Adversarial Learning for Discourse Rhetorical Structure Parsing

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3946–3957 August 1–6, 2021. ©2021 Association for Computational Linguistics 3946 Adversarial Learning for Discourse Rhetorical Structure Parsing Longyin Zhang1,2, Fang Kong1,2∗, Guodong Zhou1,2 1. Institute of Artificial Intelligence, Soochow University, China 2. School of Computer Science and Technology, Soochow University, China lyzhang9@stu.suda.edu.cn {kongfang,gdzhou}@suda.edu.cn Abstract Text-level discourse rhetorical structure (DRS) parsing is known to be challenging due to the notorious lack of training data. Although re- cent top-down DRS parsers can better leverage global document context and have achieved cer- tain success, the performance is still far from perfect. To our knowledge, all previous DRS parsers make local decisions for either bottom- up node composition or top-down split point ranking at each time step, and largely ignore DRS parsing from the global view point. Ob- viously, it is not sufficient to build an entire DRS tree only through these local decisions. In this work, we present our insight on evaluat- ing the pros and cons of the entire DRS tree for global optimization. Specifically, based on recent well-performing top-down frameworks, we introduce a novel method to transform both gold standard and predicted constituency trees into tree diagrams with two color channels. Af- ter that, we learn an adversarial bot between gold and fake tree diagrams to estimate the generated DRS trees from a global perspective. We perform experiments on both RST-DT and CDTB corpora and use the original Parseval for performance evaluation. The experimental results show that our parser can substantially improve the performance when compared with previous state-of-the-art parsers. 1 Introduction As the main linguistic theory on discourse rhetor- ical structure (DRS), Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) describes an article as a discourse tree (DT). As illustrated in Figure 1, each leaf node of the tree corresponds to an Elementary Discourse Unit (EDU), and relevant leaf nodes are connected by relation and nuclear- ity (nucleus (N) or satellite (S)) tags to form high-layer discourse units (DUs), where the ∗Corresponding author [e1: In fact,] [e2: Budget indicated] [e3: it saw some benefit] [e4: to staying involved in these programs,] [e5: in which renters earn frequent-flier miles] [e6: and fliers can get car-rental discounts.] wsj_2394 e1 e2 e3 e4 e5 e6 Same-Unit (NN) Attribution (NS) List (NN) Elaboration (NS) Elaboration (NS) Figure 1: An example RST-style discourse tree. nucleus is considered more important than the satellite. Since the RST structure can well describe the organization of an article, it has been playing a central role in various down-stream tasks like summarization (Xu et al., 2020), text catego- rization (Ji and Smith, 2017), and so on. With the release of various discourse corpora, text-level DSR parsing has been drawing more and more attention in the last decade. However, since the corpus annotation is usually time-consuming, existing DRS corpora are much limited in size. For example, the English RST-DT (Carlson et al., 2001) corpus only contains 385 WSJ …

Content truncated to 500 words