基于循环交互注意力网络的问答立场分析

For answer stance analysis task, most existing methods are difficult to extract the significant dependency between questions and answers. To this end, this paper proposes a novel method for answer stance analysis based on a recurrent interactive attention (RIA) network. By imitating the human-like learning method, the proposed model exploits the interactive attention mechanism and the recurrent training iteration for answer stance analysis, which can effectively extracts the dependency between the question and then derives the representation of the stance according to the contextual information of the answer. In addition, to address the problem that the problem text cannot clearly express the corresponding stance, the proposed method presents a novel way of enhancing the representations of the question sentences via switching the question expressions into statements. Finally, the experimental results on the Chinese social media question-answer dataset show that the proposed method achieves the state-of-the-art performance. It also verifies effectiveness of our method in extracting the dependency between questions and answers for answer stance analysis task.