基于话语主题的对话文本摘要

Dialogue Text Abstract Based on Discourse Theme

对话文本摘要任务是从一段拥有两位及以上参与者之间的交流话语中提炼出精简的内容,以便他人可快速了解对话的全过程。相对于传统的新闻式文本,对话文本通常具有结构复杂、信息来源混乱等难点。因此,传统的文本摘要模型并不能适配对话文本的结构,无法生成高质量的摘要内容。为此,文章提出了一种针对对话文本结构所改进的摘要生成方法,通过解析对话文本中的话语、说话人、话语主题三个元素,构建对话结构图。使用微调的Bi-LSTM对对话结构图中的节点以词为单位编码,通过基于Transformer模型的异构图编码器对对话结构图进行图级编码,使用带有注意力机制和指针网络的解码器实现摘要内容的生成。主要解决对话文本摘要中出现的信息来源混乱和人称指代错误问题。实验结果显示,本文的模型在生成摘要的质量上有一定提高。

The task of dialogue text summary is to extract concise content froma conversation between two or more participants,so that others can quickly understand the whole process of dialogue.Compared with traditional news text,dialogue text usually has difficulties such as complex structure and confusion of information sources.Therefore,the traditional text summary model cannot adapt to the structure of the dialogue text and cannot generate high-quality summary content.To this end,this paper proposes an improved summary generation method for the structure of the dialogue text,which constructs the dialogue structure diagramby analyzing the three elements of the discourse,the speaker and the discourse theme in the dialogue text.The fine-tuned Bi-LSTM is used to encode the nodes in the dialog structure diagram in terms of words,the heterogeneous graph encoder based on the Transformer model is used to encode the dialog structure diagramat the graph level,and the decoder with attention mechanism and pointer network is used to generate the summary content.It mainly solves the problems of confusion of information sources and personal reference errors in the dialogue text summary.The experimental results showthat the model in this paper improves the quality of the summary.