基于多视角对抗学习的开放域对话生成模型  

Open domain dialogue generation model based on multi-view adversarial learning

作　　者：张凉 杨燕[1] 陈成才 贺樑[1] Zhang Liang;Yang Yan;Chen Chengcai;He Liang(School of Computer Science&Technology,East China Normal University,Shanghai 200062,China;Shanghai Xiao’i Robot Technology Co.Ltd,Shanghai 201803,China)

机构地区：[1]华东师范大学计算机科学与技术学院,上海200062 [2]上海智臻智能网络科技股份有限公司小i机器人研究院,上海201803

出　　处：《计算机应用研究》2021年第2期372-376,共5页Application Research of Computers

基　　金：上海市教科委重点项目(18511105502)。

摘　　要：近年来,随着智能家居的普及,对话系统在生活中发挥着越来越重要的作用,基于神经网络构建的生成式对话系统由于其灵活性高受到了许多研究者的关注。以提高生成模型对话的流畅性、上下文相关性为目的,提出基于多视角对抗学习的开放域对话生成模型。其中,模型生成器是基于检索到的相似对话进行改写得到生成的对话;模型的判别器是由两个二分类器共同组成的,该二元判别器分别从句子、对话两个层面多视角地对生成句子进行判别。在中文对话语料上进行实验,该模型在人工评价和自动评测上的得分都高于目前常用的对话生成模型。实验结果表明,利用二元判别器多视角训练可以同时提高生成回复的流畅度和上下文相关性。Recently,with the emergence and popularity of intelligent applications,non-task oriented dialogue system has played an increasingly important role in daily life.Generation-based dialogue systems receive extraordinary attention of some researchers because they are more flexible.In order to improve the fluency and contextual relevance of the responses generated by models,this paper proposed an open domain dialogue generation model based on binary discriminator in terms of a multi-view adversarial learning framework.The generator of the model rewrote a retrieved response to get a generated response.While the discriminator was composed of two binary classifiers and distinguished the human-generated responses from machine-generated ones.Experiments on a Chinese dialogue corpus show that the model has higher scores on both human and automatic evaluation than baselines.Experiments also show that multi-view training with binary discriminators can improve both the fluency and contextual relevance of the generated responses.

关 键 词：对话生成 对话系统 对抗学习 改写模型 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]