Research And Application Of Text-related Sentiment Analysis Based On Grid Tags

Sentiment analysis is an analysis of people’s attitudes,opinions,and emotions towards entities such as products,services,and organizations.Text sentiment analysis means that sentiment analysis takes text as a unified carrier.At present,text emotion analysis has grown into one of the most active research fields in natural language processing.In recent years,many directions have been subdivided in this direction,such as aspect-based sentiment analysis,language-based sentiment analysis,platform-based sentiment analysis,target-based sentiment analysis,and aspect-based sentiment triplet extraction.Sentiment analysis has applications in various fields,such as life services,financial analysis,social management,and national security.This paper studies the task of aspect-based sentiment triplet extraction.Its purpose is to identify aspects,viewpoints and sentiment polarity from the text,and form them into a triplet.This task is very useful for understanding users’ views and needs.It can analyze the sentiment information in the text in a more granular manner;At the same time,this task is also very challenging and needs to deal with multiple sub-problems at the same time,such as the identification,extraction and correlation of aspects,views and sentiment polarity.This paper mainly aims at the problem of error propagation in multiple subtasks of the grid tag scheme model of aspect sentiment analysis.In order to improve the accuracy and generalization of model recognition,the following research work has been done in this paper:(1)In view of the rationality analysis of grid label scheme annotation,this paper proposes a more optimized compression annotation strategy,and uses rotary position coding to enhance the position relevance of word pairs.The original grid label annotation scheme has dimension redundancy in design,and six dimensions are used to represent six states.However,due to the difference of distribution locations,the same dimension identification can be used in different distribution areas,so dimensions can be compressed.However,the difference of different locations is not learned in the original model,so this paper uses rotary position coding to learn word to position information,This method can ensure a slight improvement in the model effect while compressing dimensions.(2)To solve the problem of poor generalization ability of the model,a mutual learning grid tag algorithm is proposed,and two data expansion schemes are proposed.The principle of this method is to associate two grid label models with different parameters,train them with the same input samples,and let the two models learn from each other during the training process.The KL divergence is used as the loss training.The experiment shows that the model has a good effect.Finally,through the two data expansion schemes proposed,the learning ability and generalization ability of the mutual learning scheme are further verified.At the same time,the method has less data The label is also obviously effective in case of poor quality.(3)Aiming at the neglect of details in grid label scheme,this paper proposes a tokenword grids label algorithm.First of all,each sentence is composed of multiple words.However,after BERT’s tokenizer,the token does not correspond to the word one by one.Many words will be split into multiple tokens.In the original grid label scheme,word level training and prediction are used.The method used in this paper is to use token level labeling network training first,and then word level prediction.This method is logical and reasonable,and the experimental results show better performance,It is proved that the method is effective.