LingMess: Linguistically Informed Multi Expert Scorers for Coreference Resolution

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2752–2760 May 2-6, 2023 ©2023 Association for Computational Linguistics LINGMESS: Linguistically Informed Multi Expert Scorers for Coreference Resolution Shon Otmazgin1 Arie Cattan1 Yoav Goldberg1,2 1Computer Science Department, Bar Ilan University 2Allen Institute for Artificial Intelligence {shon711,arie.cattan,yoav.goldberg}@gmail.com Abstract Current state-of-the-art coreference systems are based on a single pairwise scoring com- ponent, which assigns to each pair of mention spans a score reflecting their tendency to core- fer to each other. We observe that different kinds of mention pairs require different infor- mation sources to assess their score. We present LINGMESS, a linguistically motivated catego- rization of mention-pairs into 6 types of coref- erence decisions and learn a dedicated trainable scoring function for each category. This signif- icantly improves the accuracy of the pairwise scorer as well as of the overall coreference per- formance on the English Ontonotes coreference corpus and 5 additional datasets.1 1 Introduction Coreference resolution is the task of clustering tex- tual mentions that refer to the same discourse entity. This fundamental task requires many decisions. In this work, we argue that different kinds of decisions involve different challenges. To illustrate that, con- sider the following text: “Lionel Messi has won a record seven Ballon d’Or awards. He signed for Paris Saint-Germain in August 2021. “I would like to thank my family”, said the Argentinian footballer. Messi holds the records for most goals in La Liga” To correctly identify that the pronoun “He” refers to “Lionel Messi”, models need to model the discourse, while linking “my” to “I” may rely more heavily on morphological agreement. Like- wise, linking “the Argentinian footballer” to “Li- onel Messi” requires world knowledge, while link- ing “Messi” to “Lionel Messi” may be achieved by simple lexical heuristics. able1The codebase to train and run LINGMESS is avail- in https://github.com/shon-otmazgin/ lingmess-coref. Also, our recent F-COREF Python package (Otmazgin et al., 2022) includes a simple and efficient implementation of LINGMESS in https: //github.com/shon-otmazgin/fastcoref. Figure 1: Architecture of our multi expert model. Given two spans “Lionel Messi” and “He”, we sum four scores: individual mention scores (black), fm(“Lionel Messi”), fm(“He”), and pairwise scores, shared antecedent score (white) fa(“Lionel Messi”, “He”) and the relevant “expert” score (blue) f PRON-ENT a (“Lionel Messi”, “He”). Indeed, pre-neural coreference resolution works often considered the types of a mention-pair, either by incorporating this information as model features, or by tailoring specific rules or specific models for each mention pair (see related work section). However, neural-network based coreference models are all based on a single pairwise scorer that is shared for all mention pairs, regardless of the different challenges that needs to be addressed by each pair type (Lee et al., 2017, 2018; Joshi et al., 2019; Kantor and Globerson, 2019; Joshi et al., 2020; Xu and Choi, 2020; Xia et al., 2020; Toshniwal et al., 2020; Thirukovalluru et al., 2021; Kirstain et al., 2021; Dobrovolskii, 2021). In this work, we suggest that modeling different mention pairs by …

Content truncated to 500 words