An analytical approach to sentiment extraction and decision-making using fine-tuned large language models

In the competitive restaurant and food service industry, customer reviews play a crucial role in assisting others to make informed decisions. However, the large volume and linguistic complexity of these reviews make it challenging to extract actionable insights effectively. To address this issue, this study presents a LoRA-based fine-tuning workflow for end-to-end multilingual Target–Aspect–Sentiment (TAS) triplet extraction at the paragraph level. First, the SemEval-2016 Restaurant dataset was reformulated from sentence-level to paragraph-level inputs across six languages. Second, open-source large language models (LLMs) were fine-tuned, using a domain-specific instruction template incorporating Chain-of-Thought (CoT) prompting, constrained output format, and temperature optimization. Third, experiments comparing fine-tuned LLMs against non-LLM seq2seq baselines and zero-shot LLM baselines demonstrate that fine-tuning consistently yields the strongest results. The best-performing model, fine-tuned Gemma 2 9B IT, achieves a F1 score of 70.29% across six languages, significantly outperforming the strongest non-LLM baseline (mT5-base, F1 = 58.54%) and zero-shot inference (Gemma 2 9B IT zero-shot, F1 = 45.71%). Finally, based on the best-performing model, a decision support system (DSS) was developed to analyze real-world restaurant reviews from the Yelp dataset, demonstrating the practical applicability of the proposed workflow for both customer and restaurant owners. • Enhance sentiment analysis by extracting insights from multilingual customer reviews. • Improve decision-making by fine-tuning language models for sentiment and aspect detection. • Support restaurant industry decisions through refined sentiment analytics. • Enable targeted insights by capturing sentiment across key business aspects. • Optimize consumer and business strategies through data-driven sentiment modeling.