Zefan Cai

Zefan Cai

PhD student at UW-Madison. Test-time training, efficient architecture, long-context memory.

Improving Event Definition Following For Zero-Shot Event Detection

← All research

Language Understanding

ACL 2024 (Main) — Long Paper · PDF · Paper · Code · DivED

Teach event detection to follow a definition of an unseen event. ZeroED studies how training data should be constructed for zero-shot event detection. Its Diverse Event Definition dataset, DivED, provides more than 3,000 event types, with varied definitions and examples generated using event ontology information.

DivED generation pipeline with event retrieval, ontology-aware definitions and examples, and definition expansion.
Figure 2: DivED uses an event ontology to create distinct definitions and samples, expands the definitions, and prunes highly overlapping event types. Click to enlarge.

The study separately varies event types, definitions and examples. In the reported scaling experiments, diverse event types and definitions improve generalization, while simply adding more examples per type eventually stops helping. Ontology-aware definitions and hard negatives help the model distinguish neighboring event types and their triggers.

A LLaMA-2-7B model trained with DivED and Geneva is evaluated on unseen event types in ACE, M2E2 and MEE. It reaches an average trigger-classification F1 of 25.7, compared with 22.6 for the Geneva-only model and 9.0 for the paper’s GPT-3.5 baseline. The results concern these zero-shot benchmarks and prompt settings; the scaling study also shows that excessive generated data can hurt out-of-domain transfer.