Research On Text Sentiment Analysis Technology Based On Deep Learning

With the rapid development of Web2.0,users generate a lot of emotional text information through social networks,blogs,online portals,etc.These textual information are usually related to social hot issues,commodity services,product evaluation and many other fields.Analyzing and using these textual information can bring great social significance and commercial value for market research,product or service review mining,network public opinion discovery and early warning,etc.Therefore,mining text emotional information has been a hot topic in industry and academic circles in recent years.At present,the methods of sentiment analysis can be divided into three categories: the methods based on emotional dictionary,the methods based on traditional machine learning and the methods based on deep learning.It costs much for the methods based on emotional dictionary to construct and maintain emotional dictionary in diversity areas.The methods on the basis of traditional machine learning are usually based on the Bag of Words(Bo W)model,resulting in different sentences made up of the same words have the same feature representation,so the models cannot learn the text features accurately.The deep learning model can handle and use massive textual data and achieve full-scale capture of deep semantic features,which final realizes end-to-end task of text sentiment analysis.However,the existing deep learning models applied to Chinese text sentiment analysis are word-level networks,which use words as the basic processing unit,ignoring the internal structure information of the words.Those methods cannot fully learn text features from a single perspective,which causes poor performance in Chinese sentiment analysis.Based on the study of traditional sentiment analysis methods,current existing text sentiment analysis based on deep learning combined with the characteristics of Chinese,this paper analyzes and designs the following two sentiment analysis models based on deep learning:(1)The text sentiment analysis model based on dual-channel convolutional neural network and bidirectional long-short term memory network(DC-CNN-BLSTM)was constructed.At First,text is transformed into word vectors and character vectors by word embedding and character embedding and then taking word vectors and character vectors as the input of two channels respectively.In this model,convolutional neural network(CNN)is used to extract hidden features from two perspectives of word vectors and character vectors and then bidirectional long-short term memory network(BLSTM)is utilized to learn text sequence features.Finally,the merge layer is introduced to merge the two extracted features and combine them for joint learning.The experimental results show that the proposed model in this paper can effectively improve the performance of Chinese text sentiment analysis and its accuracy,recall rate and F1 values are significantly improved compared with single-channel methods,traditional machine learning algorithms and other deep learning models.(2)Since different words and characters in text have different key effects on sentiment classification,in order to highlight the influence of important information and reduce the interference of useless information in the text,this paper introduces the attention mechanism to DC-CNN-BLSTM.The model of dual-channel convolutional neural network and bidirectional long short-term memory network with attention mechanism(DC-CNN-BLSTM-Attention)is designed.The attention mechanism can help the model to obtain semantic encoding contained the attentional probability distribution,effectively highlight the words and characters in the text that are more critical to the analysis task.The experimental results show that attention mechanism can effectively improve the performance of text sentiment analysis.