Research Of Review Aspect Level Sentiment Analysis Based On Hierarchical Bi-LSTM Model

In recent years,with the rise and development of feedback websites such as shopping websites,can we accurately grasp the emotional attitudes of users from multiple aspects of products or events from the massive comment data of the Internet,and improve products and consumers from enterprises.It is extremely important to consider products and the government to accurately grasp the public opinion.Emotional analysis based on aspect level has become one of the hot research directions in the field of natural language processing.Compared with traditional sentiment analysis,Aspect-Based Sentiment Analysis(ABSA)can mine the emotional polarity corresponding to different review objects,and the level of analysis is deeper than traditional sentiment analysis.The research object and goal are different between traditional sentiment analysis and aspect-level sentiment analysis.The former research object may be a document or a sentence,the goal is to judge the emotional polarity of a whole text or a clause;and the research object based on the aspect level of sentiment analysis is a review object,and the goal is to analyze the emotional pole of multiple review objects in the text.In recent years,Recurrent Attention Network on Memory(RAM)combined with attention mechanism has achieved great success in the research of aspect-level sentiment analysis.RAM not only has the characteristics of simple structure,but also has the advantage of fast running speed because the model contains memory network.At the same time,the existence of attention mechanism makes the deep memory network get the context word in the process of inferring a certain comment object.The importance of playing.This paper proposes two improved methods based on the RAM model.The main contents are as follows:1.Since the position weights in the position weighting module in the existing RAM model are calculated by the heuristic method,the position weights obtained by this calculation method are relatively fixed,and four methods are proposed for this problem.The new position weight calculation method achieves the purpose of better describing the position information by dynamically calculating the position weight value.In this paper,the position weights are extended,and the weights based on part of speech,the weights of convolutional neural networks based on part of speech and the weights based on sentiment lexicon are proposed.By introducing part-of-speech attributes and sentiment dictionary information,the emotional words and comment objects can be more accurately obtained.The relationship between emotional polarity,thereby improving prediction accuracy.2.This paper also proposes a RAM-based improved model based on hierarchical bidirectional LSTM network,which has the advantages of retaining word order information in sentences acquired by LSTM network and retaining long-distance dependence,and has the advantages of layering.Integrate the links between other clauses on the current clause.Because the emotional polarity of the different review objects described in a review often differs little,this model can integrate the relationships between clauses.This paper selects the laptop and restaurant data sets in SemEval 2014 Task 4 as the experimental data set.Through comparative experiments,it is found that the improvements proposed in this paper can achieve better results.