A fine-grained sentiment recognition method for online Government-Public interaction texts based on large language models

In the era of e-government, fine-grained sentiment analysis of online government-public interaction texts has become increasingly crucial for accurately gauging public opinion and enhancing governmental administrative capabilities. This study proposes a novel fine-grained sentiment recognition method that integrates large language models, grounded in the arousal-valence theory of emotion. The method employs eight emotional intervals to achieve fine-grained, multi-label sentiment representation of online government-public interaction texts. To mitigate potential 'hallucinations' of large models, an innovative rule-based correction scheme was designed, synergizing the advantages of large models while circumventing their limitations. Experimental results on an online government-public interaction dataset demonstrate the superior performance of our proposed method in fine-grained sentiment analysis tasks. This study not only provides new insights and methods for cross-domain, fine-grained sentiment analysis tasks but also has the potential to advance sentiment analysis technologies in natural language processing, demonstrating both theoretical significance and practical value in e-government applications.