面向App评论响应的语义检索和生成框架  

A Semantic Retrieval and Generation Framework for App Review Response

作　　者：范国栋 陈世展[1] 肖建茂 吴洪越 张璐[1] 薛霄 王忠杰[3] 冯志勇[1] FAN Guo-Dong;CHEN Shi-Zhan;XIAO Jian-Mao;WU Hong-Yue;ZHANG Lu;XUE Xiao;WANG Zhong-Jie;FENG Zhi-Yong(College of Intelligence and Computing,Tianjin University,Tianjin 300350;School of Software,Jiangxi Normal University,Nanchang 330022;Faculty of Computing,Harbin Institute of Technology,Harbin 150001)

机构地区：[1]天津大学智能与计算学部,天津300350 [2]江西师范大学软件学院,南昌330022 [3]哈尔滨工业大学计算学部,哈尔滨150001

出　　处：《计算机学报》2022年第12期2528-2543,共16页Chinese Journal of Computers

基　　金：国家自然科学基金重点基金(61832014,62032016);国家自然科学基金(61972276,62102281);江西省教育厅科技攻关项目(GJJ210338)资助.

摘　　要：在应用程序维护过程中,移动应用(Mobile Application,App)评论的响应为应用程序开发者提供了用户反馈机制,对应用的评级产生积极影响.为了减轻响应大量用户评论的工作负担,开发者通常采用自动化的机制回复评分或跟进用户问题.当前流行使用序列到序列(Sequence to Sequence,Seq2seq)的深度生成模型或融合信息检索的方法来生成用户评论的响应.然而,现有检索方法没有考虑句子的语义相似性,生成模型没有考虑用户评论与检索到评论之间的差异,导致模型对知识的利用不佳,降低了响应质量.为了解决这些问题,本文提出了一种面向App评论响应的语义检索和生成框架(A Semantic Retrieval and Generation Framework,SRGen).首先,基于响应相似但评论不一定相似的现象,通过自监督学习方法对Sentence-BERT(SBERT)模型微调.然后,利用SBERT获得名称、评分、评论信息的向量表示,检索知识库中Top-k最相似的评论-响应对.最后,根据检索到的评论与待响应评论的差异和相应响应内容,生成评论的响应.实验表明,与现有的基线工作相比,SRGen在BLEU指标下提升了12.4%,在ROUGE指标下提升了9.4%.In the process of application maintenance,the response of mobile application(App)reviews provides developers with a user feedback mechanism,which has a positive impact on the rating of the application.To reduce the workload of responding to a large number of user comments,developers usually adopt an automated mechanism to respond to users’ ratings or follow-up user questions.Currently,it is popular to use Sequence to Sequence(Seq2 seq) generation models or fusion information retrieval methods to generate responses to user reviews.However,the existing retrieval methods do not consider the semantics of the sentences,and the generative models do not consider the difference between user reviews and retrieved reviews, which leads to poor utilization of knowledge by the model and reduces the quality of the response.To solve these problems,in this paper,we propose a Semantic Retrieval and Generation Framework named SRGen,which consists of a retrieval model and a generation model.Firstly,the Sentence-BERT(SBERT) model is fine-tuned in a self-supervised learning manner according to different reviews that may have similar responses especially in the same application.Then the Top-k most similar reviewresponse pairs are retrieved in the knowledge base using the joint vector representation of the name,rating,and review information obtained by SBERT.Finally,according to the difference between the retrieved comment and the original comment and the corresponding response content,the response of the comment is generated by a generative model.Experimental results show that the SRGen improves by 12.4% under the metric of BLEU,and 9.4% under the metric of ROUGE.

关 键 词：软件维护 用户评论 App评论响应 语义检索 Seq2seq 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]