摘要
在最新的任务导向型对话系统挑战中,有效利用主观知识(如个人见解)对于满足用户的特定需求至关重要。然而,由于此类知识具有个体主观性的特征,如何有效地整合和利用这些信息成为了研究的关键焦点。提出一种名为DynSense的方法,旨在解决从多条相关用户主观意见中生成全面且概括性回复的挑战。DynSense首先运用基于方面的情感分析(ABSA)技术来解析主观知识片段中的方面及其情感极性,并实现用户询问与知识片段的对齐。接着,利用先进对话模型结合对话上下文及经ABSA增强的信息生成回应。特别设计的DynMatch算法通过动态选择与当前查询最相似的高质量知识片段作为少样本提示(few-shot prompts),以引导模型生成更贴切的回复。实验结果表明,DynSense展现出对潜在语义特征和情感倾向的卓越捕捉能力,实现了精准、全面且高度贴合过往用户评价的回复。与现有模型相比,DynSense在SKTOD基准上的各项评估指标均有显著提升。
In the latest task-oriented dialogue system challenges,effectively utilizing subjective knowledge(e.g.,personal opinions)is crucial for addressing users’specific needs.However,due to the inherently subjective nature of such knowledge,how to effectively integrate and leverage this information has become a key focus of research.This paper proposed a method called DynSense,aimed at addressing the challenge of generating comprehensive and generalized responses from multiple relevant subjective user opinions.DynSense firstly employed aspect-based sentiment analysis(ABSA)to parse the aspects and sentiment polarities within subjective knowledge snippets,aligning them with the user’s query.Then,it utilized an advanced dialogue model that combined the dialogue context with ABSA-enhanced information to generate responses.A specially designed DynMatch algorithm guided the model to generate more relevant responses by dynamically selecting high-quality know-ledge fragments most similar to the current query as few-shot prompts.The experimental results demonstrate that DynSense exhibits exceptional ability in capturing latent semantic features and emotional tendencies,generating precise,comprehensive,and highly aligned responses based on past user reviews.Compared to existing models,DynSense shows significant improvements across various evaluation metrics on the SK-TOD benchmark.
作者
饶东宁
庄杰涛
Rao Dongning;Zhuang Jietao(School of Computers,Guangdong University of Technology,Guangzhou 510006,China)
出处
《计算机应用研究》
北大核心
2025年第6期1706-1712,共7页
Application Research of Computers
基金
广东省自然科学基金面上项目(2021A1515012556)。
关键词
任务导向型对话系统
主观知识
基于方面项的情感分析
动态少样本提示
task-oriented dialogue systems
subjective knowledge
aspect-based sentiment analysis(ABSA)
dynamic few-shot prompts
