【作者】 高静；

【导师】 姜元春；

【作者基本信息】 合肥工业大学 ， 管理科学与工程， 2025， 硕士

【摘要】 社区问答平台随着用户数量的增加,信息过载问题日益严重。如何在众多答案中准确地选择出高质量的答案,减少用户筛选答案的时间和精力,已成为社区问答领域的重要课题。当前答案选择任务的研究重点是基于深度学习技术,利用问答文本对以外的信息来增强模型的性能。但当前方法尚未充分考虑回答者专业知识在候选答案序列信息传递过程中的作用,也鲜有研究从候选答案序列信息传递和问答交互两个视角同时出发,对答案进行建模。为此,本文提出了一种基于多视角表征的答案选择方法,从候选答案序列信息传递和问答交互两个视角建模答案的表征。在候选答案序列信息传递视角下,该方法假设在给定问题下,先发布答案的回答者和后续回答者之间存在潜在的说服关系,结合回答者的专业知识,建模候选答案之间的信息传递对答案表征的影响。在问答交互视角下,该方法基于问答文本之间的语义关联以及问答概念词之间的交互作用,建模问题对答案表征的影响。本文在三个真实的社区问答数据集上进行了多组实验。实验结果表明,与基线方法相比,本文所提方法在一定程度上提升了答案排序和分类的性能。此外,该方法还为探索候选答案之间的关系对答案选择的影响提供了新的思路。更多还原

【Abstract】 As the number of users increases,the issue of information overload is becoming increasingly severe in community question answering systems.How to accurately select high-quality answers from numerous responses,thereby reducing the time and effort users spend filtering answers,has become an important topic in the field of community question answering.The current research on answer selection tasks primarily focuses on employing deep learning techniques to enhance model performance by incorporating auxiliary information beyond the question-answer text pairs.However,current methods have not fully accounted for the role of the responders’expertise in the process of candidate answer sequence information transmission.Moreover,few studies have simultaneously modeled answer representations from both the perspectives of information transmission within candidate answer sequences and question-answer interaction.To address these issues,a multi-perspective representation-based answer selection method is proposed in this paper,in which answer representations are modeled from the perspectives of candidate answer sequence information transmission and question-answer interactions.From the perspective of candidate answer sequence information transmission,the method assumes that,for a given question,there is a potential persuasive relationship between earlier and later responders.It incorporates the responders’expertise to model the impact of information transmission between the candidate answers on answer representations.From the perspective of question-answer interaction,the method utilizes the semantic relationships between the question and answer texts,as well as the interactions between concept words in the question and answer,to capture the influence of the question on answer representations.In this paper,multiple sets of experiments were conducted on three real-world community question answering datasets.The results show that,compared to the baseline methods,the proposed method improves the performance of answer ranking and classification to some extent.In addition,this method also provides new insights into exploring the impact of relationships between candidate answers on answer selection.更多还原