Meta-learning based instance manipulation for implicit discourse relation recognition

Abstract

Without discourse connectives, implicit discourse relations recognition (IDRR) remains a challenging task and has attracted increasing attention. However, most studies ignore the issues of class imbalance and data noise. To alleviate these two problems, in this paper, we propose to improve the robustness of IDRR models with two novel meta-learning based instance manipulation networks: Meta-Weight-Net (MW-Net) and Meta-Label-Net (ML-Net), which adaptively control the effect of training instances and smooth the label distributions during model training, respectively. Specifically, we use MW-Net to adaptively learn an instance weight function directly from data and integrates the instance weights into the objective function of IDRR. Meanwhile, we adopt ML-Net to dynamically refine the label distributions, leading to the smaller variance in updating gradients. We conduct experiments on the PDTB 2.0 corpus. Experimental results and in-depth analysis empirically demonstrate the effectiveness of our networks.