is difficult to resolve them with the same model. To solve this problem, the model of copula should be separated.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Copula", "sec_num": "6.1" }, { "text": "In this experiment, we regarded the predicate of \" (do) + noun\" such as \" (do arrest)\" as a single predicate. On the other hand, we regarded the predicate of \" (do) + particle + noun\" such as \" \" in Example (4) as \" (do).\". . . \u03c6 should not relegate promotion of decentralization . . . However, verbs like \"do\" in such examples do not play central roles, whereas the noun such as \" (relegation)\" carries the main meaning of the event. This phenomenon is called \"light verb construction\" (Miyamoto, 1999) .\" \" is the nominalized form of the verb \" (relegate).\" Thus we need to calculate similarity with \" \" instead of \" \". When the predicate is a light verb, we have to use the original verb to calculate the similarity.", "cite_spans": [ { "start": 487, "end": 503, "text": "(Miyamoto, 1999)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Light Verb Construction", "sec_num": "6.2" }, { "text": "A predicate may have several senses and hence have several argument distributions. For example, \"\" has two senses at least: to pack and to bring to a conclusion. \" \" in Example (5) means the latter.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicate Sense Ambiguity", "sec_num": "6.3" }, { "text": "They emphasized that \u03c6 should be brought to an conclusion as soon as possible.The distributions of arguments of such ambiguous verbs tend to have a mixture of several distributions of arguments. Therefore this makes it hard to calculate the similarity of an argument position and a co-reference chain. Additionally, it is even more difficult when the predicate is more essential verb such as \" (have)\" and \" (take).\" This suggests the close relationship between the word sense disambiguation and the predicate argument structure analysis. In fact, Meza-Ruiz and Riedel (2009) showed that the joint model for semantic role labeling and word sense disambiguation performs better than a pipeline system.Since NAIST Text Corpus is not annotated with verb senses, we are annotating the sense of verbs to allow similar analysis.", "cite_spans": [ { "start": 548, "end": 575, "text": "Meza-Ruiz and Riedel (2009)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "(5) \u03c6", "sec_num": null }, { "text": "We improved Japanese predicate argument structure analysis exploiting argument position and type. In particular, we proposed two methods: the improved similarity measure between argument positions and the selection-and-classification approach considering argument type.Experimental results show that our proposed method achieved state-of-the art accuracy for the Japanese predicate argument structure analysis. Proposed similarity between argument positions exploiting case maker is more robust than previous co-reference-based method that makes use of an unreliable automatic co-reference resolver. Furthermore, we proposed flexible approach which accounts for the minor types of arguments.Future work includes four topics: (1) to distinguish copula from other predicates; (2) to combine internal argument to take semantic argument into consideration if the verb is in light verb construction; (3) to perform word sense disambiguation before calculating similarity; (4) to conduct experiments not only on nominative case but also on other cases .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "7" } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Unsupervised Learning of Narrative Event Chains", "authors": [ { "first": "Nathanael", "middle": [], "last": "Chambers", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Jurafsky", "suffix": "" } ], "year": 2008, "venue": "Proceedings of Annual Meeting of the Association for Computational Linguistics-08 with the Human Language Technology Conference", "volume": "", "issue": "", "pages": "789--797", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nathanael Chambers and Dan Jurafsky. 2008. Un- supervised Learning of Narrative Event Chains. In Proceedings of Annual Meeting of the Association for Computational Linguistics-08 with the Human Language Technology Conference, pages 789-797.