End-to-End Neural Bridging Resolution

Proceedings of the 29th International Conference on Computational Linguistics, pages 766–778 October 12–17, 2022. 766 End-to-End Neural Bridging Resolution Hideo Kobayashi1, Yufang Hou2 and Vincent Ng1 1 Human Language Technology Research Institute, University of Texas at Dallas, USA 2 IBM Research Europe, Ireland {hideo,vince}@hlt.utdallas.edu yhou@ie.ibm.com Abstract The state of bridging resolution research is rather unsatisfactory: not only are state-of- the-art resolvers evaluated in unrealistic set- tings, but the neural models underlying these resolvers are weaker than those used for en- tity coreference resolution. In light of these problems, we evaluate bridging resolvers in an end-to-end setting, strengthen them with better encoders, and attempt to gain a better under- standing of them via perturbation experiments and a manual analysis of their outputs. 1 Introduction Bridging was used by Clark (1975) to refer to non- identity relations between anaphoric noun phrases (i.e., bridging anaphors) and their antecedents. In Example 1, “ruling – party members” is a bridging anaphor and its antecedent is “Japan”. (1) Yet another political scandal is racking Japan. But this time it’s hurting opposition as well as ruling - party members. Bridging resolution is the task of identifying bridg- ing anaphors and linking them to their antecedents. Many tasks can benefit from bridging resolution, such as textual entailment (Mirkin et al., 2010) and question answering (Tseng et al., 2021). Bridging resolution is arguably less studied but more challenging that entity coreference resolu- tion, the task of determining which entity mentions refer to the same entity in the real world. Specif- ically, while linguistic constraints on coreference exist at the grammatical (e.g., gender and number agreement), syntactic (e.g., c-command), and se- mantic (e.g., semantic type agreement) levels that can be used to filter candidate antecedents, such constraints are largely absent for bridging resolu- tion. For instance, a singular bridging anaphor (e.g., "the book") can refer to a plural antecedent (e.g., "books"), and bridging relations can be formed from mentions with different entity types (e.g., "the house" and "the window"). In fact, while many coreference relations can be identified via string matching facilities, it is not uncommon for bridging relations to be identified using background knowledge and/or sophisticated inference mech- anisms. The complexity of bridging resolution is further complicated by the lack of a large cor- pus annotated with bridging relations: while the most extensively-used coreference-annotated cor- pus, OntoNotes, contain more than 2000 docu- ments, two of the most commonly-used corpora for bridging resolution, ISNotes and BASHI, each contains only 50 documents taken from OntoNotes. The current state of bridging resolution research is rather unsatisfactory. State-of-the-art bridging resolvers are typically evaluated in unrealistic set- tings: in the bridging anaphora resolution task, the goal is to identify the antecedent of a given bridging anaphor; and in the full bridging resolu- tion task, the goal is to first identify the bridging anaphors given a set of gold mentions and then resolve each anaphor to its antecedent (Yu and Poe- sio, 2020). While such unrealistic settings have been considered unacceptable for evaluating entity coreference …

Content truncated to 500 words