Data Augmentation via Relation-Variant Discourse Recasting for Implicit Discourse Relation Recognition
Abstract
AbstractImplicit discourse relation recognition (abbr., IDRR) is a task of automatically classifying semantic relations between pairwise arguments in the absence of explicit connectives. IDRR suffers from imbalanced data distribution of different relations. The current data augmentation approaches have relieved this problem to some extent. However, the incorporated discourse samples during augmentation lack customized contrastive effects. To address this issue, we propose to construct topic-relevant relation-inconsistent arguments for augmentation. Specifically, we implement a Relation-Variant Discourse Recasting (RVDR) approach. RVDR utilizes Qwen3-4B to paraphrase discourses under the forcible constraints of preserving inherent topics and transforming relation senses toward infrequent target relations. To avoid excessive divergent thinking of RVDR, we optimize the Qwen3-4B via reinforcement learning, where Group Relative Policy Optimization (GRPO) is used along with multiple task-specific rewards. Furthermore, rule-based filtering and confidence estimation are used to ensure data quality, which additionally enable two application modes, i.e., Confidence-based and Progressive Data Augmentation (i.e., CDA and PDA). Experiments on PDTB 2.0 and 3.0 show that RVDR-based CDA consistently improves IDRR performance across multiple backbones, including both encoder-only and decoder-only cases. It also allows the strong IDRR frameworks CP-KD and SCDN to be enhanced further. The code is publicly available.