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Support-Vector Networks", "authors": [ { "first": "Corinna", "middle": [], "last": "Cortes", "suffix": "" }, { "first": "Vladimir", "middle": [], "last": "Vapnik", "suffix": "" } ], "year": 1995, "venue": "Machine learning", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Corinna Cortes and Vladimir Vapnik. 1995. Support- Vector Networks. Machine learning.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Beyond NomBank: A Study of Implicit Arguments for Nominal Predicates", "authors": [ { "first": "Matthew", "middle": [], "last": "Gerber", "suffix": "" }, { "first": "Y", "middle": [], "last": "Joyce", "suffix": "" }, { "first": "", "middle": [], "last": "Chai", "suffix": "" } ], "year": 2010, "venue": "Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "1583--1592", "other_ids": {}, "num": null, "urls": [], "raw_text": "Matthew Gerber and Joyce Y. Chai. 2010. Beyond NomBank: A Study of Implicit Arguments for Nom- inal Predicates. In Proceedings of the 48th Annual Meeting of the Association for Computational Lin- guistics, pages 1583-1592. Association for Compu- tational Linguistics.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Centering: A Framework for Modeling the Local Coherence of Discourse", "authors": [ { "first": "Barbara", "middle": [ "J" ], "last": "Grosz", "suffix": "" }, { "first": "Aravind", "middle": [ "K" ], "last": "Joshi", "suffix": "" }, { "first": "Scott", "middle": [], "last": "Weinstein", "suffix": "" } ], "year": 1995, "venue": "Computational Linguistics", "volume": "21", "issue": "2", "pages": "203--225", "other_ids": {}, "num": null, "urls": [], "raw_text": "Barbara J. Grosz, Aravind K. Joshi, and Scott Wein- stein. 1995. Centering: A Framework for Modeling the Local Coherence of Discourse. Computational Linguistics, 21(2):203-225.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Incorporating Contextual Cues in Trainable Models for Coreference Resolution", "authors": [ { "first": "Ryu", "middle": [], "last": "Iida", "suffix": "" }, { "first": "Kentaro", "middle": [], "last": "Inui", "suffix": "" }, { "first": "Hiroya", "middle": [], "last": "Takamura", "suffix": "" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 2003, "venue": "Proceedings of the 10th EACL Workshop on the Computational Treatment of Anaphora", "volume": "", "issue": "", "pages": "23--30", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ryu Iida, Kentaro Inui, Hiroya Takamura, and Yuji Matsumoto. 2003. Incorporating Contextual Cues in Trainable Models for Coreference Resolution. In Proceedings of the 10th EACL Workshop on the Computational Treatment of Anaphora, pages 23- 30.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Anaphora Resolution by Antecedent Identification Followed by Anaphoricity Determination", "authors": [ { "first": "Ryu", "middle": [], "last": "Iida", "suffix": "" }, { "first": "Kentaro", "middle": [], "last": "Inui", "suffix": "" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 2005, "venue": "ACM Transactions on Asian Language Information Processing", "volume": "4", "issue": "4", "pages": "417--434", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ryu Iida, Kentaro Inui, and Yuji Matsumoto. 2005a. Anaphora Resolution by Antecedent Identification Followed by Anaphoricity Determination. ACM Transactions on Asian Language Information Pro- cessing, 4(4):417-434.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "On the Issue of Combining Anaphoricity Determination and Antecedent Identification in Anaphora Resolution", "authors": [ { "first": "Ryu", "middle": [], "last": "Iida", "suffix": "" }, { "first": "Kentaro", "middle": [], "last": "Inui", "suffix": "" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 2005, "venue": "International Conference on Natural Language Processing and Knowledge Engineering", "volume": "", "issue": "", "pages": "244--249", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ryu Iida, Kentaro Inui, and Yuji Matsumoto. 2005b. On the Issue of Combining Anaphoricity Determi- nation and Antecedent Identification in Anaphora Resolution. In International Conference on Natural Language Processing and Knowledge Engineering, pages 244-249.