Aspect-Based Sentiment Analysis Model for Short Text with Interactive Dependent Multi-head Graph Attention

This article proposes an Interactive Dependent Multi-head Graph Attention (IDM-GAT) model for short text sentiment analysis based on dependent multi-head graph attention, which addresses the shortcomings of existing methods for short text sentiment analysis , such as non-compliance with conventional grammar and insufficient utilization of deep emotional expression between contextual words and aspect words. The model enhances the embedding of sentiment information at the sentence level by adding relative position vectors. Then, it uses the dependent graph attention structure to calculate the attention weights of individual words and dependencies between words in the sentence, and extracts sparse dependency relationships in short text as much as possible. It uses self-attention pooling and average pooling methods for contextual and aspect words at the word level, reducing the computational complexity of the model while preserving emotional features. The combination of two levels provides a more comprehensive modeling of short texts from both global and local perspectives. Through simulation experiments, the proposed model achieved accuracy of 83. 85% , 79. 15% , and 74. 39% on three benchmark datasets: Semeval2014 hotel reviews, computer reviews, and ACL14 Twitter, respectively.