Multi-channel enhanced graph convolutional network for sentiment analysis on instrumental music descriptions

Traditional single-channel feature extraction methods face challenges in instrumental music sentiment analysis, primarily due to their reliance on a single type of dependency, which overlooks the complex relationships between musical elements and emotions. While graph convolutional network (GCN)-based approaches show potential, they still struggle with aggregating both musical structure information and emotional details, especially in instrumental music without lyrics, where misinterpretation of emotional features is common. Moreover, insufficient domain knowledge hinders the model’s ability to capture subtle differences in musical terminology, further reducing sentiment analysis accuracy. To address these challenges, we propose a sentiment analysis graph neural network called KSD-GCN, where the sentiment-enhanced syntactic graph convolution module enriches the dependency graph by integrating external sentiment knowledge, thereby improving the model’s ability to capture emotions. The dependency relation embedding module focuses on capturing syntactic dependency information within the sentence. Additionally, we introduce a multi-layer interactive attention mechanism that effectively integrates syntactic, dependency, and semantic information. Through this interaction, the model can finely capture the sentiment and syntactic structure of the sentence at different layers, significantly improving the accuracy of aspect-based sentiment analysis. Experimental results show that, on multiple datasets, the model outperforms baseline models across several performance metrics.