Incorporating Centering Theory into Neural Coreference Resolution

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2996 - 3002 July 10-15, 2022 ©2022 Association for Computational Linguistics Incorporating Centering Theory into Neural Coreference Resolution Haixia Chai and Michael Strube Heidelberg Institute for Theoretical Studies gGmbH Schloss-Wolfsbrunnenweg 35 69118 Heidelberg, Germany {haixia.chai, michael.strube}@h-its.org Abstract In recent years, transformer-based coreference resolution systems have achieved remarkable improvements on the CoNLL dataset. How- ever, how coreference resolvers can benefit from discourse coherence is still an open ques- tion. In this paper, we propose to incorpo- rate centering transitions derived from center- ing theory in the form of a graph into a neu- ral coreference model. Our method improves the performance over the SOTA baselines, es- pecially on pronoun resolution in long docu- ments, formal well-structured text, and clus- ters with scattered mentions.1 1 Introduction Coreference resolution is the task to find all ex- pressions which refer to the same entity. The coreferential mentions could occur anywhere in the discourse. In recent years, many transformer- based models (Joshi et al., 2019, 2020; Kirstain et al., 2021) achieved improvements on the CoNLL benchmark (Pradhan et al., 2012). In contrast to using transformers such as BERT (Devlin et al., 2019) which learn the text input sequentially in limited chunks, how knowledge about the structure of discourse can benefit coreference resolution is less explored in the neural NLP era. Coreference plays an essential role in discourse coherence. A referring expression using a reduced linguistic form (e.g., pronoun) indicates a referen- tial relation to its antecedent in previous utterances. The referring expression connects utterances and contributes to discourse coherence implicitly. On the other hand, coreference resolution can benefit from a coherent discourse. It has long been ac- knowledged that coherence structure can impose constraints on referential accessibility from a lin- guistic perspective (Asher and Lascarides, 2003). Centering theory (Joshi and Weinstein, 1981; Grosz 1Our code and model are publicly available at: https: //github.com/HaixiaChai/CT-Coref 1. Bill wanted John to look over some important papers. 2. He/Bill had to mail him the documents by Monday. 3. Unfortunately, he/John never received the papers. 4. As a result, the whole deal fell behind schedule. continue shift shift Figure 1: An example text shows how foci change sen- tence by sentence. The words in bold are the focus of each sentence. The arrows indicate centering tran- sitions with two different transition types, continue and shift. et al., 1983, 1995; Walker et al., 1998) is a method to formally describe discourse coherence by using attentional state (i.e., the focus of attention of the participants at each utterance of the discourse). Fig- ure 1 shows how the coherence structure of an ex- ample text is built by means of tracking the changes in the local attentional state.2 By applying center- ing theory, Gordon and Scearce (1995) investigate how local coherence influences the interpretation of ambiguous pronouns. From reading-time experi- ments, they observe that utterances with pronouns were read faster in the centering …

Content truncated to 500 words