面向一致性对话生成的对抗匹配网络与目标侧注意力机制研究  

Research on towards coherent dialogue generation via adversarial matching network and target-side attention

作　　者：蔡恒毅 王成瑞 宋永浩 袁旭 张程[1] 赵晓芳[1] CAI Hengyi;WANG Chengrui;SONG Yonghao;YUAN Xu;ZHANG Cheng;ZHAO Xiaofang(Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049)

机构地区：[1]中国科学院计算技术研究所,北京100190 [2]中国科学院大学,北京100190

出　　处：《高技术通讯》2022年第2期131-142,共12页Chinese High Technology Letters

基　　金：国家自然科学基金(U1836111,U1736106);国家重点研发计划(2018YFB0904503)资助项目。

摘　　要：序列到序列(seq2seq)方法在开放域对话生成领域中备受研究学者的关注。然而,标准的序列到序列模型容易产生语义冲突和不连贯的对话回复,这种不一致性是现有系统生成的回复显著有别于人类真实对话的重要原因之一。对话生成中的一致性既包括回复内部的语义一致性,也包括上文与其回复之间的外部关联性。本文提出了一个新的对话生成框架,称为基于张量匹配的生成式对抗网络(MatchGAN),以提高对话回复与其上文之间的外部关联性。与传统的基于最大似然估计的方法不同,该框架通过基于序列到序列模型的生成器和基于张量匹配网络的判别器之间的对抗学习来生成与上文相关的回复。通过使用匹配网络对上文与回复之间的多维关系进行建模,该模型所产生的回复更加符合人类对话的特点。此外,本研究进一步引入了目标侧注意力机制来增强所产生回复的内部语义一致性。实验结果表明,本文提出的框架能够产生高质量的对话回复,在量化指标评价和人工评测方面均优于其他基线方法。The sequence-to-sequence(seq2seq)approach has received great attention in the field of open-domain dialogue generation.However,the standard seq2seq model is prone to generate meaningless and incoherent responses,making it distinguish clearly from the human-human conversations.This coherence includes both the internal consistency of the long response and the external relevance between the post and its response.This work proposes a novel dialogue generation framework called matching-based generative adversarial network(MatchGAN)to improve external relevance.Instead of imitating the ground truth with supervised learning,this model can generate post-relevant responses through the generative-adversarial learning with a seq2seq-based generator and a matching-based discriminator,allowing the generated human-machine conversation more like a human-human conversation by discriminating whether a response is matching with the post.Furthermore the target-side attention mechanism is introduced to maintain the internal consistency of the generated responses.This new framework is able to generate coherent responses with high quality.Experimental results show that the proposed model can achieve substantial improvements in both metric-based and human evaluations among various baselines.

关 键 词：生成式对话模型 神经张量网络 对抗学习 目标侧注意力机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]