A CNN-Transformer model for short Chinese texts sentiment analysis
Hengrui Hu · 2024
Sentiment analysis of short Chinese texts holds significance in preventing cyber bullying and fostering a favorable network environment. Due to the limited contextual information they contain, sentiment analysis of short Chinese texts presents a level of difficulty. Existing methods have typically focus on either traditional pipeline method or deep learning methods, e.g. Transformer. However, they are inadequate to capture specific local features and have high computational complexity when processing sequences. To resolve the above constraints, this paper introduces a convolutional neural network structure with stronger local feature extraction capability to improve Transformer. The CNN-Transformer model is proposed, which achieves a well balance between learning both local sentiment features and global sentiment information. Furthermore, we construct a novel Chinese dataset for short Chinese texts sentiment analysis specifically, which contains 80,000 comments from a variety of sources. Extensive experiments and ablation studies on three benchmarks demonstrate that the proposed CNN-Transformer model consistently outperforms the existing methods.