Exophoric Pronoun Resolution in Dialogues with Topic Regularization

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3832–3845 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 3832 Exophoric Pronoun Resolution in Dialogues with Topic Regularization Xintong Yu1, Hongming Zhang2, Yangqiu Song2, Changshui Zhang1, Kun Xu3, and Dong Yu3 1Institute for Artificial Intelligence, Tsinghua University (THUAI); 1Department of Automation, Tsinghua University, Beijing, P.R.China 2Department of CSE, The Hong Kong University of Science and Technology 3Tecent AI lab yuxt16@mails.tsinghua.edu.cn, zcs@mail.tsinghua.edu.cn, {hzhangal, yqsong}@cse.ust.hk, {kxkunxu, dyu}@tencent.com Abstract Resolving pronouns to their referents has long been studied as a fundamental natu- ral language understanding problem. Previ- ous works on pronoun coreference resolution (PCR) mostly focus on resolving pronouns to mentions in text while ignoring the exophoric scenario. Exophoric pronouns are common in daily communications, where speakers may di- rectly use pronouns to refer to some objects present in the environment without introduc- ing the objects first. Although such objects are not mentioned in the dialogue text, they can of- ten be disambiguated by the general topics of the dialogue. Motivated by this, we propose to jointly leverage the local context and global topics of dialogues to solve the out-of-text PCR problem. Extensive experiments demon- strate the effectiveness of adding topic regular- ization for resolving exophoric pronouns. 1 Introduction Grounding pronouns to objects they refer to is a challenging yet crucial natural language under- standing problem. The coreference relationship between a pronoun and its referents is categorized into endophora and exophora based on whether the referred objects appear in text or out of text, and the former case can be further divided into anaphora if the referents appear in the preceding text of the pronoun and cataphora if in the follow- ing text (Halliday and Hasan, 1976; Brown and Yule, 1983). Conventional studies on the pronoun coreference resolution (PCR) task in the NLP com- munity mainly focus on anaphora (Hobbs, 1978; NIST, 2003; Pradhan et al., 2012) and some re- cent work analyzes cataphora in machine transla- tion (Wong et al., 2020), while mostly ignoring the exophoric pronouns. However, in daily dia- logues or conversations, speakers may often use exophoric pronouns to refer to objects in the situ- ational context that all speakers and listeners are aware of without introducing them in the first place. Meal √ Topic: cooking and eating Cleaning service? I want to cook but the kitchen is in a mess. When is the house cleaning service scheduled? It is scheduled at 8 p.m.. Well, I am starving right now. Could you order that like last Friday? Figure 1: An example of resolving exophoric pro- nouns in daily dialogues with and without the help of dialogue topics. This limits the use of current PCR models in many real-world dialogue/conversation scenarios, e.g., text interpretation (Hankamer and Sag, 1976; Yule, 1979) and downstream tasks such as dialogue gen- eration (Kottur et al., 2018; Niu et al., 2019). Figure 1 shows an example of exophora. A per- son talks with his AI assistant (Siri/Alexa), “Could you order that like last …

Content truncated to 500 words