Fast Fine-Tuning Large Language Models for Aspect-Based Sentiment Analysis

Abstract

The method proposed in this study aims to reduce the execution time required for fine-tuning large language models in aspect-based sentiment analysis. To achieve efficient fine-tuning, the large-language model parameter tuning for new data is accelerated through rank decomposition. Experiments on the SemEval datasets demonstrated that our method consistently outperformed strong baselines such as GPT-ABSA and BART-ABSA across multiple metrics including accuracy, F1-score, precision, and recall while also reducing fine-tuning time by approximately 35%. The experimental results demonstrate a notable decrease in execution time with the proposed approach of the fine-tuning process while preserving the accuracy of polarity prediction.