Annotating Mentions Alone Enables Efficient Domain Adaptation for Coreference Resolution

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10543–10558 July 9-14, 2023 ©2023 Association for Computational Linguistics Annotating Mentions Alone Enables Efficient Domain Adaptation for Coreference Resolution Nupoor Gandhi, Anjalie Field, Emma Strubell Carnegie Mellon University {nmgandhi, anjalief, estrubel}@cs.cmu.edu Abstract Although recent neural models for coreference resolution have led to substantial improve- ments on benchmark datasets, transferring these models to new target domains containing out-of-vocabulary spans and requiring differ- ing annotation schemes remains challenging. Typical approaches involve continued training on annotated target-domain data, but obtaining annotations is costly and time-consuming. We show that annotating mentions alone is nearly twice as fast as annotating full coreference chains. Accordingly, we propose a method for efficiently adapting coreference models, which includes a high-precision mention de- tection objective and requires annotating only mentions in the target domain. Extensive evaluation across three English coreference datasets: CoNLL-2012 (news/conversation), i2b2/VA (medical notes), and previously un- studied child welfare notes, reveals that our ap- proach facilitates annotation-efficient transfer and results in a 7-14% improvement in aver- age F1 without increasing annotator time1. 1 Introduction Neural coreference models have made substan- tial strides in performance on standard benchmark datasets such as the CoNLL-2012 shared task, where average F1 has improved by 20% since 2016 (Durrett and Klein, 2013; Dobrovolskii, 2021; Kirstain et al., 2021). Modern coreference archi- tectures typically consist of an encoder, mention detector, and antecedent linker. All of these com- ponents are optimized end-to-end, using only an antecedent linking objective, so expensive corefer- ence chain annotations are necessary for training (Aralikatte and Søgaard, 2020; Li et al., 2020a). These results have encouraged interest in deploy- ing models in domains like medicine and child pro- tective services, where a small number of practition- 1Code is available at https://github.com/ nupoorgandhi/data-eff-coref Figure 1: Model coreference performance (avg F1) as a func- tion of continued training on limited target domain data requir- ing varying amounts of annotator time. The source domain is news/conversation (OntoNotes) and the target domain is medi- cal notes (i2b2/VA). Using our method to adapt coreference models using only mentions in the target domain, we achieve strong coreference performance with less annotator time. ers need to quickly obtain information from large volumes of text (Uzuner et al., 2012; Saxena et al., 2020). However, successes over curated data sets have not fully translated to text containing technical vocabulary, frequent typos, or inconsistent syntax. Coreference models struggle to produce meaning- ful representations for new domain-specific spans and may require many examples to adapt (Uppunda et al., 2021; Lu and Ng, 2020; Zhu et al., 2021). Further, coreference models trained on standard benchmarks are not robust to differences in anno- tation schemes for new domains (Bamman et al., 2020). For example, OntoNotes does not annotate singleton mentions, those that do not corefer with any other mention. A system trained on OntoNotes would implicitly learn to detect only entities that appear more than once, even though singleton re- trieval is often desired in …

Content truncated to 500 words