基于ABSA与动态少样本提示的主观知识对话回复生成模型
2025-08-03饶东宁庄杰涛

Subjective knowledge dialogue response generation model based on ABSA and dynamic few-shot prompting
Rao Dongning,Zhuang Jietao (School ofComputers,Guangdong UniversityofTechnology,Guangzhou 51ooo6,China)
Abstract:Inthelatest task-oriented dialoguesystem challenges,efectivelyutilizing subjective knowledge(e.g.,personal opinions)iscrucialforaddresingusers’specificneeds.However,duetotheiherentlysubjectivenatureofsuchknowledge, howto efectively integrate and leveragethis information hasbecome a key focus of research.This paper proposeda method called DynSense,aimedataddresing thechallngeof generatingcomprehensiveand generalizedresponsesfrommultiplerelevant subjective user opinions.DynSense firstlyemployedaspect-basedsentiment analysis (ABSA)to parse the aspects and sentiment polarities withinsubjective knowledge snippets,aligning them with theuser’squery.Then,it utilizedanadvanced dialoguemodel thatcombined thedialoguecontext withABSA-enhanced information to generateresponses.AspeciallydesignedDynMatchalgorithm guidedthe model to generate morerelevantresponses bydynamicallselecting high-quality knowledgefragmentsmost similartothecurrentqueryasfew-shot prompts.The experimental resultsdemonstrate thatDynSense exhibits exceptionalabilityincapturing latentsemantic featuresand emotional tendencies,generating precise,comprehensive, andhighlyalignedresponses basedonpastuserreviews.Compared toexisting models,DynSenseshowssignificantimprovements across various evaluation metrics on the SK-TOD benchmark.
Key words:task-oriented dialogue systems;subjectiveknowledge;aspect-based sentiment analysis (ABSA);dynamic fewshot prompts
0引言
经典的任务导向型对话系统主要依赖于事实性知识,例如文献[1\~4]中所使用的常见问题(FAQ)数据库。然而,Majumder等人[5]的研究表明,对话建模正在向融入主观信息的方向转变。这包括个性化的数据,如用户体验、个人见解及用户偏好,从而使对话更具情境相关性和吸引力。通过整合主观知识,现代对话系统旨在增强其响应性和个性化能力,进而提升用户满意度及对话整体质量。这一转变不仅克服了纯事实性互动的局限,还满足了对更细腻、更人性化对话不断增长的需求。
第十一届对话系统技术挑战赛(The11thDialogSystemTechnologyChallenge,DSTC11)作为最新的国际对话系统技术竞赛,发布了首个结合主观知识与客观知识的任务导向型对话系统基准数据集——SK-TOD(subjective knowledge-groundedtask-orienteddialogue)[6]。该数据集标志着对话系统研究领域的一大进步,因为它首次将用户评论等主观信息与FAQ等客观信息融合在一起,用于构建更加情境化和人性化的对话系统。图1展示了一个SK-TOD的具体实例,系统需要根据包含“cozyambiance”在内的用户评论(主观知识)以及FAQ(客观知识),综合生成对“environmentandatmosphere”这一询问(即对话中的最后一个提问)的回答。蓝色词“environmentand at-mosphere”是用户询问的方面。主观知识片段中,红色词语表示正面评价,棕色词语表示负面短语(见电子版)。

文献[7\~11]表明,尽管预训练对话模型或通过提示工程引导的大规模语言模型在经典对话生成任务中展现了令人满意的表现,但在SK-TOD基准测试中,如何有效地引导这些模型从众多主观知识中提炼出相关要素,并生成能全面反映过往评价的综合性回复,依然是一个艰巨的挑战。……
