Textual Aspect Sentiment Analysis Method Based on Llama-3 Model Optimization

Abstract

This paper proposes a textual aspect sentiment analysis method based on Llama-3 model optimization to address challenges in fine-grained sentiment classification. The approach integrates two key innovations: (1) Prompt engineering to reconstruct datasets into structured "Aspect-Polarity" pairs, enhancing data quality and reducing noise; (2) LoRA-based fine-tuning, which introduces low-rank adapters to efficiently update model weights while minimizing computational costs. Experiments on the Semeval-2016 dataset demonstrate superior performance, with the optimized Llama-3 model achieving 92.29% accuracy and 90.48% F1-score, outperforming both open-source and closed-source models. Ablation studies confirm the synergistic benefits of prompt engineering and LoRA fine-tuning. The method balances efficiency and accuracy, offering a practical solution for aspect sentiment analysis tasks.