Research On Key Technologies Of Open-Domain Question Answering Based On Textual Knowledge

The QA(Question-Answering)system can help humans to obtain the information they wanted quickly and conveniently.It has great application value in many fields.The research on QA system has profound and realistic significance for excavating the practical potential of artificial intelligence.Recently,The research of Open-Domain QA system based on textual knowledge has gradually catch more attention in the field of natural language processing.Besides,With the use of deep learning methods,it has achieved very rapid development.Humans wish computers to understand natural language like humans,and then automatically find the correct answer from textual knowledge for the question.The textual knowledge is usually massive,and it is diffcult to locate the correct answer precisely.The existing Open-Domain QA system adopts the framework with Retriever and Reader.Firstly,the traditional retrieval method is used to retrieve the candidate passages related to the question,and then the machine reading comprehension method is used to extract the answer.The traditional retrieval method utilizes the statistical features,and causes amount of noise,which will have a great negative impact for finding the correct answer.The lack of effective answer selection methods is also an important factor limiting the accuracy of the Open-Domain QA system.In order to solve the problem that amount of noise caused by Retriever of OpenDomain QA system will influence the total accuracy,this paper proposes to use the Multilayer Fusion textual matching Model(MFM)to filter the candidate passages returned by the Retriever.We try to filter the candidates accurately by extracting and analyzing semantic features at different levels,and then judging the semantic correlation between candidates and question.Thus,the scope of the correct answer has been narrowed,and the overall accuracy of the whole Open-domain QA system has been improved.For solving the inaccuracy of the traditional retrieval methods used in the existing open domain question answering system,this paper proposes a Reranker with fusion of multi-evidences semantic for candidate answer reranking.By using the deep learning method,a neural network model for reranking candidate passages is trained to not only analyze the semantic features extracted by the Reranker encoding,but also utilize the features extracted by Reader encoding,so that the two semantic features are combined to Rerank the candidate segments.Compared with other methods of Reranking,our method directly utilizes the hidden semantic information generated during the reading process as additional evidence,which can more accurately select the correct answer and its corresponding candidate passages.