Fine-tuning llama-2 and few-shot learning for ABSA

The Llama-2 model has demonstrated excellent performance in various research fields and can be adapted to specific tasks through fine-tuning techniques. This paper focuses on exploring the application of Llama-2 in aspect-based sentiment analysis (ABSA), specifically focusing on the joint tasks of aspect term extraction and polarity classification. We propose a few-shot ABSA method based on fine-tuning Llama-2. We discuss the impact of simple and complex training data instructions on model performance and find that their influence is minimal. Additionally, we investigate the performance of the fine-tuned model when using different numbers of context prompts during inference. We find that the fine-tuned Llama-2, combined with few-shot context prompts, performs well and can consistently output JSON format, achieving a maximum F1 score of 69.5%, which is a 3.8% improvement compared to GPT-3.5. Results analysis indicates that finetuning helps reduce false positives, improve model sensitivity, and specificity.