ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (9): 1915-1924.doi: 10.7544/issn1000-1239.2021.20200634

• 人工智能 • 上一篇    下一篇



  1. (哈尔滨工业大学社会计算与信息检索研究中心 哈尔滨 150001) (
  • 出版日期: 2021-09-01
  • 基金资助: 

Research on Document Grounded Conversations

Sun Runxin, Ma Longxuan, Zhang Weinan, Liu Ting   

  1. (Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, Harbin 150001)
  • Online: 2021-09-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (62076081, 61772153, 61936010) and 2030 Major Project of New Generation Artificial Intelligence of China (2020AAA0108605).

摘要: 基于文档的对话是目前对话领域一个新兴的热点任务.与以往的任务不同,其需要将对话信息和文档信息综合进行考虑.然而,先前的工作着重考虑二者之间的关系,却忽略了对话信息中的句子对回复生成的作用具有差异性.针对这一问题,提出了一种新的辩证看待对话历史的方法,在编码阶段讨论利用历史和忽略历史2种情况进行语义信息的建模,并采用辩证整合的方式进行分支信息的汇总.由此避免了在历史信息与当前对话不相关时,其作为噪声被引入进而损害模型性能,同时也强化了当前对话对信息筛选的指导作用.实验结果表明,该模型与现有基线模型相比,能够生成更为符合当前语境且信息量更加丰富的回复,从而说明其能够更好地理解对话信息并进行知识筛选.并且通过进行消融实验,也验证了各模块在建模过程中的有效性.

关键词: 基于文档的对话, 回复生成, 信息筛选, Transformer模型, 注意力机制

Abstract: Document grounded conversations is an emerging hot task in the field of dialogue system. Different from previous tasks, it needs to consider both the utterances and the given document. However, previous work focused on the relationship between the two, but ignored the utterances’ difference in the effect of response generation. To solve this problem, a new dialectical approach to the dialogue history, which means the utterances before the last one, is proposed in this paper. At the encoding step, it divides the modeling of the semantic information into two parts: using history and ignoring history, and then uses the comparative integration method to summarize the branch results. In this way, when the dialogue history is not related to the current utterance, it can avoid being introduced as noise which will damage the performance of the model. Besides, it also strengthens the guiding role of the current utterance in the information filtering process. Experimental results show that compared with the existing baselines, this model can generate responses that are more in line with the current context and more informative, indicating that it can better understand dialogue information and conduct knowledge filtering. And through the ablation study, the effectiveness of each module in the modeling process is also verified.

Key words: document grounded conversations, reply generation, information filtering, Transformer model, attention mechanism