Tackling Zero Pronoun Resolution and Non-Zero Coreference Resolution Jointly

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 518–527 November 10–11, 2021. ©2021 Association for Computational Linguistics 518 Tackling Zero Pronoun Resolution and Non-Zero Coreference Resolution Jointly Shisong Chen1, Binbin Gu3, Jianfeng Qu1, Zhixu Li2∗, An Liu1, Lei Zhao1, Zhigang Chen4 1School of Computer Science and Technology, Soochow University, Suzhou, China 2Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University 3University of California, Irvine 4IFLYTEK Research, Suzhou, China sschen777@stu.suda.edu.cn, binbing@uci.edu, zhixuli@fudan.edu.cn {jfqu,anliu,zhaol}@suda.edu.cn, zgchen@iflytek.com Abstract Zero pronoun resolution aims at recogniz- ing dropped pronouns and pointing out their anaphoric mentions, while non-zero corefer- ence resolution targets at clustering mentions referring to the same entity. Existing efforts often deal with the two problems separately regardless of their close essential correlations. In this paper, we investigate the possibility of jointly solving zero pronoun resolution and coreference resolution via a novel end-to-end neural model. Specifically, we design a gap- masked self-attention model that encodes gaps and tokens in the same space, where gaps could capture valuable contextual information according to their surrounding tokens while tokens could maintain original sequential in- formation without disturbance. Additionally, we also propose a two-stage interaction mech- anism to make full use of the exclusive re- lationship between zero pronouns and men- tions. Our empirical study conducted on the OntoNotes 5.0 Chinese dataset shows that our model could outperform corresponding state- of-the-art approaches on both tasks. 1 Introduction Zero pronoun resolution and non-zero coreference resolution are two fundamental tasks in natural lan- guage processing (NLP). Zero pronoun resolution, which is only studied in pro-drop languages such as Chinese, aims at recognizing dropped pronouns in a given text and pointing out their anaphoric mentions within the text (Chen and Ng, 2013). Coreference resolution, which is studied in all lan- guages, targets at clustering mentions referring to the same real-world entity in the text (Sukthanker et al., 2020). Both of the two tasks are vital for many downstream NLP applications including ma- chine translation (Mitkov et al., 1995), information extraction (Zelenko et al., 2004) and text summa- rization (Steinberger et al., 2007). ∗Zhixu Li is the corresponding author Given their importance, both tasks have been studied extensively. For zero pronoun resolution, most previous works assume that positions of zero pronouns are given. They encode zero pronouns and their candidate antecedents by LSTM (Yin et al., 2017), Attention (Yin et al., 2018b; Liu et al., 2017) or BERT (Song et al., 2020; Aloraini and Poesio, 2020), then measure the similarity between embeddings to find out the best antecedents. How- ever, positions of zero pronouns are usually un- known in practice. To solve this problem, some recent works (Yang et al., 2019; Song et al., 2020) treat spaces between tokens as candidate zero pro- nouns, recognize and resolve zero pronouns based on gap embeddings jointly. For non-zero coref- erence resolution, the state-of-the-art models be- long to an end-to-end paradigm or its variants. In these models, all spans in the text are candidate mentions. After representing spans by …

Content truncated to 500 words