Research And Implementation Of Conversation Application Technology Based On Machine Reading Comprehension

Conversational interaction,especially multi-turn conversation,has become a hot topic in the field of human-computer interaction.In order to further improve the user experience of their products,big companies around the world have launched their own conversation applications.Machine reading comprehension is the core technology to realize conversation application.With the development of deep learning technology,using deep learning technology to improve the performance of multi-turn machine reading comprehension model is one of the hotspots of current conversation application research.At the same time,the application of deep learning technology requires more and more computing performance and data storage space,and the compression of corresponding deep learning model is also a problem to be solved.To solve these problems,in this thesis,the methods to improve the performance and reason speed of MRC machine reading comprehension model is studied.The machine reading comprehension model based on Bert is taken as the benchmark model.The benchmark model takes the pre training language model Bert as the main model structure.It carries out the fine turning on the multi-turn machine reading comprehension data set CoQA,and uses the accurate matching accuracy(EM)and fuzzy matching accuracy(F1)indicators to evaluate the model.In order to improve the performance of the benchmark model on datasets,a multi-domain multi-task learning method(mdmt1)for machine reading comprehension is proposed.The domain data is enhanced by introducing multi-domain data in the fine-tuning stage,and the generalization of the model is improved by using the multi-task learning method,introducing additional factual basis and using the information implied in the related tasks In addition,for the multi-turn machine reading comprehension,the research difficulty lies in how to establish the relationship between the historical conversations.In order to further improve the multi-turn machine reading comprehension ability of the model and better establish the relationship between the historical conversations,based on the benchmark model,in this thesis,a new context modeling method for the fine-tuning stage is proposed,which combines the current question with the historical questions.In this method,redundant information is removed and more useful information is used to predict the model,which further improves the EM and F1 of the model.Finally,in order to reduce the parameters of the model,a compression method based on knowledge distillation is proposed to optimize the model size and improve the reasoning speed.In this thesis,a conversation application based on machine reading comprehension is designed and implemented.By the comprehensive utilization of the above key technologies,the multi-turn machine reading comprehension tasks is supported.The effectiveness of key technologies and the application are verified by a set of tests.