A new framework, CrisisKD, was announced for aspect-level sentiment and emotion analysis in crisis situations. The framework utilizes a teacher LLM to generate aspect-level labels and reasoning traces, which supervise a smaller student model. The process involves aspect extraction, syntactic parsing, opinion extraction, sentiment classification, and emotion classification. A dataset of 50,615 aspect-level labels was created and released as open-source resources. The student model, based on Qwen2.5-7B, demonstrates improved performance on several tasks.
On a 500-tweet gold set, the student model achieved a 7.9 F1 point improvement on aspect extraction, a 17.0 point improvement on emotion accuracy, and a 6.5 point improvement on sentiment accuracy compared to the untuned model. Furthermore, the CrisisKD system improved the Qwen2.5-7B ICL baseline on the ABEA benchmark by 2.8 F1 points on ATE and 3.8 F1 points on joint ATE+AEC.
This framework offers a method for reducing inference costs while maintaining accuracy in analyzing complex textual data. The open-source nature of the dataset and scripts allows for further research and development in this area. The system’s performance suggests potential applications in understanding public concerns during crisis events.
Source: https://arxiv.org/abs/2609.05757