计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (3): 104-111.DOI: 10.3778/j.issn.1002-8331.2107-0421

• 模式识别与人工智能 • 上一篇    下一篇

联合对话行为识别与情感分类的多任务网络

林鸿辉,刘建华,郑智雄,胡任远,罗逸轩   

  1. 1.福建工程学院 计算机科学与数学学院(原信息科学与工程学院),福州 350118
    2.福建省大数据挖掘与应用技术重点实验室,福州 350118
  • 出版日期:2023-02-01 发布日期:2023-02-01

Multi-Task Network for Joint Dialog Act Recognition and Sentiment Classification

LIN Honghui, LIU Jianhua, ZHENG Zhixiong, HU Renyuan, LUO Yixuan   

  1. 1.School of Computer Science and Mathematics, Fujian University of Technology, Fuzhou 350118, China
    2.Fujian Provincial Key Laboratory of Big Data Mining and Applications, Fuzhou 350118, China
  • Online:2023-02-01 Published:2023-02-01

摘要: 情感分类与对话行为识别任务常被用在对话系统中联合建模,这种联合建模的方法能够挖掘出对话文本的跨任务隐藏交互信息,提高情感分类与对话行为识别的准确性。然而,在两个任务联合建模时,对话文本有上下文信息和跨任务交互信息,对这两种信息与模型预测准确度之间的关系进行研究,需要将这两种信息协同建模,在过去的研究中对这两种信息的利用并不充分,为解决该问题提出多任务图注意力网络(multi-task graph attention network,MGAT),并且以其为核心模块搭建了多任务协同图注意力网络(multi-task synergic graph attention network,MSGAT),该模型将上下文信息与跨任务信息联合建模,同时完成情感分类与对话行为识别任务。利用两个公开数据集实验,得到了良好的效果,并且对联合模型与预训练模型组合进行了研究。

关键词: 多任务学习, 情感分类, 对话行为别, 多任务图注意力网络, 深度学习

Abstract: Dialog act recognition and sentiment classification are two interrelated tasks in dialog system, so these tasks should be jointly modeling and excavate implicit interactive information for improving the accuracy of tasks. However, the dialog contextual information and the mutual interaction information proves two crucial?factor of dialog text. The dialog contextual information shows the relation of target information with contextual utterance. The mutual interaction information shows influence of target in different task. In past studies, both types of information are underutilized. In order to solve this problem, a multi-task graph attention network(M-GAT) is proposed and the model multi-task synergic graph attention network(MSGAT) is built to integrate the dialog contextual information and the mutual interaction information. Experimental results on two public datasets show that this model successfully acquires the two sources of information and achieves the state-of-the-art performance. In addition, this paper combines MSGAT and pre-trained models and achieves well result.

Key words: multi-task learning, sentiment classification, dialog act recognition, multi-task graph attention network, deep learning