Bias-Rectified Multi-way Learning with Data Augmentation for Implicit Discourse Relation Recognition

Ziwei Zheng, Chao Liang, Wei Xiang, Bang Wang

Abstract: Implicit Discourse Relation Recognition (IDRR) is a challenging but vital task in discourse analysis that focuses on identifying and classifying the relation between two arguments without explicit connectives. Previous research has focused on constructing sophisticated argument representations or utilizing labels’ hierarchical information while neglecting the intrinsic prior relational bias, where some relation categories are significantly less frequent than others due to inherent linguistic properties. Furthermore, some works amalgamate Explicit Discourse Relation Recognition (EDRR) data with IDRR data to achieve data augmentation; however, the linguistic discrepancies between EDRR and IDRR data may mislead the relation recognition process. To address the prior relation bias and misleading from linguistic discrepancies, we propose a novel Bias-REctified Multi-way Learning (BREM) model, which incorporates prior probability knowledge to assist in understanding the discourse relations and employs a PLM-based method for pre-training on the EDRR dataset to achieve data augmentation without introducing misleading effects. Experiments on the PDTB 3.0 corpus show that BREM outperforms previous models, especially in identifying less frequent relation senses, highlighting the effectiveness of our proposed model. The source code of our proposed model is publicly available at https://github.com/zzzziwy/NLPCC2024-BREM_for_IDRR.