Integrating Large Language Models and Graph Neural Networks for enhanced Arabic sentiment analysis

Abstract Morphological complexity and dialectal variation make Arabic one of the most demanding languages for automated sentiment analysis (SA). To tackle these limitations, this work proposes a hybrid framework that couples a Large Language Model with a Graph Neural Network (LLM–GNN), combining the complementary strengths of contextual text encoding and relational structure learning. Specifically, AraBERT v2 serves as the backbone encoder, while a Graph Convolutional Network (GCN) built on cosine-similarity edges captures inter-sentence dependencies that purely sequential architectures tend to overlook. The framework is benchmarked on the publicly available Arabic 100k Reviews dataset (99,999 authentic user-generated reviews balanced equally across Positive, Negative, and Mixed sentiment classes). Against four established baselines (fine-tuned AraBERT, AraBERT-BiLSTM, AraBERT-MLP, multilingual BERT and modern Arabic-centric LLMs such as Jais-13B), the proposed model achieves an overall accuracy of 66.9% and a macro F1-score of 66.55%, representing gains of 7.6% and 4.4% over the strongest comparable baseline, respectively. Training curves indicate stable loss reduction from the earliest epochs, reflecting consistent optimization behavior throughout the 50-epoch schedule. A noted constraint is that graph construction operates at the mini-batch level, which limits the model’s exposure to corpus-wide semantic relationships. Nevertheless, the results confirm that integrating graph-based relational reasoning with transformer-derived embeddings surfaces nuanced sentiment signals that sequential models routinely miss, pointing to practical utility in Arabic social media monitoring and large-scale customer review analysis.