Research On The Factoid Question Answering Based On Attention Pooling Mechanism And External Knowledge

With the rapid development of Internet and the fast spread of smart phones,data are exploding.Facing this phenomenon,how to find useful information becomes a concern for users.Traditional search engines can retrieve relevant webpages to users through keyword matching.However,the keyword matching are the shallow matching but ignores the deep semantic meanings.In other hand,it is still time-consuming and tiring to browse the retrieved webpages.To solve this,the intelligent question answering system is proposed to answer the question directly by using information retrieval,knowledge graph and reading comprehension techniques.Recently,due to the tremendous increase both in computing power brought by GPUs and in data volume,deep learning has made great performances in several areas including speech recognition,image analysis and natural language processing.In this thesis,we propose an end-to-end factoid intelligent question answering.We use searching technology to search relevant text passages based on user's question and we take neural networks to extract factoid answer from these passages.In the meantime,in order to improve the performance of question answering,we introduce the optimization mechanisms including the attention pooling mechanism and external knowledge incorporating methods.Our contributions are as follows:1)We implement a framework of the end-to-end factoid question answering,which includes an information searching module and an answer extracting module.We compare multiple similarity scoring mechanisms in the searching module.The information searching module can provide documents for the answer extracting module,so it is the basis of the answer extracting module.2)We propose a novel attention pooling mechanism for answer extracting module.Here,we model the answer extracting task as a reading comprehension task.Therefore,based on the general reading comprehension frameworks,we introduce an attention pooling mechanism optimization algorithm which can bring the global and local attention into the answer extracting module,and improve the accuracy of answer extracting.3)We propose an external knowledge incorporating mechanism based on a two-layer attention model,which can introduce external knowledge into reading comprehension frameworks.