Study On Comparative Recognition And Extraction Of Comparative Elements Based On Deep Learning

A Comparative sentence is a special sentence pattern in product reviews,which contains information about people’s comparative views on goods or services.By analyzing the comparative sentences,we can more easily identify the emotional tendencies in the comments.Massive comment information stores a large amount of high-value information generated by users,which hides the users’ emotional attitudes and political tendencies.The resolution of such problems has an important role in providing information support for product comment mining and other information.In this thesis,by reviewing the current situation of comparative sentence research,in view of the limitations of the existing comparative sentence related research methods,the introduction of deep learning-based comparative sentence recognition and comparative relationship extraction methods is aimed at learning representation training to analyze cross-domain comparative sentences Extract the model.This thesis mainly develops the following work:(1)A Chinese comparative sentence recognition method with improved hybrid network model is proposed.The method for Chinese comparative sentence recognition relies too much on the experience of artificial experts and the difficulty of extracting the deep features.This article introduces the concept of a hybrid network model in deep learning.By using the BERT pre-training model,incorporating comparison keywords,and incorporating attention mechanisms,the hybrid network model(convolutional neural network and bidirectional long-term and short-term memory network)is improved.Improve the efficiency and performance of comparative sentence recognition.(2)A Chinese comparative relation extraction method of BiLSTM-CRF model incorporating comparative keywords is proposed.This model overcomes the shortcomings of the two-way long-short-term memory network that does not consider the label sequence and the conditional random field from the global reliance on manual extraction of features.The pre-trained word vector is used as the input of the model to extract the relationship in the comparison sentence.Compare keyword features for comparison relationship extraction.The experiment proves that on the fourth Chinese tendency evaluation corpus,based on the comparison relationship extraction of the model,good experimental results are obtained.(3)A Chinese comparison relation extraction method of BiLSTM-CRF model incorporating dependency syntax information is proposed.The main task of the extraction of comparison relations is to extract the relationship elements from the comparison sentences,and these elements have a dominant and dominated relationship in the dependency syntax.Therefore,on the basis of the BiLSTM-CRF model incorporating the comparison keywords,by incorporating the dependency syntax Information to narrow the distance between dependent words to improve the bidirectional LSTM layer’s ability to learn long-term context dependencies.Experimental results show that this method can improve the effect of comparative attribute extraction.