| With the rapid development of information technology,commenting on the Internet has become an important way for people to express their opinions and convey their experiences.By analyzing the review text,it helps enterprises to fully understand the user’s opinions and attitudes towards a product or service,so as to make targeted improvements and make better business decisions.Therefore,for aspect-level sentiment analysis tasks,the Research has important economic value.The aspect-level sentiment analysis method based on deep learning and attention mechanism is the current mainstream research direction,but the existing research also has shortcomings.First,the traditional attention mechanism makes the model pay too much attention to high-frequency words and ignore low-frequency emotions.influence of words.Secondly,most researchers currently conduct research on aspect term extraction or sentiment classification tasks alone,resulting in models that cannot adapt well to complex and changeable scenarios,and the accuracy is usually not high.Therefore,how to make the attention mechanism pay corresponding attention to low-frequency emotional words and be more suitable for comment texts in real life is of great significance for improving the effect of aspect-level sentiment analysis models.This paper focuses on these issues of aspect-level sentiment analysis.The main research work and innovations are as follows:(1)Propose an aspect category sentiment classification method based on BERT and an improved self-attention mechanism.This paper improves the ATAE-LSTM model,obtains the word embedding representation of the input text through the BERT pre-training model,and then uses the bidirectional LSTM to extract the text information to obtain the global features of the input text.In addition,in order to alleviate the problem of focusing on high-frequency emotional words and ignoring lowfrequency emotional words in the traditional attention mechanism,K-round attention training is added to the attention layer to gradually extract emotional words in sentences,and improve the model’s ability to understand low-frequency emotional words.At the same time,the emotional words are classified according to the influence of each emotional word on the emotional classification,and the attention score of the wrong emotional words is reduced,thereby improving the overall emotional classification effect of the model.The experimental results show that the models proposed in this paper have good performance.(2)An end-to-end sentiment analysis method is proposed to solve the tasks of aspect term extraction and sentiment classification.In order to improve the accuracy of the sentiment analysis task,this paper adopts the digital sequence form composed of a mixture of pointer index and sentiment class index,and converts the two sub-tasks of aspect term extraction and sentiment classification into a unified generation formula.At the same time,in order to reduce the impact of distant irrelevant sentiment words on sentiment classification tasks,this paper adopts two local context processing methods,Context Dynamic Masking(CDM)and Context Dynamic Weighting(CDW),to more accurately establish aspect terms and sentiment.relationship between words.Experiments show that on public datasets,the F1 value of the AESC-BART model proposed in this paper is improved by about 1% compared to other similar studies. |
