Improving Event Coreference Resolution Using Document-level and Topic-level Information

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6765 - 6775 December 7-11, 2022 ©2022 Association for Computational Linguistics Improving Event Coreference Resolution Using Document-level and Topic-level Information Sheng Xu, Peifeng Li and Qiaoming Zhu School of Computer Science and Technology Soochow University sxu@stu.suda.edu.cn, {pfli, qmzhu}@suda.edu.cn Abstract Event coreference resolution (ECR) aims to cluster event mentions that refer to the same real-world events. Deep learning methods have achieved SOTA results on the ECR task. How- ever, due to the encoding length limitation, pre- vious methods either adopt classical pairwise models based on sentence-level context or split each document into multiple chunks and en- code them separately. They failed to capture the interactions and contextual cues among those long-distance event mentions. Besides, high- level information, such as event topics, is rarely considered to enhance representation learning for ECR. To address the above two issues, we first apply a Longformer-based encoder to ob- tain the document-level embeddings and an en- coder with a trigger-mask mechanism to learn sentence-level embeddings based on local con- text. In addition, we propose an event topic generator to infer the latent topic-level represen- tations. Finally, using the above event embed- dings, we employ a multiple tensor matching method to capture their interactions at the docu- ment, sentence, and topic levels. Experimental results on the KBP 2017 dataset show that our model1 outperforms the SOTA baselines. 1 Introduction Within-document Event Coreference Resolution (ECR) is the task of grouping the event mentions (i.e., triggers) that occur in a document into clus- ters such that each cluster represents a unique real- world event. ECR is essential for information aggregation and can help many downstream ap- plications, such as abstractive summarization (Li et al., 2016), central event detection (Choubey et al., 2018), and discourse parsing (Lee et al., 2020). Although there studies employ clustering meth- ods (Chen and Ji, 2009; Phung et al., 2021), most work converts ECR into an event-pair classification 1Code is available at https://github.com/jsksxs360/ event-coref-emnlp2022 task (Krause et al., 2016; Huang et al., 2019; Lu and Ng, 2021a), i.e., judging whether two event mentions are coreferent. Therefore, event encoding becomes an essential step in ECR. Earlier work focuses on extracting hand-crafted features (Chen et al., 2009; Cybulska and Vossen, 2015; Lu and Ng, 2018), while recent studies encode triggers (i.e., the words that most clearly describe the event), con- textual linguistic cues (Nguyen et al., 2016; Huang et al., 2019; Lai et al., 2021), and identified ar- guments (i.e., the participants of a specific event) (Zeng et al., 2020) using various neural models. Recently, benefiting from the development of the Transformer-based models, event encoding (Krause et al., 2016; Huang et al., 2019; Lu et al., 2022; Lu and Ng, 2021a) has also changed from using CNN/RNN to BERT, SpanBERT, obtaining bet- ter event representations. However, at least two limitations exist in the above studies. First, due to the limitation of encoding length, current studies either adopt variant classical pair- wise models based on …

Content truncated to 500 words