Research Of Text Sentiment Analysis Technology Based On LSTM

Sentiment analysis is to identify the emotional polarity(positive or negative or neutral)or emotional intensity of a given piece of text.Sentiment analysis is mostly used in product reviews and statistical monitoring of public opinion,which helps to support the decision-making of businesses and the media-oriented guidance of agencies.However,the construction of sentiment lexicon is time-consuming and labor-intensive,and it is not easy to maintain the sentiment lexicon.And the manual feature extraction requires the knowledge of the expert level.However,the popular word vector model is mainly based on contextual information learning,it pays attention on semantic information without the emotional information which sentiment analysis focused on.On the other hand,The rise of network(RNN)has further solved the problem of serialization,and as an improved model of RNN network,long-term memory model(LSTM)can make better use of long-distance dependence information in sequence data and is suitable for text Sentiment classification problem.This paper analyzes the shortcomings of RNN network and the way LSTM model solving the long-distance information dependence.On this basis,we compare and analyze the derivative model of LSTM model and choose BLSTM(bidirectional LSTM)as the benchmark model proposed in this paper.The emphasis of the work is to put forward a general LSTM model of emotion analysis,which using word embedding technology and BLSTM.The texts embedding-layer processed are vectorized and sent to BLSTM for feature extraction.Finally,they are sent to the softmax classification layer for classification.On the one hand,Adam optimization algorithm is applied in the model to reduce model convergence time,on the other hand dropout technology has been added to improve the anti-over-fitting ability on small data sets.In addition,this paper proposes a hybrid model based on the generalized LSTM sentiment analysis model named AT-BL&C,which proposes a word vector that integrates sentiment information at the embedding level.The introduction of sentiment information can improve the performance of word vector technology in sentiment analysis tasks,and then added a attention mechanism that can increase the weight of key information at the bidirectional LSTM level.Finally,the AT-BLSTM and CNN are merged with the CNN's ability to extract global information,the AT-BL&C model further enhances the accuracy in sentiment analysis task.