Automatic Data Acquisition for Event Coreference Resolution

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 1185–1196 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 1185 Automatic Data Acquisition for Event Coreference Resolution Prafulla Kumar Choubey and Ruihong Huang Department of Computer Science and Engineering Texas A&M University (prafulla.choubey, huangrh)@tamu.edu Abstract We propose to leverage lexical paraphrases and high precision rules informed by news dis- course structure to automatically collect coref- erential and non-coreferential event pairs from unlabeled English news articles. We perform both manual validation and empirical evalua- tion on multiple evaluation datasets with differ- ent event domains and text genres to assess the quality of our acquired event pairs. We found that a model trained on our acquired event pairs performs comparably as the supervised model when applied to new data out of the training data domains. Further, augmenting human-annotated data with the acquired event pairs provides empirical performance gains on both in-domain and out-of-domain evaluation datasets. 1 Introduction Event coreference resolution aims to determine and cluster event mentions that refer to the same real- world event. It is a relatively less studied NLP task despite being crucial for various NLP applica- tions such as topic detection and tracking, question answering, and summarization. A typical event coreference resolution system relies on scoring similarity between two event men- tions in a document followed by clustering. How- ever, event coreference chains are very sparsely dis- tributed and only certain key events are repeated in a document, which makes manually labeling many event coreference relations very time-consuming. Furthermore, event mentions tend to appear in ex- tremely diverse contexts and few are accompanied by a full set of their arguments. The two chal- lenges, the absence of abundant human-annotated event coreference data and the high diversity of contexts containing coreferential event mentions, make it hard to build effective event coreference resolution systems. We aim to improve the effectiveness of event coreference resolution systems by automatically acquiring coreferential event pairs from many doc- uments requiring minimal supervision. Specifi- cally, coreferential event mentions are associated with discourse function of sentences in a news doc- ument (Choubey et al., 2020)1. We propose to use them to identify sentence pairs that are likely to contain coreferential event mentions as well as sentence pairs that are likely to contain non- coreferential event pairs. Consider the two exam- ple sentence pairs below, each pair having an event pair with synonymous trigger words. (1): [People living in absolute poverty in rural areas of the eight regions and provinces reduced to 14.52 million from 30.76 million over the last decade.] [Yang admitted , however , that ethnic minority regions still lagged far behind the developed eastern regions and the government still faced serious challenges to reduce poverty.] (2): [At least 30,000 war-displaced people camped in Angola’s central province of Kwanza-sul are being reset- tled in productive areas, the official news agency angop reported here on Friday.] [The resettlement is being car- ried out jointly by the local municipal …

Content truncated to 500 words