问题检索与答案排序互相促进的社区问答系统

发布时间：2018-03-03 00:36

本文选题：社区问答　切入点：问题检索　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：在社区问答(Community Question Answering,CQA)中,用户提出查询问题,CQA系统在大量已有的问题-答案对的知识库中搜索相似的问题,然后把该问题的最佳答案当作查询问题的答案返回给用户。CQA系统包括两个关键的子任务:(1)问题检索(QuestionRetrieval),通过估计问题对的语义相似性来找到和查询问题最相似的已有问题;(2)答案排序(AnswerRanking),按照答案回答问题的相关程度对多个答案进行语义相关性排序,选出最佳的答案。构建问答知识库是一项庞大而复杂的工程,一种可行的替代方案是利用互联网的庞大资源检索获得问题的答案。因此,本文的第一个工作是借助搜索引擎来构建一个网络资源辅助的社区问答系统,该系统在2015年TREC的实时问答竞赛中获得了第二名。以往关于CQA的研究多将CQA中的问题检索和答案排序两个任务分开独立解决,没有考虑它们之间的信息交互。本文的第二个工作考虑这两个任务的相互促进,并设计新的有效特征来进一步提高CQA的性能,相关工作发表在2016年IJCNN会议。传统CQA系统采用专家精心设计的特征,泛化性差,而深度学习的优势是能够自动学习特征。因此,本文的第三个工作探索了深度学习模型在问题检索和答案排序任务上特征自动学习的性能,相关工作发表在2016年的SemEval会议。在本文第二个和第三个工作的启发下,本文的第四个工作深入研究了深度学习框架下的CQA系统。本文提出一个基于门机制的深度神经网络模型,该门机制能够自动学习问题检索和答案排序任务间的交互信息,从而帮助进一步提高CQA性能。本文广泛而深入地研究了采用传统自然语言处理技术与深度学习方法的问题检索和答案排序相互促进的CQA系统,大量的实验结果表明,本文提出的两个任务相互促进的策略在传统方法和深度学习方法中都能够有效地提高CQA系统的性能。
[Abstract]:In Community Question answering and answering (CQA), users ask queries and CQA systems search for similar questions in a large number of existing questions-answer pairs of knowledge bases. Then the best answer to the question is returned to the user. CQA system including two key sub-tasks: 1) QuestionRetrieval is retrieved by estimating the semantic similarity of the question pairs to find the most similar to the query question. The answer is sorted by AnswerRanking.According to the degree of relevance of the answer to the question, the multiple answers are sorted in terms of semantic correlation. Choose the best answer. Building a question-and-answer knowledge base is a huge and complex project, and a viable alternative is to use the vast resources of the Internet to retrieve the answer to the question. The first work of this paper is to build a community Q & A system assisted by network resources with the help of search engine. In 2015, the system won the second place in the real-time quiz of TREC. In the past studies on CQA, the two tasks of question retrieval and answer sorting in CQA were solved separately and independently. The second work of this article considers the mutual promotion of the two tasks and designs new valid features to further improve the performance of CQA. The related work was published at the IJCNN Conference in 2016. The traditional CQA system adopts the characteristics carefully designed by experts and has poor generalization, while the advantage of deep learning is the ability to learn automatically. The third work of this paper explores the performance of feature automatic learning of deep learning model in question retrieval and answer sorting tasks. The related work was published at the SemEval Conference on 2016. Inspired by the second and third work of this paper, In the fourth work of this paper, we deeply study the CQA system under the framework of deep learning. In this paper, we propose a deep neural network model based on gate mechanism, which can automatically learn the interactive information between question retrieval and answer sorting tasks. In order to further improve the performance of CQA, this paper extensively and deeply studies the CQA system which adopts the traditional natural language processing technology and the deep learning method, the question retrieval and the answer ranking promote each other. A large number of experimental results show that, The strategies proposed in this paper can effectively improve the performance of CQA systems in both traditional methods and depth learning methods.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】