Abstract
Argument unit recognition and classification (AURC) is a promising and critical research topic in argument mining, which aims to extract the argument units that express support or opposing stance in a given argumentative text under controversial topics. Existing studies treated the AURC as a sequence labeling problem and designed a unified approach to predict argument unit boundary and argument unit stance simultaneously. In this paper, we propose a general framework hierarchical neural network (HNN) for AURC, by fusing two different approach: divide-and-conquer approach and unified approach. The divide-and-conquer approach considers the correlation of the two tasks inherent in AURC (task 1: argument unit recognition, AUR and task 2: argument unit classification, AUC), and jointly optimize them for prediction by a novel probability transition matrix. Finally, we used a token-level attention mechanism to efficiently fuse probability distributions obtained by our proposed divide-and-conquer approach and existing unified approach. Experimental results on two benchmark datasets demonstrate the effectiveness of our proposed framework.