Discourse Parsing Enhanced by Discourse Dependence Perception

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 354–363 November 20–23, 2022. ©2022 Association for Computational Linguistics 354 Discourse Parsing Enhanced by Discourse Dependence Perception Yuqing Xing, Longyin Zhang, Fang Kong⇤, Guodong Zhou School of Computer Science and Technology, Soochow University, China {yq_xing,zzlynx}@outlook.com {kongfang,gdzhou}@suda.edu.cn Abstract In recent years, top-down neural models have achieved significant success in text-level dis- course parsing. Nevertheless, they still suffer from the top-down error propagation issue, es- pecially when the performance on the upper- level tree nodes is terrible. In this research, we aim to learn from the correlations in be- tween EDUs directly to shorten the hierarchi- cal distance of the RST structure to alleviate the above problem. Specifically, we contribute a joint top-down framework that learns from both discourse dependency and constituency parsing through one shared encoder and two independent decoders. Moreover, we also ex- plore a constituency-to-dependency conversion scheme tailored for the Chinese discourse cor- pus to ensure the high quality of the joint learn- ing process. Our experimental results on CDTB show that the dependency information we use well heightens the understanding of the rhetori- cal structure, especially for the upper-level tree layers. 1 Introduction According to the representative Rhetorical Struc- ture Theory (RST) (Mann and Thompson, 1988), a text can be presented as a hierarchical discourse tree (DT) built on a set of elementary discourse units (EDUs). Given a piece of text, RST-style discourse parsing identifies such a DT with EDUs serving as terminal nodes. Moreover, it labels the rhetorical relations and nuclearity attributes associ- ated with each non-terminal node of the DT. Due to its far-reaching effects on text understanding and downstream NLP applications, text-level discourse parsing has been drawing more and more attention in the past decade. From the early bottom-up approaches (Feng and Hirst, 2014; Ji and Eisenstein, 2014; Heil- man and Sagae, 2015; Li et al., 2016; Braud et al., 2017; Yu et al., 2018; Mabona et al., 2019) to ⇤Corresponding author the more recent top-down frameworks (Lin et al., 2019; Kobayashi et al., 2020; Zhang et al., 2020, 2021; Koto et al., 2021), previous studies gradu- ally switch from feature-based machine learning methods to deep neural models and have achieved particular success. Among current neural models, top-down parsers, in most cases, perform better than bottom-up ones due to their capability of cap- turing global context information. Nevertheless, due to the long-distance dependencies in between textual units and the notorious lack of training data, top-down text-level discourse parsing still faces the following possible bottlenecks: • At the initial parsing stage, top-down parsers consider each entire text to determine the upper- level DT nodes. However, the whole text segment usually consists of diverse information, too much for the machine to understand thoroughly. As a result, our experimental statistics show that the parsing performance decreases by about 30% when the DT level is greater than 5. • …

Content truncated to 500 words