Researchers propose CLUE, a contrastive learning framework for unsupervised event type induction and annotation

Researchers from an unnamed institution propose CLUE, a contrastive learning framework designed to automate event type induction, naming, and annotation without requiring manual ontology creation. The work addresses the resource-intensive nature of traditional event ontology development, which relies on domain experts to define guidelines and event types, often introducing human biases.

The framework introduces a novel joint embedding representation that jointly expresses both verb-centric and semantic views of events. This dual-view representation enables more comprehensive event characterization than prior methods that typically focus on a single perspective.

Event type induction in CLUE is performed using contrastive loss guided clustering, a technique that leverages contrastive learning to group similar event representations while separating dissimilar ones. The authors report that experiments demonstrate the effectiveness of this approach in unsupervised settings.

Beyond induction, CLUE includes mechanisms for event naming and refinement, aiming to produce interpretable and consistent event type labels. The authors provide an implementation of the framework, including code and resources, via a public GitHub repository.

The paper is published on arXiv under the Computation and Language category with identifier arXiv:2610.12580 and was submitted on October 8, 2026. The authors—Pratik Maitra, Lauren Kwon, and Qi Li—describe the contributions as a step toward reducing reliance on expert-driven ontology creation while maintaining or improving performance in event extraction tasks.