Distance-Free Modeling of Multi-Predicate Interactions in End-to-End Japanese Predicate-Argument Structure Analysis

Proceedings of the 27th International Conference on Computational Linguistics, pages 94–106 Santa Fe, New Mexico, USA, August 20-26, 2018. 94 Distance-Free Modeling of Multi-Predicate Interactions in End-to-End Japanese Predicate-Argument Structure Analysis Yuichiroh Matsubayashi♠♢ and Kentaro Inui♠♢ ♠Graduate School of Information Sciences, Tohoku University ♢RIKEN Center for Advanced Intelligence Project {y-matsu, inui}@ecei.tohoku.ac.jp Abstract Capturing interactions among multiple predicate-argument structures (PASs) is a crucial issue in the task of analyzing PAS in Japanese. In this paper, we propose new Japanese PAS analysis models that integrate the label prediction information of arguments in multiple PASs by extending the input and last layers of a standard deep bidirectional recurrent neural network (bi-RNN) model. In these models, using the mechanisms of pooling and attention, we aim to directly capture the potential interactions among multiple PASs, without being disturbed by the word order and distance. Our experiments show that the proposed models improve the prediction accuracy specifically for cases where the predicate and argument are in an indirect dependency relation and achieve a new state of the art in the overall F1 on a standard benchmark corpus. 1 Introduction A predicate-argument structure (PAS) is a structure that represents the relationships between a predicate and its arguments. Identifying PASs in Japanese text is a long-standing challenge chiefly due to the abundance of omitted (elliptical) arguments. In the example in Figure 1, the dative relation between answer and reporters is not explicitly indicated by the syntactic structure of the sentence. We regard such arguments as elliptical and call those argument slots Zero cases. 25% of the obligatory arguments in Japanese newspaper articles are reported to be elliptical.1 The accuracy of identifying the fillers of such Zero cases remains only around 50% in terms of F1 even if the task is restricted to the identification of intra-sentential predicate-argument relations (Matsubayashi and Inui, 2017). 答え た 。 記者 [団] が 尋ねる と [首相] が [reporters]-NOM ask then [prime minister]-NOM answer-PAST ‘The prime minister answered when the reporters asked.’ DAT DAT NOM NOM Figure 1: Example of PAS analysis. The dashed lines represent the predicate-argument relations. “[reporters]-NOM ask then” constitutes a subordinate clause and “[prime minister]-NOM answer- PAST” constitutes a matrix clause. One promising approach for addressing this problem is to model argument sharing across multiple predicates (Iida et al., 2015; Ouchi et al., 2015; Ouchi et al., 2017). In Figure 1, for example, one can find very limited syntactic clues for predicting the long-distance dative relation between answer and reporters. However, the relation must be easy to identify for human readers who know that the person who asks a question is likely to be answered; namely, the nominative argument of ask is likely to be shared with the dative argument of answer. Capturing such inter-predicative dependencies has, therefore, been considered crucial of Japanese PAS analysis. This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http: //creativecommons.org/licenses/by/4.0/ 1Statistics from the NAIST Text Corpus 1.5. (Iida et al., 2017) 95 Word emb. Target predicate Input sentence Predictions …

Content truncated to 500 words