Research On The Method Of Text Feature Extraction

Text sentiment classification is widely used in public opinion analysis,electronic commerce,information interception,financial investment and other fields.The traditional text sentiment classification method based on manually-designed feature is time-consuming,laborious and poor generalization ability.As Deep learning can automatically learn the characteristics of text,it overcomes the defects of the traditional text sentiment classification and has better performance in text sentiment classification.Word2vec tools which train distributed word vectors is widely used in text sentiment classification,but the distributed word vectors only contains semantic information and ignores emotional and attribute information of words.A lot of text feature extraction research only revolves the word and ignores the structure information of the sentence.To overcome the above shortcomings,work of this research is as follows:1.This research presents the W-P sentiment representation model and the fusion sentiment feature expression model to improve the method of text sentiment feature representation.Using word2vec tools and sentiment dictionary,the W-P sentiment representation model trains distributed word vectors which contains semantic information,emotional information and part-of-speech information.Based on the W-P sentiment representation model and using Bi-LSTM networks,the fusion sentiment feature expression model trains distributed word vectors which contain contextual information semantic information,emotional information and part-of-speech information.The effectiveness of the two models is verified by experiments.2.Based on the fusion sentiment feature expression model,this research presents sequence structured text feature extraction model to improve text sentiment feature extraction.Sequence structured text feature extraction model uses attention mechanism and CNN networks to extract the text information.This model can realize the multi-level extraction of text information and reduce the loss of information.The experiment shows that the effect of the model classification is greatly improved.