Research on Customer Emotion Recognition on the Side of State Grid Power Customer Service Agents based on Large Language Models

Abstract

In the service scenarios on the side of power customer service agents, it is difficult for agents to accurately capture and respond to changes in customer emotions, resulting in poor service experience. Fluctuations in customer emotions directly affect their communication methods and expression of needs. To address this problem, this paper proposes a three-stage emotion recognition and response scheme of "small model pre-screening - large model emotion grading - recommended soothing scripts". Firstly, a domain-specific sensitive lexicon for power customer service is constructed, and a lightweight small model is adopted to achieve rapid sensitive word recognition. If a sensitive word is matched, manual agent intervention is triggered and the call is transferred to a manual agent. Secondly, based on the State Grid Guangming Large Model, the model is fine-tuned to build an emotion recognition model, which judges the emotional polarity (positive, neutral, negative) of the dialogue text after the manual agent is connected, and subdivides negative emotions into three levels: mild, moderate, and severe. Finally, according to the recognition results of different levels of negative emotions, soothing scripts are recommended to assist agents in making accurate responses. This can effectively improve the efficiency of agents' emotion perception and response speed, enhance customer experience, and provide corresponding technical support for the intelligent upgrading of power customer service.