Adapting Coreference Resolution for Narrative Processing

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2262–2267, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Adapting Coreference Resolution for Narrative Processing Quynh Ngoc Thi Do1, Steven Bethard2, Marie-Francine Moens1 1Katholieke Universiteit Leuven, Belgium 2University of Alabama at Birmingham, United States quynhngocthi.do@cs.kuleuven.be bethard@cis.uab.edu sien.moens@cs.kuleuven.be Abstract Domain adaptation is a challenge for super- vised NLP systems because of expensive and time-consuming manual annotated re- sources. We present a novel method to adapt a supervised coreference resolution system trained on newswire to short narra- tive stories without retraining the system. The idea is to perform inference via an In- teger Linear Programming (ILP) formula- tion with the features of narratives adopted as soft constraints. When testing on the UMIREC1 and N22 corpora with the-state- of-the-art Berkeley coreference resolution system trained on OntoNotes3, our infer- ence substantially outperforms the original inference on the CoNLL 2011 metric. 1 Introduction Coreference resolution is the task of partitioning the set of mentions of discourse referents in a text into classes (or ‘chains’) corresponding to those referents (Stede, 2011). To solve the problem, con- textual and grammatical clues, as well as semantic information and world knowledge are necessary for either learning-based (Bengtson and Roth, 2008; Stoyanov et al., 2010; Haghighi and Klein, 2010) or rule-based (Haghighi and Klein, 2009; Lee et al., 2011) coreference systems. These systems draw on diverse information sources and complex heuristics to resolve pronouns, model discourse, determine anaphoricity, and identify semantically compati- ble mentions. However, this leads to systems with many hetorogenous parts that can be difficult to interpret or modify. Durrett and Klein (2013) propose a learning- based, mention-synchronous coreference system to 1http://dspace.mit.edu/handle/1721.1/57507 2http://dspace.mit.edu/handle/1721.1/85893 3https://catalog.ldc.upenn.edu/LDC2011T03 tackle the various aspects of coreference by using the simplest possible set of features. Its advantage is that the system can both implicitly model impor- tant linguistic effects and capture other patterns in the data that are not easily teased out by hand. With a simple set of features including head/first/last words, preceding/following words, length, exact string match, head match, sentence/mention dis- tance, gender, number etc. and an efficient training using conditional log-likelihood augmented with a parameterized loss function optimization they report state-of-the-art results on CoNLL 2011 data. But while CoNLL 2011 training data (OntoNotes) includes a few different source domains (newswire, weblogs, etc.), we witness significant drops in performance when systems trained on CoNLL 2011 are applied to new target domains such as narratives. Some linguistic effects and patterns that are very important for the target domain were never seen in the source domain on which the model was trained. In such cases, when adapting a coreference system to a new domain, it is necessary to incorporate these more complex linguistic features and patterns into the model. We propose a novel method to adopt the tar- get domain’s features to a supervised coreference system without retraining the model. We present a case of transferring the system of (Durrett and Klein, 2013), which is trained on OntoNotes, to short …

Content truncated to 500 words