TFMPHGNN: Two-Fold multi-perspective heterogeneous graph neural network for sentiment analysis

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Sentiment analysis faces challenges due to complex relationships between sentiment, context, and emotions across diverse data.
• Existing deep learning and transformer models often isolate sentiment expressions, missing multi-perspective interactions.

Purpose of the Study:

• To introduce a novel Two-Fold Multi-Perspective Heterogeneous Graph Neural Network (TFMPHGNN) for comprehensive sentiment analysis.
• To effectively model joint dependencies among sentiment, emotion, and context using a dual-stage heterogeneous graph framework.

Main Methods:

• A meta-path-based encoder with a capsule network captures hierarchical semantic relationships.
• A multi-channel graph convolutional network (MC-GCN) learns complementary representations.
• A variational autoencoder (VAE) refines latent embeddings for denoising.

Main Results:

• TFMPHGNN achieved superior performance over eight state-of-the-art baselines on the VaKSent-2025 corpus.
• Significant improvements were observed in accuracy (4.67%), F1-micro (2.7%), and F1-weighted (4.2%).
• Ablation studies confirmed the synergistic value of integrating multiple graph perspectives, with collaborative fusion yielding the highest accuracy.

Conclusions:

• TFMPHGNN effectively captures intricate sentiment-emotion interdependencies.
• The proposed model offers a robust and interpretable framework for advanced sentiment understanding.
• This research advances the field by providing a more nuanced approach to analyzing sentiment in complex data.