Generalized Category Discovery in Event-Centric Contexts: Latent Pattern Mining with LLMs
Yi Luo, Qiwen Wang, Junqi Yang, Luyao Tang, Zhenghao Lin, Zhenzhe Ying, Weiqiang Wang, Chen Lin
TL;DR
This work defines Event-Centric Generalized Category Discovery (EC-GCD) as a realistic GCD setting with long, narrative texts and severe class imbalance. It introduces PaMA, a framework that uses LLM-driven event-pattern mining to align cluster prototypes with known classes, complemented by a ranking-filtering-mining pipeline to mitigate minority-class bias. PaMA demonstrates state-of-the-art performance on EC-GCD benchmarks, including a newly constructed Scam Report dataset, and maintains strong generalization on standard GCD benchmarks. The approach combines instance- and prototype-level contrastive learning with a robust pseudo-label reassignment mechanism, offering a practical solution for open-world text classification tasks with complex narratives.
Abstract
Generalized Category Discovery (GCD) aims to classify both known and novel categories using partially labeled data that contains only known classes. Despite achieving strong performance on existing benchmarks, current textual GCD methods lack sufficient validation in realistic settings. We introduce Event-Centric GCD (EC-GCD), characterized by long, complex narratives and highly imbalanced class distributions, posing two main challenges: (1) divergent clustering versus classification groupings caused by subjective criteria, and (2) Unfair alignment for minority classes. To tackle these, we propose PaMA, a framework leveraging LLMs to extract and refine event patterns for improved cluster-class alignment. Additionally, a ranking-filtering-mining pipeline ensures balanced representation of prototypes across imbalanced categories. Evaluations on two EC-GCD benchmarks, including a newly constructed Scam Report dataset, demonstrate that PaMA outperforms prior methods with up to 12.58% H-score gains, while maintaining strong generalization on base GCD datasets.