Research On Convolutional Deep Neural Networks For Document-Based Question Answering

As the Internet has become more and more powerful,people are increasingly relying on Internet technology in their daily lives.How to obtain useful information accurately and efficiently in a huge amount of data is becoming more and more important,academia and industry.The more and more circles are focused on the automatic question-and-answer technology in the field of natural language processing.However,with the explosive growth of data,traditional automatic question-and-answer technology has been unable to meet the requirements of people who want to acquire accurate information intelligently.Therefore,more and more scholars use the distributed representation technology of words and learn through neural network models.The semantic features of the sentence solve this problem.The core goal of the document-based automated question-and-answer task is to perform text matching and answer selection.The essence is to select the answer that matches the question by calculating the correlation between the problem and the candidate document,and how to include the vocabulary and text.Deeper semantic features are more accurately represented by vectors,which is the key to solving this problem,and thus can improve the accuracy of the model.In order to improve the accuracy of the text matching and answer selection model,the main work done in this paper is as follows:This paper proposes an automatic question-and-answer model combining multivariate features,by adding overlapping information(Overlap)between question and answer pairs in the word embedding layer,word position information(Position)in the sentence,and reverse document frequency(IDF)characteristics.The word vector matrix can contain more semantic information,which can improve the subsequent rich neural network model to learn more rich and accurate text vector representation.In this paper,a convolutional neural network model based on attention mechanism is used.By focusing on the convolutional layer output question and answer,the feature matrix is weighted,and the relationship between the question and answer pairs can be established,so that the key information in the text will be played.A bigger role.In addition,using a variety of filters of different sizes to capture the abstract semantic features of different lengths in the sentence and combine them,the answer selection accuracy of the convolutional neural network can be improved.In this paper,the experimental verification of the NLPCC 2016 DBQA data set is carried out.Compared with the traditional method and baseline,and the extended features of this paper are analyzed by single analysis and combination analysis.The experimental results show that the proposed model The results of MAP,MRR and other indicators have been improved to some extent,which can prove the effectiveness of the above methods.