Design And Implementation Of Multi-Passage Reading Comprehension Based On Deep Learning

With the development of the Internet,more and more Internet users get information from massive unstructured web pages.However,the current mainstream search engines can only return multiple text level results for the problems raised by users,and there are many interference information in these returned results,so users need to judge and filter the returned results to get the information they need.This process not only increases the difficulty of user search,but also increases the time of user search.Therefore,this paper realizes a multi-passage reading comprehension system,which can give users more accurate answers.The algorithm of multi-passage reading comprehension refers to that the given question and multi-passage machine automatically extract the answer from the given paragraph.However,there are some problems in the multi-passage reading comprehension algorithm,such as the length of the text is too long,which leads to the poor reading comprehension effect,and the candidate answer sorting algorithm can't synthesize the multi-faceted text semantic information.Therefore,in order to improve the performance and usability of the multi-passage reading comprehension algorithm,this paper studies and improves the multi-passage reading comprehension algorithm,and implements the multi-passage reading comprehension system based on the improved algorithm.The main research contents include the following three aspects:(1)First of all,this paper implements a reading comprehension model based on dependency syntax and attention mechanism.In the long text,the traditional attention mechanism calculates the attention weight of all words,which causes noise interference.Based on the Bert output,the model only computes the attention weight for the dependent syntactic phrases,which effectively reduces the misleading of the useless information to the semantic expression.Experiments show that the model proposed in this paper achieves better prediction results than match LSTM,BiDAF and Bert.(2)Then,the paper implements a candidate answer ranking model which integrates the features of text implication relationship and many other text features.To solve the problem that the candidate answer sorting algorithm can't integrate many aspects of semantic information.The model proposed in this paper integrates the deep semantic relationship among the questions,articles and paragraphs.Experiments show that the model proposed in this paper has better prediction effect than DrQA,R3,HAS-QA and RankQA(3)Finally,a multi-passage reading comprehension system is designed and implemented by using Vue,flask,JavaScript and other technologies.The system includes five modules,which are web service module,algorithm module,data acquisition module,basic support module and log storage module.At the same time,combined with the functional and non-functional requirements of the system,each module is designed and implemented in detail.The test results show that the multi-passage reading comprehension system meets the system requirements and achieves the expected goals.