Neural Japanese Zero Anaphora Resolution using Smoothed Large-scale Case Frames with Word Embedding

Neural Japanese Zero Anaphora Resolution using Smoothed Large-scale Case Frames with Word Embedding Souta Yamashiro Hitoshi Nishikawa Tokyo Institute of Technology / W8-73, 2-12-1 Ookayama, Meguro, Tokyo 152-8552, Japan {yamashiro.s.aa@m, {hitoshi, take}@c}.titech.ac.jp Takenobu Tokunaga Abstract This paper presents a Japanese zero anaphora resolution model which deals with both intra- and inter-sentential zero anaphora. Solving inter-sentential anaphora needs to consider a large number of antecedent candidates beyond the sentence boundaries, which is a crucial ob- stacle for training the model and resolving the anaphora. To cope with this problem, we pro- pose an effective candidate pruning method using case frame information. Also, we in- troduce a local single-attention RNN for inter- sentential anaphora resolution, allowing the model to consider the distant context from the target predicate. We evaluated the proposed models with a Japanese balanced corpus and confirmed the effectiveness of the candidate pruning by showing 0.056 point increase of accuracy. 1 Introduction Zero anaphora resolution is the task to detect omit- ted arguments (zero anaphors) of the predicate in a given text and to identify their antecedents. The an- tecedents might or might not appear in the text. In the latter case (exophora), the antecedents exist out- side of the text, e.g. the writer of the text. In the former case, they appear within the same sentence as the predicate (intra-sentential anaphora) or appear in the preceding sentences (inter-sentential anaphora)1. (1) Tikaku no syˆotengai-niDAT (watasi-gaNOM) osyarena tatemono-woACC mikakeruv1 yˆoninatta. Kafeteria-gaNOM tokuni ˆoku, kongetu mo 1We do not deal with cataphora in this study. (kafeteria-gaNOM) (Tikaku no syˆotengai-niDAT) ˆopunsiteiruv2. (INOM) seev1 fashionable buildingsACC in the nearby shopping districtDAT recently. There are many cafeteriasNOM in particular, and (a cafeteriaNOM) has openedv2 this month (in the nearby shopping disctictDAT). In the example (1), the nominative argument of v1 (see) and the nominative argument and dative argu- ment of v2 (open), which are enclosed by the paren- theses, are omitted from the sentences. The nomina- tive argument of “open” is “cafeteria” which appears in the same sentence (intra-sentential zero anaphora) and the dative argument is “the nearby shopping dis- trict” which appears in the previous sentence (inter- sentential zero anaphora). On the other hand, the nominative argument of v1 (see) is the writer of this text who is not explicitly mentioned in the text (ex- ophora). This study focuses on zero anaphora resolution of Japanese texts, but we observe such pronoun- dropping phenomenon in other languages as well, e.g. Chinese, Italian, Turkish and so on. There have been many studies on the task similar to the Japanese zero anaphora resolution in other languages (Iida and Poesio, 2011; Rello et al., 2012; Chen and Ng, 2016; Yin et al., 2017). The zero anaphora resolution is one of the ac- tive research areas in the Japanese language pro- cessing as it is crucial for improving the perfor- mance of various natural language processing appli- cations such as automatic text summarisation (Ya- mada et al., 2017), information extraction (Sudo et al., 2001) and machine translation (Kudo et …

Content truncated to 500 words