Research On Conversational Machine Reading Comprehension Based On Deep Learning

In recent years,with the rapid development of Internet technology and artificial intelligence,the amount of data has exploded.When people face such complex data every day,how to obtain the information they want has become a major problem.The purpose of machine reading comprehension is to allow machines to read texts like humans.The main process is to give correct answers to given articles and questions on the basis of comprehension,which is an important task of natural language processing.At present,the pace of people’s life is accelerating,which has caused time to become fragmented.If machines can understand human language,it will have a huge impact on human-computer interaction.However,traditional machine reading comprehension is mostly a single round of interaction,and the interaction is poor.Conversational machine reading comprehension will be combined with dialogue and machine reading comprehension.When answering questions,it is necessary to consider the article and dialogue history,which has become a hot research topic.At present,deep learning models have achieved good results on multiple machine reading comprehension tasks,and have even surpassed human performance on some data sets,but such a complex model cannot be achieved without sufficient data training.For the best results,pre-training models based on large-scale unsupervised corpora have become the current mainstream.Most of the existing models are difficult to deal with the history of dialogue.They simply splice questions,articles,and dialogue history and then encode them,which increases the computational cost but the effect is mediocre.The current conversational machine reading comprehension data set not only predicts answers,but also predicts dialogue behavior.If the two tasks are modeled separately,the complexity of the model will increase.In view of the above situation,the CoBERT-BiGRU(Concat Bidirectional Encoder Representation from Transformers-Bidirectional Gate Recurrent Unit)model is proposed.First,split the input sequence into multiple variants,mark the conversation history in each variant in the original article,and input different marked articles and questions into the CoBERT model to obtain a vectorized representation of multiple sequences;Second,the representations of multiple sequences are input into the historical attention network,and multiple results are merged into a vectorized representation of a sequence.Because multiple variant sequences are essentially the same sequence and contain more redundant information,the simplified model reduces the complexity of the model;Finally,the fusion result is input into BiGRU to extract more advanced semantic information,and then the vector dimension is adjusted through the fully connected layer.The model adopts a multi-task learning method to jointly train answer prediction and dialogue behavior prediction,which improves the accuracy and training speed of the model,and enhances the generalization ability.The experimental results on the QuAC data set show that the CoBERT-BiGRU model can effectively handle multiple turns of dialogue history.Compared with the baseline model and others that have been published on this dataset,the HEQ-Q,HEQ-D and F1 values have all improved.