Adapting Coreference Resolution for Processing Violent Death Narratives

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4553–4559 June 6–11, 2021. ©2021 Association for Computational Linguistics 4553 Adapting Coreference Resolution for Processing Violent Death Narratives Ankith Uppunda, Susan D. Cochran, Jacob G. Foster Alina Arseniev-Koehler, Vickie M. Mays, Kai-Wei Chang∗ University of California, Los Angeles Abstract Coreference resolution is an important compo- nent in analyzing narrative text from admin- istrative data (e.g., clinical or police sources). However, existing coreference models trained on general language corpora suffer from poor transferability due to domain gaps, especially when they are applied to gender-inclusive data with lesbian, gay, bisexual, and transgender (LGBT) individuals. In this paper, we an- alyzed the challenges of coreference resolu- tion in an exemplary form of administrative text written in English: violent death nar- ratives from the USA’s Centers for Disease Control’s (CDC) National Violent Death Re- porting System. We developed a set of data augmentation rules to improve model perfor- mance using a probabilistic data programming framework. Experiments on narratives from an administrative database, as well as existing gender-inclusive coreference datasets, demon- strate the effectiveness of data augmentation in training coreference models that can better handle text data about LGBT individuals. 1 Introduction Coreference resolution (Soon et al., 2001; Ng and Cardie, 2002) is the task of identifying denotative phrases in text that refer to the same entity. It is an essential component in Natural Language Pro- cessing (NLP). In real world applications of NLP, coreference resolution is crucial for analysts to ex- tract structured information from text data. Like all components of NLP, it is important that coreference resolution is robust and accurate, as applications of NLP may inform policy-making and other de- cisions. This is especially true when coreference systems are applied to administrative data, since results may inform policy-making decisions. In this paper, we describe an approach to adapt- ing a coreference model to process narrative text ∗kwchang@cs.ucla.edu from an important administrative database writ- ten in English: the National Violent Death Re- porting System (NVDRS), maintained by the Cen- ters for Disease Control (CDC) in the USA. Vio- lent death narratives document murders, suicides, murder-suicides, and other violent deaths. These narratives are complex, containing information on one or more persons; some individuals are vic- tims, others are partners (heterosexual or same-sex), family members, witnesses and law enforcement. Specifically, we apply the End-to-End Coreference Resolution (E2E-Coref) system (Lee et al., 2017, 2018), which has achieved high performance on the OntoNotes 5.0 (Hovy et al., 2006) corpus. We observe that when a model trained on OntoNotes is applied to violent death narratives, the performance drops significantly for the following reasons. First, despite the fact that OntoNotes contains multiple genres1, it does not include administra- tive data. Administrative text data is terse and contains an abundance of domain-specific jargon. Because of the gap between training and admin- istrative data, models trained on OntoNotes are poorly equipped to handle administrative data that are heavily skewed in vocabulary, …

Content truncated to 500 words