Research On Chinese Textual Entailment Recognition Based On Deep Learning

The purpose of recognizing textual entailment is to judge the semantic and logical relationship between two texts.The reasoning process involves syntactic analysis,vocabulary comprehension,logical reasoning,social experience,common sense and so on.Recognizing textual entailment is a challenging research task to judge whether the computer understands the text semantics to a certain extent,and is one of the most important benchmark tasks in the field of natural language processing.Deep learning method has been widely used in the field of recognizing textual entailment in recent years.However,there is a lack of research work on recognizing Chinese textual entailment,and there are still some challenges: the model is limited to extract the deep semantics of sentences,complex external knowledge is difficult to be integrated,the reasoning process does not follow the one-way principle and so on.Therefore,this thesis uses deep learning technology to explore and study a series of problems existing in the research of recognizing Chinese textual entailment.The main research work is summarized as follows:Firstly,to address the problems that the ability of self-attention mechanism to capture complex sentence semantics is insufficient and the dataset is small scale as well as large noise,we propose a recognizing Chinese textual entailment method which combines semantic role and self-attention mechanism.Based on the structure of Transformer,the method integrates the results of the semantic role labeling with the self-attention mechanism,and improves the ability of self-attention mechanism to capture the semantic meaning of sentences.In addition,using large-scale pre-trained language model,Bertwwm-ext,can significantly improve the recognition performance of the model on smallscale dataset.Secondly,to address the problems that complex external knowledge is difficult to be integrated and semantic role labeling information is difficult to be encoded completely,we propose a recognizing Chinese textual entailment method based on knowledge graph network.To begin with,the graph attention network is used to encode the ConceptNet knowledge network and the semantic role labeling information.Next,the graph embedding results are processed by local inference as well as global inference.Finally,the relationship categories are obtained by fusing the features.The experimental results show that the graph attention network can effectively integrate the complex ConceptNet structure into the recognizing textual entailment model,thereby improving the reasoning ability in common sense based on external knowledge.Moreover,the graph attention network can more completely encode the semantic role information of complex sentences,so as to improve the overall recognition performance of the model.Thirdly,to address the problems that recognizing textual entailment model is lack of one-way reasoning ability and high complexity,we propose a recognizing Chinese textual entailment method based on one-way deep fusion.This method constructs a new sentence interaction module based on the one-way reasoning,and eliminates a large number of alignment mechanisms and the dense connection in the previous model.The experimental results show that this method can not only effectively improve the performance of recognizing Chinese textual entailment,but also speed up the training speed of the model.