Research Of Chinese Textual Entailment Recognition Method Based On Attention Mechanism

Recognizing textual entailment,also known as natural language inference,is a basic task in the field of natural language processing.Its goal is to determine the semantic reasoning relationship between two sentences.It has important application value in machine translation,question answering systems and other tasks.With the development of deep learning,various methods of recognizing textual continuously spring up.Based on the research and analysis of many recognition models in recent years,this paper proposes a kind of textual entailment model combining mutual attention mechanism and a new aggregation mechanism.This model surpasses the baseline model in Chinese text entailment data sets.In this paper,the model is optimized on embedding layer and it is conducted feature fusion experiments.Beyond that,the model is applied to question and answer task and sentence matching task respectively.The specific research contents are as follows:(1)Propose a recognizing textual entailment model based on mutual attention mechanism and aggregation mechanism.The current deep learning recognizing textual entailment method is mainly composed of three parts: coding,information interaction,and aggregation classification.Based on this architecture,this paper proposes a text entailment model that combines mutual attention and a new aggregation mechanism.The model uses mutual attention mechanism to identify the semantic reasoning relationship between sentences,uses an improved aggregation mechanism to retain more coding features and reduce noise,and uses pooling operations to enhance the feature extraction capability of the model classification layer.The model in this paper achieves an accuracy rate of 92.38% on the corpus of CCL2018 Chinese text,which is better than the common benchmark models,such as Bi-LSTM,ESIM,Decomp-Att and so on.(2)Recognizing textual entailment model based on optimized embedding layer.The model of this paper focuses on the information interaction and aggregation classification part to improve,so the embedding layer of the model is to further optimized.This paper uses MLP,CNN,LSTM,Self-Attention and other mechanisms as embedding layer of the model for comparative experiments.The experiment result shows that using Self-Attention as embedding layer has better performance improvement effect.Using the coding optimization model,the accuracy rate on the NLPCC2016 question and answer data set reaches to 81.31%.(3)Recognizing textual entailment model based on fusion vector features.Considering that the smallest semantic unit of Chinese is character,and one of the basic attributes of word is part of speech,so this paper conducts character vector and part of speech vector fusion experiments on multiple text entailment methods including the model this paper proposes.The result show that character vector fusion has an enhancement effect on all experimental models,while part of speech vectors have a slight improvement effect on some models.This paper conducts an experiment on CHIP2018 sentence matching task.The experiment result shows that character vector fusion can improve the sentence matching task result.