Study On Aspect-level Sentiment Classification Based On Retaining Progressive Self-supervised Attention Mechanism

Contemporary e-commerce is developing rapidly,5G era will bring us more and more data.A large number of consumers will post their comments on various products on the Internet platform.The main purpose of aspect-level comment sentiment analysis is to use the neural model with attention mechanism to obtain the importance of each context word in a given aspect according to the comments of Internet users,so as to show that consumers are different about a product or event.The aspect of sentiment evaluation tendency,aspect-level sentiment classification mainly includes two aspects:the extraction of aspects in a sentence and the sentiment classification of a given aspect.However,existing research tends to focus on some frequently-occurring emotion words and ignore less-occurring emotion words.This thesis conducts an in-depth study on this issue,the main work is as follows:(1)Propose a progressive self-supervised attention emotion classification algorithm,which can automatically mine useful attention supervision information from the training set to optimize the attention mechanism.The core idea of the algorithm is to apply this method of automatically mining information to iteratively predict emotions on each training set sample.In each iteration,the word with the largest attention weight will be extracted if the sentiment prediction is correct in this sentence The extracted words are extracted as positive words.If the sentiment prediction of this sentence is wrong,the extracted words are extracted as misleading words,and the misleading words are given very low weight to participate in the next iteration.Finally,after the iteration of the entire training set is completed,after reducing the weight of misleading words,the weight of the positive words extracted is equally distributed,and then the aspect-level sentiment classification is performed.(2)Propose a classification model based on the above algorithm AS+TNet-ATT model,the model uses the three types of public data sets to buy laptops,Twitter,restaurants for comparative experiments,using supervision The information memory network and the conversion network accept input information.Because the attention mechanism will give higher weight to the emotional words that appear more frequently,before training the data set,an effective set of words to assist subsequent training must be extracted.This method is mainly by repeatedly iterating the words with the highest weight in each data set sample,and then extracting them to give a lower weight,so that its attention is reduced,and then subsequent iterations extract words,other than other Words with low and low attention are more visible to the model,thus extracting more effective information of the sentence,and providing more effective semantic support for the final aspect-level sentiment classification.Finally,the effectiveness of the algorithm and model is verified through experiments.