Discourse Component Recognition via Graph Neural Network in Chinese Student Argumentative Essays

Identifying and classifying original discourse components is a prerequisite task for constructing the knowledge graph structure. Previous work suffers from the following problems. (i) Existing methods only rely on the discourse components themselves to extract the features of the text, and do not fully take into account the potential help of the context information of the discourse to consummate its own features. (ii) Most of the current methods usually combine multiple tasks for joint inference, with complementary effects among them, lacking methods focusing on the single-target task of discourse component recognition. To address these issues, we propose a graph neural network-based discourse component recognition model (DCRGNN), which enhances the interaction of sentence-level discourse component features through graph structure. Our experimental results show that DCRGNN achieves a relative improvement of up to 6% on Macro-F1 for specific discourse component types compared to the previous state-of-the-art methods on the Chinese dataset, and exceeds the baseline model in the single-target task on the English dataset.