Design And Implementation Of Question Answering System Based On Reading Comprehension

Faced with the massive amount of information on the Internet,people are increasingly relying on search engines to obtain information.Traditional search engines return web pages related to user queries,and users need to spend a lot of time and efforts to get the information they need.Different from traditional sear-ch engines,Retrieval-based Question Answering,(Q&A)System obtains relevant documents through Information Retrieval technology and uses Q&A algorithm to extract answers from relevant documents,which can provide users with short and accurate results.Traditional Retrieval-based Q&A System adopts a pipeline-based Q&A algorithm.This kind of Q&A algorithm can only capture the simple semantic associations in the input information,and there are problems such as high optimization cost and high optimization difficulty.Considering the advantages of the end-to-end Reading Comprehension model,including the advantages of simple training,intuitive optimization and the ability to capture the complex semantic associations in the input information,we have designed and implemented a Q&A System based on Reading Comprehension.For the Q&A System designed and implemented in this thesis,it includes five modules:Web service module.Information Retrieval module.Chinese preprocessing module,answer prediction module and log storage module.Among them,the Q&A algorithm of answer prediction module is implemented based on Reading Comprehension model.In particular,in view of the fact that the current Reading Comprehension models are difficult to meet the computational efficiency requirements of the Q&A algorithm in Retrieval-based Q&A System and can't process the case where the number of answers in relevant documents is not fixed,we have designed a Reading Comprehension model for Retrieval-based Q&A System(RQA-RC).RQA-RC model consists of four parts:question encoding structure,document encoding structure,Attention Mechanism and prediction structure.Considering that Retrieval-based Q&A System pays more attentions to the computational efficiency of the Q&A algorithm and the characteristics of the Q&A data in Retrieval-based Q&A System scenario.we have designed and combined a question encoding structure based on Bi-directional Long Short-term Memory and a document encoding structure based on Convolutional Neural Network,which makes the model significantly improve the computational efficiency while maintaining good natural language text encoding.In this Q&A scenario,considering that the number of answers in relevant document is not fixed,we have desianed a prediction structure based on Sequence Labeling and have proposed a loss function with reference to Focal Loss idea,which enables the model to efficiently process the case w-here the number of answer's in relevant document is not fixed.The results of experiment show that RQA-RC model has a significant improvement in computation efficiency and F1 score compared to current Reading Comprehension models in Q&A task for Retrieval-based Q&A System.Through system testing,all modules of the Q&A System designed and implemented in this thesis meet the design expectations,and the system performance meets application requirements.