Predicate-Argument Structure Analysis with Zero-Anaphora Resolution for Dialogue Systems

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 806–815, Dublin, Ireland, August 23-29 2014. Predicate-Argument Structure Analysis with Zero-Anaphora Resolution for Dialogue Systems Kenji Imamura, Ryuichiro Higashinaka, and Tomoko Izumi NTT Media Intelligence Laboratories, NTT Corporation 1-1 Hikari-no-oka, Yokosuka, 239-0847, Japan {imamura.kenji,higashinaka.ryuchiro,izumi.tomoko}@lab.ntt.co.jp Abstract This paper presents predicate-argument structure analysis (PASA) for dialogue systems in Japanese. Conventional PASA and semantic role labeling have been applied to newspaper arti- cles. Because pronominalization and ellipses frequently appear in dialogues, we base our PASA on a strategy that simultaneously resolves zero-anaphora and adapt it to dialogues. By incor- porating parameter adaptation and automatically acquiring knowledge from large text corpora, we achieve a PASA specialized to dialogues that has higher accuracy than that for newspaper articles. 1 Introduction Semantic role labeling (SRL) and predicate-argument structure analysis (PASA) are important analysis techniques for acquiring “who did what to whom” from sentences1. These analyses have been applied to written texts because most annotated corpora comprise newspaper articles (Carreras and M`arquez, 2004; Carreras and M`arquez, 2005; Matsubayashi et al., 2014). Recently, systems for speech dialogue between humans and computers (e.g., Siri of Apple Inc. and Shabette Concier of NTT DoCoMo) have become familiar with the popularization of smart phones. A man-machine dialogue system has to interpret human utterances to associate them with system utter- ances. The predicate-argument structure could be an effective data structure for dialogue management. However, it is unclear whether we can apply the SRL/PASA for newspaper articles to dialogues because there are many differences between them, such as the number of speakers, written or spoken language, and context processing. For example, the following dialogue naturally includes pronouns, and thus anaphora resolution is necessary for semantic role labeling. A: [I]ARG0 want [an iPad Air]ARG1. B: [When]ARGM will [you]ARG0 buy [it(=an iPad Air)]ARG1? Similar phenomena exist in Japanese dialogues. However, most pronouns are omitted (called zero- pronouns), and zero-anaphora resolution is necessary for Japanese PASA. A: [iPad Air]NOM -ga hoshii-na. iPad Air NOM. want “φ want an iPad Air.” B: itsu φNOMφACC kau-no? when buy? “When will φ buy φ?” This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http://creativecommons.org/licenses/by/4.0/ 1Recent SRL systems assign labels of predicates and their arguments as semantic roles. Consequently, SRL and PASA are very similar tasks. We use the term predicate-argument structure analysis in this paper because most Japanese analyzers use this term. 806 This paper presents predicate-argument structure analysis with zero-anaphora resolution for Japanese chat dialogues. Here, we regard the task of constructing PASA for dialogues as a kind of domain adap- tation from newspaper articles to dialogues. M`arquez et al. (2008) and Pradhan et al. (2008) indicated that the tuning of parameter distribution and reducing the out-of-vocabulary are important for the do- main adaptation of SRL. We also focus on parameter distribution and out-of-vocabulary to construct a PASA adapted to dialogues. To the best of our knowledge, this is …

Content truncated to 500 words