Cantonese Sentiment Analysis with LLMs: A Study on OpenRice Restaurant Reviews

The research in this paper is a Cantonese sentiment analysis task based on a large language model, which is of great significance in the field of natural language processing for low-resource languages. This paper constructs a positive and negative emotion dataset based on Cantonese restaurant reviews on the OpenRice platform, and uses three emoticons processing strategies. Zero-shot and fine-tuning experiments are conducted on multiple Cantonese pre-trained models and compared with traditional machine learning and deep learning methods. The results show that fine-tuning significantly improves the performance of the model, and some models have good emotion recognition ability in Cantonese. Emoji processing is less influential and the model mainly relies on semantic information. The fine-tuned models (such as CantoSent-Llama3/3.1 and CantoSent-Qwen-7B series) have been published on Hugging Face, which are helpful to promote Cantonese sentiment analysis and provide technical support for customer sentiment analysis in Cantonese-speaking areas of the catering industry, and help businesses analyze customer sentiment and understand customer feedback.