Research On Targeted Aspect-based Sentiment Analysis Based On Deep Learning

With the rapid popularization of smart terminals and the rapid development of information technology,the number of user-reviewed texts has grown exponentially.Analysing the opinions that contain user sentiments can not only promote the development of public opinion monitoring,marketing strategies and other applications,but also be used for the precise optimization of conversational recommendation systems.How to accurately perform sentiment analysis has become an urgent requirement in the current industrial and academic circles,which has also greatly promoted sentiment analysis to become a hot research direction in the field of natural language processing.The previous deep learning-based sentiment analysis models mainly deal with the sentence as a whole,but it is difficult to effectively recognize the sentiment of the sentence in terms of fine-grainedness.At the same time,the user reviews are mostly short texts,which makes the existing models unable to fully mine the sentiment features information.Therefore,this paper focuses on fine-grained sentiment analysis tasks,and researches from shallow to deep based on deep learning technology and previous research results.The main research work is as follows:Firstly,for the aspect-based sentiment analysis task,a deep network model GRCNN-HBLSTM based on the gate mechanism and cascade network is proposed.The model first fuses character-level convolutional networks with traditional word vectors as input to the gate mechanism in part of the word vector representation to obtain the final word vector representation,and then uses the word vector along with aspect words and topic words as inputs to subsequent models.Then,through the combination of the regional CNN and the bidirectional LSTM network based on the hierarchical mechanism to obtain the sentiment feature information of specific aspects,it also improves the training speed.Compared with the experimental results of existing typical models,it is verified that the model has better performance on aspect-based sentiment analysis tasks.Secondly,for the targeted sentiment analysis task,a dynamic adaptive hierarchical network model GC-HLSTM is proposed.The model continues to optimize based on the aspect-based sentiment analysis model.The main optimizations are in two aspects: one is the word vector representation part,which replaces the traditional word2 vec with the ELMO mechanism,and then obtains the final word vector representation through the gate mechanism.The other is a hierarchical attention network is established for this task by combining the hierarchical mechanism and the attention mechanism to obtain the sentiment characteristics of the specific target at the word level and the sentence level.Compared with the experimental results of existing typical models,the effectiveness and better performance of the model on targeted sentiment analysis tasks are verified.Thirdly,for more fine-grained targeted aspect-based sentiment analysis tasks,considering that the sentiment analysis model at the aspect level and the target level cannot capture the emotional connection of the target and the association at the same time,a deep network model SRC-TBERT based on BERT pre-training and TextCNN is proposed.The model first proposes a BERT-based mask data enhancement method for the shortcomings of today's small samples that lead to insufficient semantic understanding,further expanding the sample size.Then the model transforms this sentiment analysis task into sentiment reading comprehension task and fine-tuning based on the constructed TBERT pre-trained model.Finally,compared with the experimental results of existing advanced models,it is verified that the model has stronger performance on the targeted aspect-based sentiment analysis task and also improves the training speed.