Research On Aspect Level Emotion Analysis Based On Multi Semantic Learning

With the rapid development of the Internet,more and more netizens can express their preferences and emotional attitudes to affairs through the platform,such as restaurant reviews,notebook reviews,etc.The text of these reviews contains a lot of valuable information.For consumers,they can make judgments on the quality of products based on existing reviews,and for merchants,it can help them improve product quality and enhance their competitiveness.In this case,sentiment analysis came into being.Sentiment analysis is an important research direction of natural language processing.It is becoming more and more important to sort out and analyze complicated data with the method of sentiment analysis.However,traditional sentiment analysis generally divides people's emotional attitudes into three aspects: positive,negative and neutral,but this kind of coarse-grained sentiment analysis is increasingly unable to meet people's needs,while aspect-level sentiment analysis for specific entities can better solve this problem.As a subtask of sentiment analysis,aspect-level sentiment analysis can analyze the refined sentiment in sentences.Compared with document-level sentiment analysis,aspect-level sentiment analysis requires more semantic features for more refined semantic analysis.At present,the research methods at home and abroad still have some shortcomings.Supervised aspect-level sentiment analysis is limited by the conflict between the sophistication of sentiment classification and the small number of corpora.In addition,the corpus with many neutral words has always been a difficulty in aspect-level sentiment analysis,and the current research on the perception of aspect words is obviously insufficient.In order to better obtain refined emotion representation,this paper proposes the following specific research methods.(1)By migrating the extensively pre-trained BERT model to aspect-level sentiment analysis,the problem of the small aspect-level sentiment analysis corpus is effectively addressed.At the same time,in order to realize aspect-level sentiment analysis centered on aspect words,a multi-semant-ic learning model based on BERT(Bidirectional Encoder Representation from Transformers)is proposed,which consists of left local semantics,right local semantics,aspect target semantics and global semantic learning modules.(2)In order to better capture the semantic dependencies between context words and target aspect words,this paper proposes an aspect-level perception enhancement method based on BERT and multi-head attention mechanism.First,perform the average pooling operation on the aspect words to obtain the same aspect semantic information as the BERT hidden layer state,and then link the average aspect semantics with the left local semantics,right local semantics and global semantics to each hidden layer state respectively.Finally,it is incorporated onto each semantic learning module through a linear transformation layer and a multi-head attention mechanism operation,resulting in aspect-target-aware enhancement of emotional semantics.(3)In order to realize the complementarity between semantic features,two methods of semantic fusion are proposed.One is multi-level semantic fusion.First,the corresponding hidden states are spliced in the left local semantics and the right semantics,and then the fused local semantics and all semantic information are spliced according to the state of the corresponding hidden layer;the other is a gated semantic combination,which uses a gating mechanism to measure the degree of different contributions of the three semantics to sentiment polarity classification,and highlights key semantic information.It can better deal with corpora with more neutral words.Based on this model,this paper proposes an improved method,which introduces a semantic refinement layer similar to CNN(Convolution Neural Networks)to obtain a more refined semantic representation.Through the above methods,multi semantic learning model based on multi-channel convolution proposed in this paper improves the accuracy by an average of 0.5% in five internationalized corpora.