| With the development of the Internet,a large amount of text data has been accumulated,and understanding text data and returning information required by users has become a hot issue in the field of natural language processing research.However,the current search engine can't fully understand the user's intention by only returning search results based on the relevance between the web page and the questions.With the continuous development and innovation of deep learning technology,machine reading comprehension models have become the preferred algorithm for people to efficiently obtain effective information,which can meet users' fast and efficient search needs,thereby saving time and costs.Machine reading comprehension is an important task for measuring text understanding.Its development is closely related to deep learning technology and has achieved good results to a certain extent.However,there are still the following problems: in terms of models,the current machine reading comprehension models are mostly faced with the loss of original information with the deepening of the network,which reduces the prediction ability of the model to the answer.In terms of data sets,facing more and more large-scale reading comprehension data sets oriented to real application scenarios,many machine reading comprehension technologies are unable to give a good answer to such data sets.In addition,although many reading comprehension techniques have achieved good results in English datasets,they have not been fully verified in Chinese scenarios.Based on the above problems,in order to better complete the task of machine reading comprehension,this paper proposes a Chinese machine reading comprehension model based on deep learning based on the research of the existing machine reading comprehension model and the span extraction reading comprehension task.The main research contents and contributions of this paper are as follows:(1)For the traditional stacked BiLSTM network,with the deepening of the network,information loss and attention mechanism will be more unstable in the deeper the network,this paper proposes the BiDAF-DRCN reading comprehension model.In the BiDAFDRCN model,this paper proposes a DRCN encoder by combining DenseNet and BiLSTM.The DRCN encoder connects the original information of the article and the problem,the attention weight,and the information of the BiLSTM layer in a DenseNet manner,obtains the representation of the article and the problem,strengthens the transfer of features in the network,and makes full use of the original information.On the other hand,in terms of BiDAF,which only extracts the bidirectional information from the article to the question and the question to the article,and ignores the internal structural relationship between the article and the question.Thus,the BiDAFMHT-Reader model is proposed in this paper.In the BiDAF-MHT-Reader model,this article first combines Highway and a Convolutional Neural Network to initially extract the features of articles and questions and enrich the input information.Then,based on BiDAF,the multi-head self-attention mechanism is integrated to capture the internal information of the article and the problem,and fully understand the information of the article and the questions.Compared with the results of BiDAF on DuReader-Search and DuReader-Zhidao,the BLEU-4 value of BiDAF-DRCN increased by 1.4% and 1.9%,and the Rouge-L value increased by 1.4% and 0.9%,respectively,BiDAF-MHT-Reader's BLEU-4 value increased by 1.7% and 2.0% respectively,and Rouge-L value increased by 2.1% and 1.0% respectively.(2)Aiming at the problem that each word in the machine reading comprehension article has different importance to answering questions and strengthen internal information interaction,this paper proposes an Attention-Reader model by combining the gated attention-based Recurrent Neural Network in the R-NET model.The model first uses a multi-head self-attention mechanism to obtain contextual information about articles and questions,and facilitates the extraction of internal structural features of articles and questions.Then through the gated attention-based Recurrent Neural Network in R-NET to promote understanding of important information in articles and questions,and improve the model's ability to predict answers.At the fusion layer of the model,the multi-head self-attention mechanism is also used to understand the article information after the question information is fused again,and to strengthen the interaction of the information inside the article.The experimental results show that compared with other mainstream models,Attention-Reader shows good performance on DuReader and MS MARCO data sets. |
