BERT-OntoSent: combining BERT language model with sentiment ontology for enhanced sentiment analysis on social media

Sentiment analysis on social media is vital but challenged by language complexity and context dependency. Existing methods often fall short. This paper presents BERT-OntoSent, a hybrid approach synergizing BERT's contextual power with ontology-based structured knowledge. We provide a detailed methodology where the DCWEB-SOBA process transforms text into (aspect, opinion, sentiment) triplets, populating a domain-specific BERT-OntoSent ontology used for fine-tuning BERT. SWRL rules within the ontology then refine BERT's predictions, enhancing robustness, particularly for negation and mixed sentiments, as formalized in our proposed refinement algorithm. We present a comprehensive evaluation demonstrating BERT-OntoSent's effectiveness not only on the SemEval 2016 restaurant benchmark but also across diverse domains, including product reviews (Amazon) and microblogs (Twitter). Our results consistently show significant improvements in accuracy and F1-score compared to BERT-only and ontology-only baselines across these datasets. Furthermore, we report on experimental optimizations for scalability, demonstrating practical performance gains. The findings confirm the robustness and broader applicability of the BERT-OntoSent hybrid approach, establishing the value of LLM-ontology synergy for advanced sentiment analysis.