Reproduce “Keyword-Centric Prompting for One-Shot Event Detection with Self-Generated Rationale Enhancements” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ziheng Li, Zhi-Hong Deng

Peking University researchers present KeyCP++, a framework for one-shot event detection using large language models, which integrates keyword-centric prompting with automatically generated rationales. This approach improves trigger understanding and reduces over-interpretation, outperforming prior in-context learning and supervised fine-tuning methods by 7.1% to 12.8% in F1 score.

Key insights

  • Integrating automatically generated keywords into LLM prompts effectively mitigates over-interpretation and improves trigger profiling for one-shot event detection.
  • Self-generated rationales, particularly those derived from comparative judgments of both keywords and LLM-proposed candidates, significantly enhance LLMs' ability to discriminate and accurately identify event triggers.
  • Prioritized negative sampling, which focuses on examples with more potential trigger candidates, is crucial for teaching LLMs robust detection rules and preventing over-conservative predictions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.