Annotation-Inspired Implicit Discourse Relation Classification with Auxiliary Discourse Connective Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 15696–15712 July 9-14, 2023 ©2023 Association for Computational Linguistics Annotation-Inspired Implicit Discourse Relation Classification with Auxiliary Discourse Connective Generation Wei Liu and Michael Strube Heidelberg Institute for Theoretical Studies gGmbH {wei.liu, michael.strube}@h-its.org Abstract Implicit discourse relation classification is a challenging task due to the absence of discourse connectives. To overcome this is- sue, we design an end-to-end neural model to explicitly generate discourse connectives for the task, inspired by the annotation process of PDTB. Specifically, our model jointly learns to generate discourse connectives between arguments and predict discourse relations based on the arguments and the generated connect- ives. To prevent our relation classifier from being misled by poor connectives generated at the early stage of training while alleviating the discrepancy between training and inference, we adopt Scheduled Sampling to the joint learning. We evaluate our method on three benchmarks, PDTB 2.0, PDTB 3.0, and PCC. Results show that our joint model significantly outperforms various baselines on three datasets, demonstrat- ing its superiority for the task. 1 Introduction Discourse relations, such as Cause and Contrast, describe the logical relation between two text spans (Pitler et al., 2009). Recognizing discourse rela- tions is beneficial for various NLP tasks, includ- ing coherence modeling (Lin et al., 2011), read- ing comprehension (Mihaylov and Frank, 2019), argumentation mining (Habernal and Gurevych, 2017; Hewett et al., 2019), and machine translation (Meyer, 2015; Longyue, 2019). Discourse connectives (e.g., but, as a result) are words or phrases that signal the presence of a dis- course relation (Pitler and Nenkova, 2009). They can be explicit, as in (1), or implicit, as in (2): (1) [I refused to pay the cobbler the full $95]Arg1 because [he did poor work.]Arg2 (2) [They put the treasury secretary back on the board.]Arg1 (Implicit=However) [There is doubt that the change would accomplish much.]Arg2 When discourse connectives are explicitly present between arguments, classifying the sense of a dis- course relation is straightforward. For example, Pitler and Nenkova (2009) proved that using only connectives in a text as features, the accuracy of 4-way explicit discourse relation classification on PDTB 2.0 can reach 85.8%. However, for implicit cases, there are no connectives to explicitly mark discourse relations, which makes implicit discourse relation classification challenging (Zhou et al., 2010; Shi et al., 2017). Existing work attempts to perform implicit discourse relation classification directly from arguments. They range from design- ing linguistically informed features from arguments (Lin et al., 2009; Pitler et al., 2009) to modeling in- teraction between arguments using neural networks (Lei et al., 2017; Guo et al., 2018). Despite their impressive performance, the absence of explicit dis- course connectives makes the prediction extremely hard and hinders further improvement (Lin et al., 2014; Qin et al., 2017). The huge performance gap between explicit and implicit classification (85.8% vs. 57.6%) (Liu and Li, 2016) motivates recent studies to utilize implicit connectives for the training process of implicit rela- tion classifiers. For …

Content truncated to 500 words