Research On Conversational Machine Reading Comprehension Algorithm Integrate History Information

Conversational machine reading comprehension is an essential task in natural language processing,and its purpose is to evaluate the machine’s understanding of the text through the form of conversation.Studying this task makes it possible to improve the accuracy of models such as dialogue systems,question answering systems in answering questions.With the popularity of deep learning technology in recent years,much research on conversational machine reading comprehension has also begun based on deep learning,selecting appropriate conversation history information is crucial to the accuracy of machine answering questions and the validity of judging questions.This thesis does the following two works on how to select the conversation history information:(1)To improve the accuracy of the machine in judging whether the follow-up question is valid,propose a strategy to integrate the history information and questions.Specifically,different historical position embeddings are set for each round of conversational history information and the current question.The spliced position embedding and the spliced original sentence embedding are added to obtain a sentence embedding representation that integrates the conversational history information and the current question.Then,the feature extraction based on the self-attention mechanism is designed,and the discrimination method based on the convolutional neural network and the recurrent convolutional neural network for feature classification and the discrimination method based on the Bert model are respectively designed.Finally,three models are tested on the LIF data set.The F1 and Acc scores are used as evaluation indicators to test the effectiveness of each model when selecting the number of conversational history information rounds as 1,2,and 3 rounds.Accuracy,and carried out a contrast experiment.The results show that the strategy of fusing the conversational history information and the question can improve the machine’s ability to judge the question’s validity.(2)To improve the accuracy of a machine answering questions,propose a semantic similarity discrimination method,a conversational history information screening strategy based on the combination of Glove and cosine similarity.Then a Bert model-based answer prediction method is designed to implement the strategy.Finally,the strategy based on the Bert model is tested on the Co QA dataset,and the F1 score is used as the evaluation index.In the experiment,the Bert model combined with the conversational history information is compared with the Bert model without conversational history information and carried out a contrast experiment.The results show that the conversational history information screening strategy based on the combination of Glove and cosine similarity can improve the accuracy of machine answering questions.