摘要
文本生成是实现可解释推荐系统的有效技术途径之一,有利于提升用户对平台的满意度和信任感.然而,现有方法忽略了用户历史评论与目标物品之间的情感一致性问题,使得所生成的解释文本差强人意.以电商推荐场景为例,提出一种基于情感可控文本生成的可解释推荐框架.该框架由评分回归模型与解释生成模型串联而成,前者输出的预估评分作为情感查询,用于辅助后者从历史评论中甄选出情感一致的评论语料,并产生情感可控的解释文本.通过建立多任务联合学习机制,实现了评分回归模型与解释生成模型之间的双向互通和协同优化.四个电商场景下的实验结果表明,所提出方法在评分预测精度和文本生成质量两类指标上均具有显著的性能优势.
Text generation plays a key role in building explainable recommender systems that are beneficial to enhance users'satisfaction and trust to platforms.However,current schemes ignore the emotional inconsistency between her past textual reviews and the target item,resulting in unsatisfactory interpretive text.Taking E-business item recommendation as examples,this paper presents an explainable recommendation framework based on emotion controlling text generation,which is composed of a rating regressor and a cascaded interpretation generator.Concretely,the regressor first estimates the rating score posted by the active user on the target item,and such a score is used to query the emotionally relevant reviews which are then used to generate interpretive text in an emotion controlling fashion.Under the action of multi-task joint learning,regressor and generator can be reinforced each other naturally and gradually.Extensive experiments under four E-Business recommendation scenarios show encouraging results in terms of both rating prediction accuracy and interpretation generation quality.
作者
邬俊
刘林
卢香葵
罗芳媛
WU Jun;LIU Lin;LU Xiangkui;LUO Fangyuan(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China;Engineering Research Center of Integration and Application of Digital Learning Technology,Ministry of Education,Beijing 100039,China;MoE Key Laboratory of Big Data&Artificial Intelligence in Transportation,Beijing 100044,China)
出处
《闽南师范大学学报（自然科学版）》
2023年第4期24-34,共11页
Journal of Minnan Normal University：Natural Science
基金
中央高校基本科研业务费(2023YJS022)
数字化学习技术集成与应用教育部工程研究中心创新基金重点项目(1321004)
香港中文大学(深圳)广东省大数据计算基础理论与方法重点实验室开放课题(B10120210117-OF01)
国家重点研发计划(2021YFF0901001)。
关键词
可解释推荐系统
情感可控文本生成
评分回归
预训练语言模型
explainable recommender systems
emotion controlling text generation
rating regression
pre-trained language models
