Research On Sentiment Text Classification Based On Deep Learning

With the development of the Internet,various online music applications have emerged.Due to individual differences in music preferences,conducting sentiment analysis and research on music reviews can effectively target specific users and improve music products and services.Compared to sentiment analysis in other domains,music reviews possess the following characteristics.Firstly,the semantic expressions in music reviews are diverse,with the use of metaphors,metaphysics,exaggeration,and other rhetorical techniques.Therefore,accurate identification of emotions requires a high level of semantic analysis capability.Secondly,the language styles in music reviews are highly diverse.Sentiment analysis in the field of music reviews needs to consider the processing of different language styles,such as internet slang or specific jargon,requiring the model to possess corresponding language processing abilities.Additionally,deep learning models commonly used in the field of Natural Language Processing(NLP)are often large-scale and require substantial computational resources for training and inference,making it challenging to deploy them effectively on certain embedded or mobile devices.(1)To address the issues of semantic diversity and complex sentiment tendencies in Chinese music reviews,this study proposes a Chinese sentiment analysis method based on BERT-wwm-Bi LSTM-SVM.Given the scarcity of relevant datasets in the music review domain,this research employs web scraping techniques to collect relevant comments on popular songs from mainstream music applications.After employing data cleaning techniques and sentiment polarity labeling through relevant software,a music dataset is obtained for sentiment analysis using the proposed model.Building upon the existing BERT-based sentiment classification,this study utilizes the BERTwwm model,which is more advantageous for Chinese processing,and incorporates sentiment polarity features into the input vectors to generate word embeddings containing emotional and hierarchical semantic information.Additionally,a Bi LSTM model is employed to further extract feature information,followed by SVM for classification.Experimental results demonstrate that compared to other models,the BERTwwm-Bi LSTM-SVM model shows improvements in accuracy,recall rate,and other aspects when applied to sentiment analysis in music reviews.(2)To address the challenges posed by the large scale and deployment of deep learning models,this study explores knowledge transfer from BERT to Bi LSTM based on the knowledge distillation method of joint training with soft and hard labels.The teacher model and student model for Chinese sentiment classification tasks are optimized.The BERT model is replaced with BERT-wwm in the teacher model to enhance the model’s performance and Chinese processing capabilities.An attention mechanism is incorporated into the Bi LSTM student model to improve its representational capacity and accuracy.Furthermore,two data augmentation methods are introduced during the training of the student model to increase data diversity and richness,enhancing the model’s generalization and robustness.Experimental results indicate that these improvements and optimizations can further enhance the distillation effect of the knowledge transfer task from BERT to Bi LSTM based on joint training with soft and hard labels,enabling better application of sentiment classification tasks while significantly compressing the model’s scale.