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Annotating a Japanese Text Corpus with Predicate-Argument and Coreference Relations", "authors": [ { "first": "Ryu", "middle": [], "last": "Iida", "suffix": "" }, { "first": "Mamoru", "middle": [], "last": "Komachi", "suffix": "" }, { "first": "Kentaro", "middle": [], "last": "Inui", "suffix": "" }, { "first": "Yuji", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the Linguistic Annotation Workshop", "volume": "", "issue": "", "pages": "132--139", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ryu Iida, Mamoru Komachi, Kentaro Inui, and Yuji Matsumoto. 2007. Annotating a Japanese Text Cor- pus with Predicate-Argument and Coreference Re- lations. In Proceedings of the Linguistic Annotation Workshop, pages 132-139. Association for Compu- tational Linguistics.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Nihongo Goi Taikei :A Japanese Lexicon", "authors": [ { "first": "Satoru", "middle": [], "last": "Ikehara", "suffix": "" }, { "first": "Masahiro", "middle": [], "last": "Miyazaki", "suffix": "" }, { "first": "Satoshi", "middle": [], "last": "Shirai", "suffix": "" }, { "first": "Akio", "middle": [], "last": "Yokoo", "suffix": "" }, { "first": "Hiromi", "middle": [], "last": "Nakaiwa", "suffix": "" }, { "first": "Kentaro", "middle": [], "last": "Ogura", "suffix": "" } ], "year": 1997, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai, Akio Yokoo, Hiromi Nakaiwa, and Kentaro Ogura. 1997. Nihongo Goi Taikei :A Japanese Lexicon. Iwanami Shoten.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Discriminative Approach to Predicate-Argument Structure Analysis with Zero-Anaphora Resolution", "authors": [ { "first": "Kenji", "middle": [], "last": "Imamura", "suffix": "" }, { "first": "Kuniko", "middle": [], "last": "Saito", "suffix": "" }, { "first": "Tomoko", "middle": [], "last": "Izumi", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Joint conference of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing", "volume": "", "issue": "", "pages": "85--88", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kenji Imamura, Kuniko Saito, and Tomoko Izumi. 2009. Discriminative Approach to Predicate- Argument Structure Analysis with Zero-Anaphora Resolution. In Proceedings of the Joint conference of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Pro- cessing, pages 85-88.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Case Frame Compilation from the Web using High-Performance Computing", "authors": [], "year": null, "venue": "Proceedings of the 5th International Conference on Language Resources and Evaluation", "volume": "", "issue": "", "pages": "1344--1347", "other_ids": {}, "num": null, "urls": [], "raw_text": "Case Frame Compilation from the Web using High- Performance Computing. In Proceedings of the 5th International Conference on Language Resources and Evaluation, pages 1344-1347.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Jointly identifying predicates, arguments and senses using markov logic", "authors": [ { "first": "Ivan", "middle": [], "last": "Meza", "suffix": "" }, { "first": "-", "middle": [], "last": "Ruiz", "suffix": "" }, { "first": "Sebastian", "middle": [], "last": "Riedel", "suffix": "" } ], "year": 2009, "venue": "Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "155--163", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ivan Meza-Ruiz and Sebastian Riedel. 2009. Jointly identifying predicates, arguments and senses us- ing markov logic. In Proceedings of Human Lan- guage Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 155-163.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "The Light Verb Construction in Japanese: the role of the verbal noun", "authors": [ { "first": "Tadao", "middle": [], "last": "Miyamoto", "suffix": "" } ], "year": 1999, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tadao Miyamoto. 1999. The Light Verb Construction in Japanese: the role of the verbal noun. John Ben- jamins Publishing Company.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Grammar for Ellipsis Resolution in Japanese", "authors": [ { "first": "Sigeko", "middle": [], "last": "Nariyama", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 9th International Conference on Theoretical and Methodological Issues in Machine Translation", "volume": "", "issue": "", "pages": "135--145", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sigeko Nariyama. 