Research And Implementation Of English Textual Entailment Recognition Based On Hybrid Neural Network

With the wide application of computers and the rapid development of Internet technologies,today's society is advancing toward the era of big data at a rapid speed,and the text data in the form of electronic documents is also increasing.Just letting the computer process the surface information of the text is far from meeting the needs of modern people.How to make the computer understand the natural language text deeply and obtain valuable information accurately and efficiently becomes more and more important.Recognizing textual entailment is a task of judging directional semantic relationship between texts.It is required to judge whether the semantics of H can be inferred from T for a given two texts,text(T)and hypothesis(H).Recognizing textual entailment aims to promote the semantic study of texts and improve the computer's ability to understand natural language texts.It has very important research value and application value.This thesis mainly studies the textual entailment recognition method for English.Through the deep study of the problems in semantic representation and entailment methods,this thesis proposes an English textual entailment recognition method using capsules and an English textual entailment recognition method which combines the capsules with the pre-trained BERT model.The accuracy of the former on the SNLI,MultiNLI-matched,MultiNLI-mismatched and SciTail datasets is 89.2%,77.4%,76.4%and 78.4%,respectively;The accuracy of the latter on the SNLI and SciTail datasets is 89.4%and 86.3%,respectively.The main contributions of this thesis are as follows:1.An English textual entailment recognition method using capsules is proposed.In terms of the semantic representation of words,the convolutional neural network is used to extract the local feature information of words from the character level,which alleviates the problem of non-obvious semantic features caused by random initialization of OOVs.Secondly,in the process of recognition,it constructs a capsule for each kind of relationship by combining interactive attention mechanism,and completes the identification of the final relationship by means of sub-category learning.2.An English textual entailment recognition method which combines the capsules with the pre-trained BERT model is proposed.Based on the capsule-based English textual entailment recognition method,it further combines the aggregated features of the text T and the hypothesis H coding by the pre-training BERT model to judge the relationship between the two texts.3.The verification experiments are carried out on three standard datasets.The results show that the proposed methods are comparable to other advanced English textual entailment recognition methods in this field.The visual analysis of the attentional relationship matrix also validates the effectiveness of the capsule in this task.4.According to the English textual entailment recognition method which combines the capsules with the pre-trained BERT model,the corresponding system is designed and implemented.The system includes a data preprocessing module,a neural network module,and an ensemble learning module,which can automatically identify the relationship for a given text pair.