Research On Aspect-level Sentiment Analysis Method Based On Bi-LSTM

With the progress and development of the Internet,people are accustomed to freely expressing their feelings or expressing their opinions on the Internet,and at the same time,they also hope to obtain references to certain aspects of things through the Internet.Aspect-Level Sentiment Analysis can help people obtain a certain aspect of sentiment information,which in turn affects people's attitudes towards something,so it has become one of the popular areas of Sentiment Analysis.In recent years,with the popularity of deep learning,deep learning methods such as recurrent neural networks and convolutional neural networks have achieved great success in the research of Aspect-Level Sentiment Analysis.But these methods all lack the degree of attention to the aspect information to varying degrees.To solve this problem,this paper proposes an Aspect-Level Sentiment Analysis Method based on Bi-LSTM.This method studies two subtasks of Aspect-Level Sentiment Analysis.For Aspect Extraction Task,it can be regarded as a sequence labeling problem.This paper proposes a method based on CNN and Bi-LSTM-CRF for the characteristics of sequence labeling.This method first integrates part-of-speech features into the CNN model,uses the CNN model to extract local features,and then uses the Bi-LSTM model to extract contextual features.After the three features are combined,the CRF is used to predict the tag sequence to get more accurate aspect category extraction.As for Aspect-Level Sentiment Classification Task,previous studies have shown that attention is paid to aspect information.However,when a sentence is composed of multiple aspects of information,the emotion prediction will still be inaccurate.For this reason,this paper proposes a method based on Bi-LSTM and Attention.This method first adds three gating units about Aspect Information on the basis of Bi-LSTM,then introduces attention mechanism,and fuses Bi-LSTM hidden layer features integrated with aspect information with global attention features and aspect attention features,so as to improve the accuracy of sentiment prediction.In order to verify the effectiveness of the method proposed in this paper,a comparison experiment with other classic models is done on the data set provided by Sem Eval and the Twitter data set.The experimental results show that the method proposed in this paper has achieved higher accuracy and F1 value than the existing methods,which proves that the methods in this paper are both practical and effective.