A Study on Sentiment Analysis of Movie Reviews based on ALBERT-TextCNN-HAN

Abstract

With the proposal of Albert model and the fusion application of Textcnn model, the ability of emotion analysis has been greatly improved. However, the Albert-TextCNN model is limited in its capacity to identify key words and sentences in the text, impeding the formation of a comprehensive text representation through these important elements. To address the shortcomings of the Albert-TextCNN model, we have proposed adding an attention mechanism to the model. However, since single-word-level attention fails to fully consider the significance of text hierarchy and inter-sentence relationships, the Albert-TextCNN model is combined with the Han model (Hierarchical Attention Network) to enable the model to allocate different levels of attention to sentences and words of varying importance in the text. This two-level attention mechanism better extracts key information in comments. Compared to the original model, the Albert-TextCNN-Han model more accurately distinguishes between different emotions and improves the performance of emotional analysis in film reviews. This improved accuracy can help individuals select movies that align with their preferences, and manufacturers can precisely obtain customer feedback.