Enabling Explainable NLP Through Graph Neural Network Architectures

The growing need for transparency in Natural Language Processing (NLP) exposes the interpretability limitations of Transformer-based models, where token-level attention mechanisms often obscure deeper semantic understanding. This paper introduces a hybrid framework that converts raw text into structured graph representations for relational reasoning via Graph Neural Networks (GNNs), while also employing Large Language Models (LLMs) to generate fluent, human-aligned explanations. Our approach is evaluated on benchmark datasets including AG-News and SST-2, demonstrating superior performance in terms of fidelity, sparsity, and interpretability when compared to state-of-the-art methods such as GNNExplainer, SubgraphX, and PGM-Explainer. Furthermore, we propose future directions involving heterogeneous graphs and the integration of external knowledge sources. This work establishes a new standard for trustworthy and explainable NLP by synergizing symbolic structure with semantic fluency.

Paper

Full text

PDF

Enabling Explainable NLP Through Graph Neural Network Architectures

Semantic Scholar · 2025

Abstract

The growing need for transparency in Natural Language Processing (NLP) exposes the interpretability limitations of Transformer-based models, where token-level attention mechanisms often obscure deeper semantic understanding. This paper introduces a hybrid framework that converts raw text into structured graph representations for relational reasoning via Graph Neural Networks (GNNs), while also employing Large Language Models (LLMs) to generate fluent, human-aligned explanations. Our approach is evaluated on benchmark datasets including AG-News and SST-2, demonstrating superior performance in terms of fidelity, sparsity, and interpretability when compared to state-of-the-art methods such as GNNExplainer, SubgraphX, and PGM-Explainer. Furthermore, we propose future directions involving heterogeneous graphs and the integration of external knowledge sources. This work establishes a new standard for trustworthy and explainable NLP by synergizing symbolic structure with semantic fluency.

Similar papers

© 2026 NYSGPT2525 LLC