Modeling Context-sensitive Selectional Preference with Distributed Representations

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2829–2838, Osaka, Japan, December 11-17 2016. Modeling Context-sensitive Selectional Preference with Distributed Representations Naoya Inoue Yuichiro Matsubayashi Masayuki Ono∗ Graduate School of Information Sciences Tohoku University 6-6-05 Aramaki Aza Aoba, Aobaku, Sendai, Miyagi 980-8579, Japan {naoya-i,y-matsu,masayuki.ono,okazaki,inui}@ecei.tohoku.ac.jp Naoaki Okazaki Kentaro Inui Abstract This paper proposes a novel problem setting of selectional preference (SP) between a predicate and its arguments, called as context-sensitive SP (CSP). CSP models the narrative consistency be- tween the predicate and preceding contexts of its arguments, in addition to the conventional SP based on semantic types. Furthermore, we present a novel CSP model that extends the neural SP model (Van de Cruys, 2014) to incorporate contextual information into the distributed representa- tions of arguments. Experimental results demonstrate that the proposed CSP model successfully learns CSP and outperforms the conventional SP model in coreference cluster ranking. 1 Introduction Selectional Preference (SP) of predicates is a term denoting a bias in co-occurrence of a predicate and its argument. Predicates tend to take a particular semantic type of phrase as an argument. For example, the object slot of eat is generally filled by a noun phrase denoting food such as an apple; it is rarely filled by a phrase that is not food such as a watch. As the knowledge of SP has been recognized as key for many natural language processing tasks, including semantic role labeling and anaphora resolution, automatic acquisition of SP knowledge has persisted as a popular research topic. In literature, a variety of computational models for SP have been proposed, ranging from thesaurus-based approaches (Resnik, 1996), to probabilistic latent variable models (Rooth et al., 1999; S´eaghdha and Korhonen, 2014), and distributed approaches (Van de Cruys, 2014). Conventionally, SP is defined as the context-independent acceptability of a word as a filler of a predi- cate in the sense of a semantic type. Suppose that we must identify the referent of him(j): (1) John(i) beat Bob(j). Mary comforted him(j). Henceforth, we call a predicate (e.g., comfort) and an argument (e.g., John and Bob) to be examined as a query predicate and query argument, respectively. Conventional SP models judge the appropriateness of John(i) and Bob(j) in terms of whether comfort can take each noun as its object. However, it ignores the information signified by the preceding context, namely John(i) beat Bob and Bob(j), whom John beat. Therefore, conventional approaches cannot determine the preference between John and Bob, both of whom can fill the object of comfort, with the same semantic type. In this paper, we propose a context-sensitive version of SP (CSP), a novel task setting in which SP is considered in discourse. In text (1), for instance, Bob(j) is considered to be a more plausible filler than John(i) in terms of contextual relatedness: one who was beaten is more likely to be comforted than one who beat someone. In this paper, we want to discriminate (2a) from (2b) in addition to the conventional SP-based discrimination …

Content truncated to 500 words