2002. Grammar for Ellipsis Resolu- tion in Japanese. In Proceedings of the 9th Interna- tional Conference on Theoretical and Methodologi- cal Issues in Machine Translation, pages 135-145.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Acquisition of Verb Entailment from Text", "authors": [ { "first": "", "middle": [], "last": "Viktor Pekar", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the main conference on Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics", "volume": "", "issue": "", "pages": "49--56", "other_ids": {}, "num": null, "urls": [], "raw_text": "Viktor Pekar. 2006. Acquisition of Verb Entailment from Text. In Proceedings of the main conference on Human Language Technology Conference of the North American Chapter of the Association of Com- putational Linguistics, pages 49-56. Association for Computational Linguistics.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "A Japanese Predicate Argument Structure Analysis Using Decision Lists", "authors": [ { "first": "Hirotoshi", "middle": [], "last": "Taira", "suffix": "" }, { "first": "Sanae", "middle": [], "last": "Fujita", "suffix": "" }, { "first": "Masaaki", "middle": [], "last": "Nagata", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "523--532", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hirotoshi Taira, Sanae Fujita, and Masaaki Nagata. 2008. A Japanese Predicate Argument Structure Analysis Using Decision Lists. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 523-532.", "links": null } }, "ref_entries": { "TABREF2": { "html": null, "type_str": "table", "num": null, "content": "
shows the list of nouns that have direct
dependency arcs in syntactic dependency struc-tures along with case markers ga (nominative case), wo (accusative case) and ni (dative case) ex-
tracted from the WEB corpus described in Section
4. According to Table 1, the distributions of nouns of \" (surrender)\" following ga and \"
", "text": "" }, "TABREF3": { "html": null, "type_str": "table", "num": null, "content": "
inative (ga) case of \"(surrender)\" (E 1 )
", "text": "An example of similarities between argument positions calculated with WEB corpus.encoding such information as Jensen-Shannon divergence between argument distributions of argument positions. A sample of the calculated similarities is shown inTable 2. This table illustrates the most similar argument positions are the nom-" }, "TABREF5": { "html": null, "type_str": "table", "num": null, "content": "
Italic texts
in (b) and (c) refer most likely argument which (a)
and (b) selected respectively. Non-italic texts refer
to the correct argument.
", "text": "Examples made for training." }, "TABREF6": { "html": null, "type_str": "table", "num": null, "content": "", "text": "Examples made for training with humanannotated data. Generated examples depend on the argument type." }, "TABREF8": { "html": null, "type_str": "table", "num": null, "content": "
", "text": "Baseline features of classifiers in the 'Selection' step and the 'Classification' step." }, "TABREF9": { "html": null, "type_str": "table", "num": null, "content": "
SectionINTRA DINTRA ZINTER
P -85.2 5.3, 5.4 27 R F1 P R F1 P R F1 5.1 BL : 5.2 TA : Taira et al. 2008 -75.53 --30.15 --23.45 IM : Imamura et al. 2009 88.8 87.0 58.8 43.4 50.0 47.5 7.6 13.1 5.4 SC+D+E 89.70 91.55 90.62 65.08 61.37 63.17 24.86 21.57 23.08 5.5 ALL (SC+A+B+C+E+D) 89.93 91.70 90.81 67.39 62.18 64.68 25.86 22.93 24.30 ALL-A 90.44 91.34 90.89 68.12 61.95 64.89 25.72 23.77 24.69 ALL-B 90.48 91.48 90.98 66.83 62.06 64.35 25.48 22.71 24.01 ALL-C 89.30 91.65 90.46 65.19 61.92 63.50 25.71 22.22 23.83 ALL-D 87.65 90.48 89.04 66.14 61.21 63.57 26.03 22.85 24.32 ALL-E 89.66 90.73 90.19 62.47 59.04 60.71 25.56 22.49 23.92
", "text": "Our baseline 84.06 50.74 63.24 27.02 56.13 36.46 16.44 13.70 14.89 SC : Our proposed method 80.71 85.35 82.96 47.57 45.74 46.64 23.79 15.93 19.07 SC+A (CHAIN LENGTH) 85.39 88.79 87.05 51.64 52.22 51.93 25.31 18.63 21.44 SC+B (USED) 85.59 88.44 86.99 54.40 53.32 53.86 26.09 21.27 23.43 SC+C (SIM COREF NEWS) 86.82 88.90 87.85 54.07 52.89 53.47 25.83 20.08 22.58 SC+D (SIM CS NEWS) 88.42 91.10 89.74 59.05 58.12 58.58 24.81 19.91 22.08 SC+E (SIM CS WEB) 87.00 90.44 88.69 64.76 60.27 62.43 25.63 21.32 23." } } } }