Deep Learning Based Sentiment Analysis Research

Sentiment analysis is one of the fundamental task in the area of Natural Language Processing,and its purpose is to predict the sentiment polarity of text.In recent years,due to the development of the Internet and mobile Internet,various social networks and online shopping platforms have accumulated a large amount of user-generated texts.Analyzing these text data has a positive effect on improving the service quality of social networks and online shopping platforms.Compared with traditional statistics or machine learning based sentiment analysis models,a series of deep learning based models have greatly improved the performance and robustness of sentiment analysis models.This thesis focus on deep learning based sentiment analysis.Our works can be introduced as following:1.Domain adaptation based Coarse-grained sentiment analysis.Insufficient training data for deep learning models is a basic problem.This thesis attempts to solve this problem by using domain adaptation.Adversarial training based sharing-private model,shared feature extractors and specific feature extractors are trained on the source domain data.Among them,the shared feature extractor can be directly applied to the unlabeled target domain.Then select the appropriate source domain to transfer the specific feature knowledge to the target domain through adversarial training.Finally,combining shared feature and specific feature to complete the coarse-grained sentiment analysis.Compared with the previous method,the information loss of specific feature is compensated by domain adaptation.Both Amazon reviews and FDU-MTL have achieved competitive results on both datasets.2.Aspect based sentiment analysis is a branch of fine-grained sentiment analysis.This thesis uses the LSTM to encode the context and aspect information respectively,introduces the position weight vector,and introduces word-level attention mechanism to reduces the information loss caused by average pooling operation in the previous attention mechanism calculation.Experiments on three dataset SemEval restaurant 2014,SemEval laptop 2014 and Twitter get a significant improvement over all baseline models.3.Aspect term sentiment analysis involves aspect term extraction and term sentiment analysis two subtasks.Normal researcher deal with these two subtasks independently,which affect their practical use.This thesis attempts to solve a complete aspect term sentiment analysis task through an end-to-end model.The unified labeling mode is adopted,and the complete task is treated as a sequence labeling task.Experiments show that the unified labeling scheme based on sequence labeling performs better than the pipeline model.BERT based models can achieve significant performance improvements by finetuning.Experiments show that BERT based model for downstream task characteristics,networks based on self-attention mechanisms such as BERT-SAN and BERT-TFM get better performance.