【作者】 王涛；

【导师】 张兮；

【作者基本信息】 天津大学 ， 管理科学与工程， 2020， 硕士

【摘要】 近年来,随着移动终端的普及和Web 3.0的发展,问答社区作为典型的知识共享平台发展迅速。但是所有用户都可以在问答社区中发布内容导致平台中内容质量难以保证,这极大地影响平台用户体验和留存。为保证问答社区中答案质量,结合深度神经网络模型和自然语言处理的问答匹配技术受到广泛研究和关注。目前结合注意力机制的深度神经网络由于获得较高准确率,是问答匹配领域最热门的研究方向之一。现有注意力深度神经网络模型大多基于词向量相似度构建注意力机制,这种注意力机制构建方式能很好的考虑问题和答案间的相关性信息。但是,这种方式忽略了其他对于答案质量评价同样重要的语义信息,例如客观性、理解性和说服力等,这限制了模型准确度和模型可解释性的进一步提升。为将更多答案质量相关语义信息纳入模型中以提高模型的准确性和可解释性,本研究提出一个基于外部化的注意力神经网络模型。本模型将知识创造理论中的外部化过程与问答匹配技术进行结合,使本模型的注意力模块能将外部化所代表的语义信息特征纳入注意力构建中。通过构建由主任务和辅助任务共同组成的整体模型,并对问答匹配的主任务和识别答案外部化的辅助任务进行软参数共享,使得模型在保留现有基于词向量相似度模型的大部分特征同时更多地利用数据中的语义信息。本模型还设计一个复合损失函数来同时训练主任务和辅助任务。除此之外,本研究对问答社区上的金融和教育数据集进行数据标注、实验分析和注意力结果的可视化。实验结果表明该模型不仅能提升模型的准确度,而且能加强注意力模型在解决问答匹配问题中的可解释性。本研究通过提出基于外部化的注意力神经网络模型,将知识创造理论中的外部化与问答社区中的问答匹配在算法层面进行结合,这对知识共享平台的实践和知识创造的理论研究都具有十分重要的意义。更多还原

【Abstract】 In recent years,with the popularity of mobile devices and Web 3.0,Community Question Answer(CQA)has grown rapidly.CQA is a typical knowledge sharing platform.However,the quality of content in the CQA is unwarranted,because everyone can freely publish content in the platform.The quality of content greatly affects the user experience and retention.In order to ensure the quality of answers in the CQA,the question-answer matching technology,which combines the deep neural network with natural language processing,has received extensive research in recent years.The question-answer matching can utilize more unstructured information.At present,the attention based deep neural network is one of the most popular research directions in the field of question-answer matching.Most studies indicate that text feature similarity between the question and answer can determine whether or not they match.However,these models may ignore the latent semantic information such as objectivity,understanding and persuasion,which limits further improvements in model accuracy and interpretability.To integrate more semantic information in modeling and improve the accuracy and interpretability,this study proposes an Externalization-Based Attention Neural Network(EBANN).This model equips the auxiliary task with the externalization from the process of knowledge creation in order to take more semantic information into the attention network.More specifically,EBANN focuses on modeling and transferring semantic information via soft parameter sharing between the main task of questionanswer matching and the auxiliary task of identifying degree of externalization for each answer.Moreover,this model designs a composite loss function to train the main and auxiliary tasks at the same time.Furthermore,this study carries out data annotation,experimental analysis and visualization of attention results for the finance and education data sets in the CQA.Experimental results demonstrate that this model can not only improve the accuracy,but also enhance the interpretability of the attention network in question-answer matching problem.The Externalization-Based Attention Neural Network combines externalization in knowledge creation theory with question-answer matching in CQA at the algorithm level.This combination is of great significance to the practical of knowledge sharing platform and the theoretical research of knowledge creation.更多还原