Research On Domain-Dependent Automatic Question Answering Method

In the era of mobile Internet, the way to get information for people is more and more convenient, and people's demand for information is also growing. In order to meet the needs of different domains for the people of different levels, search engine is facing enormous challenges in the era of mobile Internet. The question answering technology, which has experienced a long evolutional process, can overcome the drawbacks of the search engine greatly so that people have a more c omfortable human-computer interaction. The question answering system can understand the natural language questions of people more accurately and return a concise and comprehensive answer to people instantly and efficiently after retrieving the knowledge base. The question answering system meets people's needs more efficiently. With the advancements in technology of artificial intelligence and natural language processing, many new question answering systems of different type have been developed because of the different data forms. In recent years, many domestic and foreign research institutions began working on human-like intelligence technology, the question answering technology will be applied to the college entrance examination. The research aims to build a question answering system for the history short-answer questions of college entrance examination.The main contents of the thesis include:Data preprocessing and system building. After conducting a sample analysis on the exam papers of college entrance examination, this paper concluded the types of questions and the difficulties for solving these questions. On basis of the main problems of all kinds of questions, this paper accomplished the data collection and resource storage of domain knowledge base and built a history retrieval system to make sure the normal operation of the question answering system. Because there might have classical materials in the question, the question answering system should be capable of translating these materials from classical Chinese to vernacular Chinese. For that, this paper collected a certain amount of parallel corpus.Candidate answers retrieving based on knowledge base. In order to obtain the documents related to the question from the knowledge base accurately, this pap er performed the traditional operations of keywords extraction, information retrieval and confidence calculation. After that, on account of the specificity, this paper tried the matching method based convolutional neural network. The method converted the candidate answers finding question into a sequence prediction problem to achieve a deeper semantic matching.Answer generation based on multi-documents. Afters obtaining the candidate documents set that contains the key elements of answer from the knowledge base, we want to generate a concise and accurate answer. According to the idea of the multi-documents summarization, this paper proposed a method to handle it. Firstly, we need to get some clusters containing many sentences. Next, we use the method of multi-sentences compression for information extraction and then the answer is generated.To facilitate the experimental analysis of the automatic question answering system performance, this paper defined a uniform evaluation standard. After testing on the exam papers of college entrance examination, it proves the validity of the system.