Research On Chinese Multi-text Reading Comprehension Model Based On Neural Network

Machine reading comprehension is an important research direction in the field of natural language processing.According to the number of texts provided,it can be divided into single text reading comprehension model and multi-text reading comprehension model.In view of the higher challenge and practical application value of the latter,this dissertation focuses on the research of multi-text reading comprehension model in Chinese reading comprehension data sets.The model architecture includes paragraph selector and text reader.For a given question,documents and reference answers,firstly,a paragraph selector is used to select paragraphs with high relevance to the reference answers from the documents.Then,the text reader completes reading work on the selected paragraphs and predicts the range of the answer.According to the tasks of paragraph selector and text reader,the main research work of this dissertation includes:1.Propose two kinds of paragraph selectors based on convolutional neural network,which are two-class version and multi-class version respectively.Both selectors transform paragraph selection into a classification task.First,the word vectors of question and paragraphs are pre-trained.Then,bi-directional long short term memory network is used to encode context information in the question and paragraphs respectively.Next,attention mechanism layer is used to match semantic information between question and paragraphs,emphasizing paragraph content with high relevance to the question.After that,convolution neural network is used to extract important features in paragraphs and generate feature vectors of paragraphs.Finally,the two-class classifier uses the paragraph feature vectors to predict the probability that the paragraphs contains the answer,and selects the paragraphs with higher probability.The multi-class classifier further calculates the semantic similarity between paragraphs and question,and then selects the paragraphs with higher similarity.2.For the task of text reader,the basic neural network text reader based on attention mechanism and the R-net model incorporating matching features are implemented,and the effectiveness of these models in Chinese data sets is verified.The experiments are completed in Baidu knows and Baidu search subdatasets of the Dureader Chinese reading comprehension dataset.Compared with the baseline method,the results obtained by the model have significantly improved the Bleu-4 and Rouge-L evaluation scores.