Multi-granular Text Sentiment Classification For Method Research Based On Machine Learning

In recent years,as the Internet has gradually penetrated people's lives,users have become accustomed to sharing and expressing their opinions on major platforms.These comment texts containing a high volume of user behavior data provide strong data support for text sentiment classification research.As a hot research topic in the field of natural language processing,text sentiment classification has been attracting a lot of researchers' attention for model research and application expansion.Text sentiment classification can be divided into sentence-level text sentiment classification and target-level text sentiment classification on the basis of different text sentiment granularity.When a sentence contains multiple target entities and sentiments,the sentence-level sentiment classification task will have a large sentiment judgment bias.To address this limitation of the sentence-level sentiment classification task,this paper conducts a study on the target-level text sentiment classification task that is more in line with people's cognitive habits.Although many results have been achieved in the model research of text sentiment classification tasks,there is still much room for improvement.First of all,previous sentence-level sentiment classification models have paid limited attention to the information loss caused by the multi-layer mapping of a large number of nonlinear functions in the neural network model.Second,a large number of model researches in the target-level text sentiment classification have not fully utilized the task characteristics of dividing a sentence into three parts by target words.At the same time,previous studies have not considered the feasibility and application value of sentiment classification models.In this paper,we design two deep learning sentiment classification models based on the shortcomings of existing models for sentence-level and target-level text sentiment classification tasks,respectively,and obtain the overall score of each store and the results of word-of-mouth analysis within the region through case studies.The main work and contributions of the paper are summarized as follows.(1)In the sentence-level text sentiment classification task,this paper proposes a feature fusion-based adversarial recurrent neural network model(FARNN-Att).Aiming at the problem of information loss caused by the sparse activation of the nonlinear activation function,the FARNN-Att model is based on the bidirectional long short-term memory network(Bi-LSTM),which not only makes full use of the pre-trained word embedding information by constructing a feature connection layer but well alleviates the information loss of the model in the process of forwarding propagation.In addition,the performance and robustness of the FARNN-Att model are improved by adding an attention mechanism and an adversarial training module.Finally,the experiments verify the excellent performance and effectiveness of the FARNN-Att model.(2)In the task of target-level text sentiment classification,this paper proposes a three-channel feature-enhanced deep interaction network model(TFEI).The previous research failed to introduce the task characteristics of three-part sentences of the target words into the pre-trained language model and fully excavate it.To thoroughly stimulate the application potential of this feature in target-level text sentiment classification tasks,the TFEI model is based on the BERT language model and cleverly used this task characteristic to construct a three-channel feature extraction method.At the same time,the TFEI model also uses different channel input forms and interactive learning to more fully explore the semantic relationship between the target word and the context.Finally,the experiments verify the outstanding performance and effectiveness of the TFEI model on all data sets.(3)In the model application and case analysis,the application of the FARNN-Att model in word-of-mouth evaluation is studied.Based on the prediction results of the FARNN-Att model,the prediction results of the model and the number of reviews are combined to construct an overall evaluation model of the store,thereby obtaining the score of each store.Finally,we analyze the relationship between store scores and per capita consumer prices and the number of reviews and visualize the spatial trends and differences of store scores within the scope of the research.