Bridging Anaphora Resolution as Question Answering

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1428–1438 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1428 Bridging Anaphora Resolution as Question Answering Yufang Hou IBM Research Europe yhou@ie.ibm.com Abstract Most previous studies on bridging anaphora resolution (Poesio et al., 2004; Hou et al., 2013b; Hou, 2018a) use the pairwise model to tackle the problem and assume that the gold mention information is given. In this paper, we cast bridging anaphora resolution as ques- tion answering based on context. This allows us to find the antecedent for a given anaphor without knowing any gold mention informa- tion (except the anaphor itself). We present a question answering framework (BARQA) for this task, which leverages the power of transfer learning. Furthermore, we propose a novel method to generate a large amount of “quasi-bridging” training data. We show that our model pre-trained on this dataset and fine-tuned on a small amount of in-domain dataset achieves new state-of-the-art results for bridging anaphora resolution on two bridging corpora (ISNotes (Markert et al., 2012) and BASHI (R¨osiger, 2018)). 1 Introduction Anaphora accounts for text cohesion and is cru- cial for text understanding. An anaphor is a noun phrase (NP) that usually refers back to the same or a different entity (the antecedent) in text. Anaphora resolution is the task to determine the antecedent for a given anaphor. While direct anaphora resolu- tion attracts a lot of attention in the NLP commu- nity recently, such as Winograd Schema Challenge (Rahman and Ng, 2012; Opitz and Frank, 2018; Ko- cijan et al., 2019), indirect anaphora resolution or bridging anaphora resolution is less well studied. In this paper, we focus on bridging anaphora resolution where bridging anaphors and their an- tecedents are linked via various lexico-semantic, frame or encyclopedic relations. Following Hou et al. (2013b) and R¨osiger et al. (2018), we mainly consider “referential bridging” in which bridging anaphors are truly anaphoric and bridging rela- tions are context-dependent. In Example 11, both “her building” and “buildings with substantial dam- age” are plausible antecedent candidates for the bridging anaphor “residents” based on lexical se- mantics. In order to find the antecedent (buildings with substantial damage), we have to take the mean- ing of the broader discourse context into account. (1) In post-earthquake parlance, her building is a “red”. After being inspected, buildings with substantial damage were color-coded. Green al- lowed residents to re-enter; yellow allowed lim- ited access; red allowed residents one last entry to gather everything they could within 15 minutes. Most previous studies on bridging anaphora res- olution (Poesio et al., 2004; Lassalle and Denis, 2011; Hou et al., 2013b; Hou, 2018a) tackle the problem using the pairwise model and assume that the gold mention information is given. Most work (Poesio et al., 2004; Lassalle and Denis, 2011; Hou et al., 2013b) uses syntactic patterns to measure semantic relatedness between the head nouns of an anaphor and its antecedent. Hou (2018a) proposes a simple deterministic algorithm that also consid- ers …

Content truncated to 500 words