The era of big data has brought convenience to people's basic necessities of life,but a great deal of useful and useless information has also affected people's life to a certain extent,how to obtain practical text information from big data has become an issue in the field of natural language processing to be solved.Recognizing Textual Entailment(RTE)is a technology that has an important role in understanding the meaning of text and can help people get useful information quickly.Therefore,research textual entailment recognition has a profound meaning.Recognizing Textual Entailment refers given text T and hypothesis H identified whether T is implied H,in other words,when a person reads T and deduces whether H is true,Chinese textual entailment recognition is identified whether there is implication between Chinese sentence pairs(T-H pairs).Aiming at the task of Chinese textual entailment recognition,this dissertation proposes a Chinese textual entailment recognition method based on deep neural network,and constructs a Chinese textual entailment recognition system based on CNN(Convolutional Neural Network)and BiLSTM(Bidirectional Long Short-Term Memory).The system improves the MacroF1 result of 61.74%on the dataset of RITE-VAL in 2014 NTCIR-11,better than BUPT which is 61.51%.The experiment proves that the method is effective for Chinese textual entailment recognition.The main contributions of this dissertation are as follows:1.Propose a feature learning method based on the deep neural network.The method is performed by a variety of neural networks for Chinese text feature extraction,selection,learning and automatic mining shallow and deep levels of information.To some extent avoid artificial feature engineering and error accumulation problem.2.Propose a Chinese textual entailment recognition method based on the deep neural network.This method combines the advantages of deep learning and traditional rules to make the textual entailment recognition more accurate.3.Construct a Chinese textual entailment recognition system based on CNN-BiLSTM-parallel.This dissertation evaluates on the dataset of RITE-VAL in 2014 NTCIR11 and obtains 61.74%which exceeds optimal value 61.51%by BUPT. |