Research On Machine Reading Comprehension Based On Multiple Documents And Multiple Answers

Machine reading comprehension is one of the main research directions of question answering system task in natural language processing field.Its goal is to combine given text to get the answer of the question,and it belongs to supervised learning task.Because in the traditional question answering system,the extraction method of answer will cause a lot of information redundancy,in many cases,it can not extract structured information effectively from the original text.The machine reading comprehension technology can extract the parts related to the questions from the limited text information,so as to give the answers that match the questions accurately.It can help and promote a more sound question answering system,and lay the foundation for the realization of advanced artificial intelligence.At present,the task of machine reading comprehension is mostly based on Bi DAF,which is the Bi-Directional Attention Flow Model.It adopts multi-stage and hierarchical processing,so that it can capture the features of different levels of the original text.At the same time,it uses the bi-directional attention mechanism without memory to strengthen the learning between the original text and the question features.However,as a popular model for machine reading comprehension,Bi DAF also has its limitations.In general,the output layer of the original Bi DAF model uses the Softmax function to find the beginning and the end of the answer in the original text,which is only applied to the reading comprehension task with single answer.This paper focuses on the machine reading comprehension task with multiple documents and multiple answers,involving complex technologies such as understanding and reasoning.Based on the Bi DAF model,it makes targeted modifications to the model according to specific tasks.The used data is a large-scale Chinese reading comprehension data set for military application scenarios.The data set contains 24445 question answer pairs and 62789 articles.Each of question answer pairs corresponds to 5 articles,and the number of answers may not be unique,and all of answers are text information extracted from the original document.The main research work of this paper is as follows.First of all,this paper makes a preliminary analysis of the data set used,including the statistics of the length of documents,the information source documents and the number and length of the answers,the improvement of word segmentation process and the statistics of word frequency,and the exploration of the similarity between the text of each data field.Especially,this work obtains an initial conclusion that compared with the non answer information source documents,the cosine similarity between the answer information source documents and their corresponding questions is significantly higher than the latter,which provides the basis for the later work of filtering input documents.Secondly,the output layer of Bi DAF model is improved to make it suitable for machine reading comprehension tasks with multiple answers.In order to apply the model to the data set of this paper,we consider using the word vector matrix of the original text which has fused the question information,mapping the semantic information contained in it to the scalar corresponding to each word of the original text through the trainable weight vector,and obtaining the predicted answer by the given threshold value and the hypothesis test.However,there may be some answers that can not be obtained by threshold screening,but they are more prominent in their local sequences,so we further build a more flexible local change point detection method,and also use t-test to obtain the answers.In addition,according to the improved output layer,the appropriate loss function is redefined to optimize the model.Through experiments,compared with the single threshold screening method,the model using the local change point detection method has significantly improved the prediction effect of the answer.Finally,in order to further improve the accuracy of the model prediction,the following measures are taken to improve the model: first,the Softmax function and the corresponding threshold value are used to filter the 5 articles of each question answer pairs,so as to reduce the amount of text input model without losing the key semantic information to improve the efficiency of model training;second,the Self-Attention Flow Layer is introduced behind the Bi-Directional Attention Flow Layer of the model,which makes the model learn the word dependency within the sentence,capture the internal structure of the sentence,and strengthen its learning on the original text and question characteristics;third,the character embedding is used to replace the word embedding,so as to eliminate the segmentation boundary error caused by the poor segmentation effect caused by a large number of proper nouns.By comparing the test results,it can be found that the introduction of self-attention mechanism can slightly improve the accuracy of the model prediction answers,but character embedding instead of word embedding will make the prediction effect of the model decline.