Evolutionary Expert Model Merging with Task-Adaptive Iterative Self-Improvement Process for Large Language Modeling on Aspect-Based Sentiment Analysis

In recent years, natural language processing (NLP) has achieved notable advancements, particularly with large language models (LLMs) enhancing tasks such as sentiment analysis. Aspect-based sentiment analysis (ABSA), which involves identifying sentiment related to specific aspects within text, poses a more complex challenge compared to simple sentiment classification. To address this complexity, integrating multiple expert models has emerged as a promising approach. In this study, we propose the Evolutionary Expert Model merging with Task-adaptive Iterative Self-improvement Process (EEM-TISP) to improve ABSA task performance. EEM leverages evolutionary algorithms to merge expert models, optimizing task-specific outputs, while TISP iteratively refines these outputs, enhancing accuracy. Our results demonstrate that EEM significantly improves performance in simpler tasks, such as Aspect Term Sentiment Analysis (ATSA) and Aspect Category Sentiment Analysis (ACSA), surpassing individual expert models. TISP further boosts accuracy in these simpler tasks, though performance declines were observed in more complex ones. These findings indicate that while TISP is effective for simpler tasks, further optimization is required to tackle complex ABSA challenges, calling for future research into flexible prompt design and task-specific approaches.