DiCoRe: Divergent-Convergent Reasoning for Zero-shot Event Detection

This repository contains the implementation of DiCoRe [paper] (Divergent-Convergent Reasoning), a novel approach for zero-shot event detection that leverages a two-stage reasoning process to improve event extraction performance.

Overview

DiCoRe comprises three components:

1. Dreamer: This component conducts divergent open reasoning to find all possible solutions and emphasizes on improving the recall.
2. Grounder: This component grounds and maps the open predictions from the dreamer to the task constraints.
3. Judge: This component is a simple verification that ensures that the predictions are highly precise.

This approach enables effective zero-shot event detection across multiple domains without requiring task-specific training data. Below we show an illustration.

Folder Structure

• /code/: Contains the core implementation files:

• /scripts/: Contains shell scripts for running experiments:

• /data/: Contains datasets and processed data for evaluation:

⚠️ Data Disclaimer: The data provided in this repository is for research purposes related to our work. If you use any of these datasets, please:

• Acknowledge the original license/policy of the respective datasets
• Cite the original dataset papers appropriately
• Follow any usage restrictions specified by the original dataset creators

Quick Start

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (recommended)
• VLLM for efficient LLM inference

Installation

# Create conda environment from environment.yml
conda env create -f environment.yml

# Activate the environment
conda activate dicore

Running Experiments

1. DiCoRe

# Run DiCoRe on ACE05 dataset with Llama-3 model
./scripts/run_dicore.sh ace llama3_8b 2 1 0.4 "" 8

2. Base DiCoRe (w/o FSM)

# Run baseline model on ACE05 dataset
./scripts/run_dicore_base.sh ace llama3_8b 2 1 0.4 "" 8

3. DiCoRe w/ Chain-of-thought

# Run DiCoRe with Chain-of-Thought reasoning
./scripts/run_dicore_cot.sh ace llama3_8b 2 1 0.4 "" 8

Script Parameters

• dataset: Dataset name (ace, maven, casie, genia, fewevent, speed)
• llm: Language model (llama3_8b, gpt4, etc.)
• n_gpu: Number of GPUs to use (default: 2)
• n_runs: Number of runs (default: 1)
• temp: Temperature for generation (default: 0.6)
• suffix: Output suffix (default: "")
• batch_size: Batch size for inference (default: 8)

Explicit Evaluation

We utilize TextEE for the evaluation setup. We extract a single script and provide it here for faster evaluation.

python code/evaluate_llm_preds.py --pred_file <pred-file> --gold_file <gold-file>

Citation

If you use this code or find it helpful, please cite our paper:

@inproceedings{parekh2025dicore,
    title={DiCoRe: Enhancing Zero-shot Event Detection via Divergent-Convergent LLM Reasoning},
    author={Tanmay Parekh and Kartik Mehta and Ninareh Mehrabi and Kai-Wei Chang and Nanyun Peng},
    booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
    year={2025}
}

Contact

For questions or issues, please contact the lead author Tanmay at tparekh@g.ucla.edu

Acknowledgments

We thank the creators of the datasets used in this work and the open-source community for the tools and frameworks that made this research possible.