卷积深度神经网络在基于文档的自动问答任务中的应用与改进

基于文档的自动问答,尤其是语义匹配,其目标是计算两个文本之间的相似度。这是自然语言处理中的典型任务,并且用以衡量对自然语言的理解程度。深度学习方法得益于可以自动化地学习到给定任务的最优特征表示,在许多研究中取得成功,也包括文本匹配。针对基于文档的自动问答,提出一个基于卷积深度神经网络的语义匹配模型,以便对每一对问题和文档提取特征,并据此计算它们的得分。通过问题和文档之间的交互计算,利用重叠词等文本特征,在中文开放域上的自动问答任务中取得的实际效果证明了该模型的有效性。

APPLICATION AND IMPROVEMENT OF CONVOLUTIONAL DEPTH NEURAL NETWORK IN DOCUMENT-BASED QUESTION ANSWERING TASK

Document-based automatic question answering,especially semantic matching,aims to compute the similarity or relevance between two documents.It is a typical task in NLP and considered as a touch-stone of natural language understanding.Deep learning has been successful in many studies,including text matching,because it can automatically learn the optimal feature representation of a given task.Aiming at document-based question answering,I proposed a structure based on convolution depth neural network to extract features from each pair of questions and documents and calculate their scores.The effectiveness of the model was proved by the interactive computation between questions and documents,as well as the use of overlapping words and other text features in the automatic question answering task in the Chinese open domain.