Aspect Based Sentiment Analysis Based On Gating Mechanism And Adversarial Learning

In recent years,the rapid development of the domestic Internet industry,including social network,e-commerce,information flow and other fields,has profoundly changed the public life.User reviews play an important role in a large number of Internet services.As the carrier to express user viewpoints,user reviews contain a lot of emotional information and huge business opportunities.Therefore,text sentiment analysis has become a hot research topic and industrial application in recent years.Most of the existing researches on sentiment analysis focus on the coarse-grained analysis to judge the overall sentiment of the text,but there are often multiple aspects for the same target in the same text,and the corresponding sentiment polarities are not the same.Therefore,this paper studies the aspect level sentiment analysis.The main contents include:The existing methods often take the evaluation object as a whole and lack of finegrained modeling for different aspects.This paper proposes the gated interactive networks for aspect level sentiment analysis.This method separately models the attention mechanism for aspect entities and attributes under the same context,and analyzes the interaction between aspect entities and attributes.Finally,independent and interactive information are fused through gating mechanism to obtain the final judgment of sentiment polarity.The experimental results on Sem Eval 2015 and Sem Eval 2016 aspect level sentiment analysis datasets show that compared with the highest results of non-BERT model,the evaluation index F1 value is improved by 1.65% and 1.50% respectively,compared with the model based on BERT,the F1 value is improved by 1.68% and 1.25%respectively.The method based on deep learning needs a large number of annotation data for training,but the cost of obtaining annotation data is often high,and the scale of public dataset of aspect level sentiment analysis task is small.In order to solve this problem,this paper further proposes the adversarial multi-task learning method for aspect level sentiment analysis,through adversarial training to learn aspect sentiment features which aim to improve the classification performance of the model.Firstly,the method constructs pseudo samples by exchanging aspect terms,and then designs feature extractor and discriminator to judge whether the samples are pseudo.Finally,based on multi task learning,auxiliary tasks including judging the sentiment polarity and aspect category,are added to learn the aspect independent and interactive features in the samples,which aims at improving the performance of aspect sentiment analysis model.The experimental results on Sem Eval 2015 and Sem Eval 2016 aspect level sentiment analysis datasets show that compared with the existing models,the proposed model improves the F1 value by3.39% and 3.96% respectively,reaching the state-of-the-art performance.