Research On Question Answering System Based On Attention Mechanism And Answer Verification

Unlike traditional search engines that match and sort web results based on input keywords,question-answering(QA)system based on machine reading comprehension can directly and accurately answer users' questions and return the text of answers,which provides users with an efficient and accurate method of information retrieval.And it is widely used in application scenarios such as auxiliary decision,community QA,chat robot and intelligent service,which has the great development prospect and application value.With the release of various QA datasets and the significant improvement of GPU computing efficiency,the research of QA system has gradually changed from traditional shallow semantic analysis to deep semantic comprehension based on deep learning models.Good achievements have been made in the QA research of Chinese and English,but there are still some deficiencies and difficulties.Aiming at some key problems existing in the research of QA task,this thesis uses attention mechanism,pre-trained language model,text representation and other methods to improve the QA model.Thus,it can improve the accuracy of the predict answer,and apply the improved QA model to an online QA system.The main research work of this thesis is as follows:1.For answer verification,this thesis has proposed an answer verification method called V4 ES based on the pre-trained language model.V4 ES simulates the process of human progressive reading comprehension based on the layer-by-layer encoded by the pre-trained language model,and designs an answer verification mechanism including sketchy reading,intensive reading and verification.By making full use of the encoded features of the text at the high and low levels from the pre-trained language model,V4 ES simulates the process that human beings grasp the general idea of the text and the intention of the problem from shallow to deep in the real world.Finally,a series of experiments on the QA datasets of Chinese and English verify the performance and effectiveness of the answer verification method V4 ES based on the pre-trained language model.2.For text representation,this thesis has proposed a text representation method called FMF based on the fusion of multi-granularity features.FMF uses character features,word features and topic features to represent and learn the context and question text respectively,and realizes the combination encoding features of the three by using word-formation operation,vector-combination operation and attention mechanism.FMF integrates the multi-granularity and multi-level features of the text,and makes full use of the semantic,grammatical features and context interaction contained in the text,which enhances the expression ability of text features.Finally,a series of experiments on internal and external evaluation task and QA task verify the effectiveness of the text representation method based on on the fusion of multi-granularity features.3.For model application,this thesis has designed and implemented an online automatic QA system.The QA system deploys the QA service based on the QA model proposed in the above two points.The system development adopts B/S architecture for design and implementation,provides a concise and beautiful visual interface,receives the user's QA request,and directly returns the text of answer corresponding to the user's question.