Research on Fine-Grained Sentiment Classification

Abstract

Aiming at the fine-grained sentiment classification that distinguishes the emotional intensity, the commonly used dataset SST-1 is analyzed in depth. Through the analysis, it is found that the dataset has serious problems such as data imbalance and small overall scale, which seriously restricts the classification effect. In order to solve the related problems, data augmentation method is adopted to realize the optimization of the dataset. The IMDB and other data which are relatively homologous to the original dataset are annotated, and the focus is to expand the categories with fewer numbers. By this way, the problem of data imbalance is effectively alleviated and the original data scale is expanded. Then, based on the Bidirectional Encoder Representations from Transformers (BERT) model, which has good overall performance on natural language processing, the benchmark classification model is built. Through multiple comparison experiments on the original dataset and the enhanced data, the influence of the deficiency of the original dataset on the classification effect is verified. And, it is fully demonstrated that the enhanced data can effectively improve the test results and solve the problem of large differences in performance between different categories well.