Evaluation of Chinese Sentiment Analysis for Lightweight LLM

Large language models have demonstrated impressive performance in natural language processing tasks. However, their extensive parameter scale necessitates substantial computational resources and presents various challenges regarding portability and application scenarios, thereby hindering the widespread adoption and utilization of this technology. This study examines domestic large language models characterized by high portability and a smaller parameter scale, particularly focusing on their performance in sentiment analysis tasks. Accordingly, we designed five Chinese sentiment analysis tasks based on seven public datasets, evaluated the tasks using popular lightweight domestic large language models, and compared their capabilities with deep learning models and ChatGPT. The results indicate that the performance of lightweight domestic large language models on Chinese sentiment analysis tasks surpasses that of deep learning models and approaches the performance of ChatGPT. Furthermore, we assessed enhancement techniques such as prompt word engineering and large model fine-tuning, revealing that the enhanced model's parameter count is merely 3.45% of ChatGPT's, while achieving 95.2% of ChatGPT's performance.