Incorporating Context and External Knowledge for Pronoun Coreference Resolution

Proceedings of NAACL-HLT 2019, pages 872–881 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 872 Incorporating Context and External Knowledge for Pronoun Coreference Resolution Hongming Zhang♣∗, Yan Song♠, and Yangqiu Song♣ ♣Department of CSE, The Hong Kong University of Science and Technology ♠Tencent AI Lab hzhangal@cse.ust.hk, clksong@gmail.com, yqsong@cse.ust.hk Abstract Linking pronominal expressions to the correct references requires, in many cases, better anal- ysis of the contextual information and exter- nal knowledge. In this paper, we propose a two-layer model for pronoun coreference res- olution that leverages both context and exter- nal knowledge, where a knowledge attention mechanism is designed to ensure the model leveraging the appropriate source of external knowledge based on different context. Experi- mental results demonstrate the validity and ef- fectiveness of our model, where it outperforms state-of-the-art models by a large margin. 1 Introduction The question of how human beings resolve pro- nouns has long been of interest to both linguistics and natural language processing (NLP) communi- ties, for the reason that pronoun itself has weak semantic meaning (Ehrlich, 1981) and brings chal- lenges in natural language understanding. To ex- plore solutions for that question, pronoun coref- erence resolution (Hobbs, 1978) was proposed. As an important yet vital sub-task of the gen- eral coreference resolution task, pronoun coref- erence resolution is to find the correct reference for a given pronominal anaphor in the context and has been shown to be crucial for a series of downstream tasks (Mitkov, 2014), including machine translation (Mitkov et al., 1995), sum- marization (Steinberger et al., 2007), information extraction (Edens et al., 2003), and dialog sys- tems (Strube and M¨uller, 2003). Conventionally, people design rules (Hobbs, 1978; Nasukawa, 1994; Mitkov, 1998) or use fea- tures (Ng, 2005; Charniak and Elsner, 2009; Li et al., 2011) to resolve the pronoun coreferences. ∗This work was done during the internship of the first author in Tencent AI Lab. Figure 1: Pronoun coreference examples, where each example requires different knowledge for its resolution. Blue bold font refers to the target pronoun, where the correct noun reference and other candidates are marked by green underline and brackets, respectively. These methods heavily rely on the coverage and quality of the manually defined rules and features. Until recently, end-to-end solution (Lee et al., 2017) was proposed towards solving the general coreference problem, where deep learning mod- els were used to better capture contextual informa- tion. However, training such models on annotated corpora can be biased and normally does not con- sider external knowledge. Despite the great efforts made in this area in the past few decades (Hobbs, 1978; Mitkov, 1998; Ng, 2005; Rahman and Ng, 2009), pronoun corefer- ence resolution remains challenging. The reason behind is that the correct resolution of pronouns can be influenced by many factors (Ehrlich, 1981); many resolution decisions require reasoning upon different contextual and external knowledge (Rah- man and Ng, 2011), which is also proved in other NLP tasks (Song et al., 2017, 2018; Zhang et al., 2018). Figure …

Content truncated to 500 words