TransS-Driven Joint Learning Architecture for Implicit Discourse Relation Recognition

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 139–148 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 139 TransS-Driven Joint Learning Architecture for Implicit Discourse Relation Recognition Ruifang He1,2,3, Jian Wang1,2, Fengyu Guo1,2∗, and Yugui Han1,2 1College of Intelligence and Computing, Tianjin University, Tianjin, China 2Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin, China 3State Key Laboratory of Cognitive Intelligence, iFLYTEK, China {rfhe,jian wang,fengyuguo,yghan}@tju.edu.cn Abstract Implicit discourse relation recognition is a challenging task due to the lack of connectives as strong linguistic clues. Previous methods primarily encode two arguments separately or extract the specific interaction patterns for the task, which have not fully exploited the anno- tated relation signal. Therefore, we propose a novel TransS-driven joint learning architecture to address the issues. Specifically, based on the multi-level encoder, we 1) translate discourse relations in low-dimensional embedding space (called TransS), which could mine the latent geometric structure information of argument- relation instances; 2) further exploit the seman- tic features of arguments to assist discourse un- derstanding; 3) jointly learn 1) and 2) to mutu- ally reinforce each other to obtain the better argument representations, so as to improve the performance of the task. Extensive experimen- tal results on the Penn Discourse TreeBank (PDTB) show that our model achieves compet- itive results against several state-of-the-art sys- tems. 1 Introduction Discourse relation describes how two adjacent text units (e.g., clauses, sentences, and larger sentence groups) are connected logically to one another. A discourse relation instance is usually defined as a connective taking two arguments (as Arg1 and Arg2, respectively). Implicit discourse rela- tion recognition without explicit connectives (Pitler et al., 2009) is still a challenging problem of dis- course analysis, which needs to infer the discourse relation from a specific context. It is beneficial to many downstream natural language processing (NLP) applications, such as machine translation (Meyer and Popescu-Belis, 2012) and text summa- rization (Gerani et al., 2014). ∗Corresponding author. The existing neural network-based models have shown great success in recognizing implicit dis- course relations. It mainly includes 1) Basic neural networks (Braud and Denis, 2015; Zhang et al., 2015; Liu et al., 2016) can learn the dense vector representations of discourse arguments, which can capture the semantic information to some extent. Further studies exploit different attention or mem- ory mechanisms (Liu and Li, 2016; Zhang et al., 2016) to capture the critical information of argu- ment pairs. 2) Complex neural models (Chen et al., 2016; Lei et al., 2017; Guo et al., 2018) utilize gated relevance networks or neural tensor networks to capture the deeper interactions between two dis- course arguments. 3) Joint learning architectures (Qin et al., 2017; Bai and Zhao, 2018; Xu et al., 2019) exploit implicit connective cues, different granularity of text, or topic-level relevant informa- tion to improve the discourse relation prediction. However, these approaches still have the following drawbacks: 1) do not make full use of the annotated discourse relation signal to explore the argument- relation features; 2) neglect …

Content truncated to 500 words