A Simple Unsupervised Approach for Coreference Resolution using Rule-based Weak Supervision

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 79 - 88 July 14-15, 2022 ©2022 Association for Computational Linguistics A Simple Unsupervised Approach for Coreference Resolution using Rule-based Weak Supervision Alessandro Stolfo1 Chris Tanner2 Vikram Gupta3 Mrinmaya Sachan1 1ETH Zürich 2Harvard University 3Sharechat {alessandro.stolfo, mrinmaya.sachan}@inf.ethz.ch christanner@g.harvard.edu vikramgupta@sharechat.co Abstract Labeled data for the task of Coreference Res- olution is a scarce resource, requiring signifi- cant human effort. While state-of-the-art coref- erence models rely on such data, we propose an approach that leverages an end-to-end neu- ral model in settings where labeled data is unavailable. Specifically, using weak super- vision, we transfer the linguistic knowledge encoded by Stanford’s rule-based coreference system to the end-to-end model, which jointly learns rich, contextualized span representa- tions and coreference chains. Our experiments on the English OntoNotes corpus demonstrate that our approach effectively benefits from the noisy coreference supervision, producing an improvement over Stanford’s rule-based sys- tem (+3.7 F1) and outperforming the previous best unsupervised model (+0.9 F1). Addition- ally, we validate the efficacy of our method on two other datasets: PreCo and Litbank (+2.5 and +5 F1 on Stanford’s system, respectively). 1 Introduction Coreference resolution is an important problem in language understanding. In the recent years, sig- nificant progress has been made on this task with coreference annotated corpora (Hovy et al., 2006) and deep neural network architectures (Wiseman et al., 2015; Clark and Manning, 2016a,b; Lee et al., 2017). Further gains have been obtained by lever- aging contextualized text encoders like ELMo (Lee et al., 2018), BERT, SpanBERT, and Longformer (Kantor and Globerson, 2019; Joshi et al., 2019, 2020; Wu et al., 2020; Kirstain et al., 2021). The progress in supervised coreference resolu- tion has not been accompanied by analogous im- provements in unsupervised methods. The best performing work in this domain is the unsuper- vised mention-ranking systems proposed by Ma et al. (2016). Approaches that do not rely on gold annotation are highly desirable for this task, as coreference corpora are expensive to create. Ad- dressing this issue, weak supervision has been used for multilingual coreference resolution to automati- cally obtain labels for languages with no annotated datasets (Wallin and Nugues, 2017). In this paper, we introduce a simple yet effec- tive approach for unsupervised coreference resolu- tion, which leverages an end-to-end span-ranking coreference model (Lee et al., 2018) and contextu- alized span representations. The end-to-end model is trained with weak supervision from Stanford’s coreference system (Lee et al., 2011), which, in turn uses a set of linguistic rules for coreference. Previous works have used Stanford system’s rules as feature extractors (Fernandes et al., 2012; Wise- man et al., 2015; Ma et al., 2016). However, our approach uses Stanford’s rule-based sieves to pro- duce noisy labels that are subsequently used to train the neural end-to-end resolver. The rationale behind the use of Stanford’s re- solver for producing noisy labels lies in its ease of use and its modular structure, which allows us to interpret the value of the linguistic knowledge …

Content truncated to 500 words