A Hybrid Sentiment Classification Model for Course Comments Based on 'Unstable Interval' Correction - An Applied Study Integrating SVM and BiLSTM

Abstract

Aiming at the limitations of mainstream text sentiment analysis methods (poor generalization of dictionary-based methods, machine learning relying on labeled data and being prone to overfitting), this paper proposes a hybrid model combining sentiment dictionaries and machine learning. The 'unstable interval' refers to the interval where the sentiment dictionary score is close to 0 (the boundary between positive and negative emotions) with low classification accuracy, and its calculation and derivation are based on the test set score distribution, and the interval range is determined by traversal. The model fusion method is: the sentiment dictionary and SVM/BiLSTM output polarities respectively, the consistent part is retained, and the inconsistent part in the 'unstable interval' is based on the SVM/BiLSTM result. Taking 3119 course comments from the NetEase Cloud Classroom platform as the experimental data source, the experimental results show that the accuracy of the hybrid model is increased to 88.9% (SVM sub-model) compared with the single SVM model (82.3%), and to 91.2% (BiLSTM sub-model) compared with the single BiLSTM model (85.6%); at the same time, it is clarified that positive emotions mainly come from 'practical course content' and 'clear explanation by lecturers', while negative emotions focus on causes such as 'delayed course updates' and 'untimely customer service feedback', and the key factors in each cluster highlight the course's strengths and weaknesses. Researchers can utilize these typical emotional factors as evidence for dynamically adjusting course content.