Multi-view Hierarchical Graph Neural Network for Argumentation Mining

Yang Sun, Jianzhu Bao, Geng Tu, Bin Liang, Min Yang, Ruifeng Xu

Abstract: Argumentation mining (AM) aims to detect the arguments and their relations from argumentative texts. Generally, AM contains three key challenging subtasks, including argument component type classification (ACTC), argumentative relation identification (ARI), and argumentative relation type classification (ARTC). Most previous studies solve these three subtasks separately, neglecting the rich interrelation information among the three tasks. In this paper, we propose a multi-view hierarchical graph neural network (MHGNN) for AM, which resolves the three interacted subtasks in a unified multi-task learning framework. Concretely, MHGNN learns graph embeddings from multiple views (i.e., word view and semantic view) that often provide more comprehensive information. Each graph view is equipped with a two-level graph structure: (i) the first level is the argumentation graph with each argumentation component (AC) as a graph node, which learns the inter-AC knowledge from the input text; (ii) the second level is the AC graph with each word or semantic role as graph node respectively, which learn the fine-grained intra-AC knowledge within each AC from the word level or semantic level. The multi-view hierarchical GNN makes our model more effective to utilize the rich information among and within the ACs in the input text. Then, we transform ACTC, ARI, and ARTC into node classification, edge prediction, and edge type classification on the argumentation graph by devising novel graph attention mechanisms to learn comprehensive and relation-aware graph embeddings. These three subtasks are integrated into a unified model through multi-task learning and partial parameters sharing. Extensive experiments on two benchmark datasets demonstrate that the proposed MHGNN framework outperforms the strong baseline methods for all three subtasks.