| Fine-grained sentiment classification is an important method for text sentiment analysis.To solve the problems of lack of attribute location information in existing finegrained sentiment classification methods,joint training of multiple tasks and the vulnerability of the algorithm to unbalanced sample distribution,this paper investigates and proposes two sentiment classification methods from the perspectives of contextual dynamic weight adjustment and gradient coordination based subtask interactive learning,and verifies the effective feasibility of the methods through experiments.At the same time,the proposed sentiment classification methods are applied to the restaurant evaluation analysis scenario,and a restaurant evaluation analysis system based on end-to-end fine-grained sentiment classification is designed and implemented.The main research work and results include:1.An aspect sentiment classification method based on BERT and contextual dynamic weighting strategy is proposed.Considering the impact caused by the location of attributes in the context on sentiment feature extraction,the method uses BERT for word embedding and extracts global contextual semantic features,designs a contextual dynamic weighting strategy for extracting local contextual semantic features,and introduces a multi-headed self-attention mechanism for fusion of the two features to extract effective sentiment features and perform aspect sentiment classification.Experimental results on Restaurant,Laptop,and MAMS datasets show that the proposed strategy and method can effectively improve the sentiment classification accuracy and performance.2.An end-to-end fine-grained sentiment classification method based on shared features and gradient coordination mechanism is proposed.The method draws on the design ideas of Encoder-Decoder framework and designs ATE component and ASC component to output attribute label sequences and sentiment polarity label sequences respectively to ensure all attribute and sentiment polarity binaries in the output utterance.Meanwhile,the interaction between attribute extraction and sentiment polarity prediction subtasks is enhanced by sharing the hidden layer vector representation among subtasks,the mean pooling technique is used to solve the problem of sentiment inconsistency,and the gradient coordination mechanism is introduced to improve the loss function to alleviate the impact of unbalanced sample distribution.experimental results on the datasets of Restaurant,Laptop,and Twitter show that the proposed method,by sharing features and introducing a gradient coordination mechanism,can effectively learn the correlation between attribute extraction and sentiment polarity classification subtasks,achieve end-to-end finegrained sentiment classification,and effectively improve the performance of joint attribute and sentiment polarity classification.3.An end-to-end fine-grained sentiment classification-based restaurant evaluation and analysis system is developed using Spring Boot and Vue technologies.The system includes the main functions of login and registration,data collection and pre-processing,fine-grained sentiment classification,user management,calculation engine,evaluation data management,etc.The sentiment classification method proposed in this paper is successfully used for practical restaurant evaluation analysis.Tests show that the system has certain advantages in mining the emotional information contained in the texts of consumers’ reviews of restaurant merchants,etc.,and meets the actual needs of users. |
