Question answering in multi-party conversation typically focuses on exploring discourse structures or speaker-aware information but ignores the interaction between questions and conversations. To solve this problem, a new model which integrates various information is proposed. In detail, to hierarchically model the discourse structures, speaker-aware dependency of interlocutors and question-context information, the proposed model leverages above information to propagate contextual information, by exploiting graph convolutional neural network. Besides, the model employs a reasonable interaction layer based on attention mechanism to enhance the understanding of multi-party conversations by selecting more helpful information. Furthermore, the model is the first to pay attention to the explicit interaction between question and context. The experimental results show that the model outperforms multiple baselines, illustrating that the model can understand the conversations more comprehensively.

目前基于多方对话文本的自动问答任务侧重于探索对话结构信息或说话者角色信息, 忽视问题文本和对话文本的交互。针对这一问题, 提出一个融合多信息的全新模型。该模型使用图卷积神经网络, 对多方对话文本中的话语结构、说话者角色以及问题–上下文信息进行分层次建模, 并设计合理的基于注意力机制的交互层, 通过选择更有帮助的信息, 加强对多方对话文本的理解。此外, 该模型首次对问题和上下文间的显式交互给予关注。实验结果表明, 所提模型的性能优于多个基线模型, 实现对多方对话文本的深层次理解。