ArgMining 2025 Submission – Team "gottfried-wilhelm-leibniz"

Critical Question Generation (CQs-Gen)

This repository contains the complete pipeline developed by Team gottfried-wilhelm-leibniz for the Critical Question Generation (CQs-Gen) Shared Task at ArgMining 2025.

Task Overview

The goal of the task is to generate critical questions that challenge the reasoning, assumptions, or conclusions in argumentative interventions. These questions are essential for fostering critical thinking and deepening understanding in argumentative discourse.

Our Approach

We implemented a fully automated two-stage pipeline: first, prompting open-weight LLMs to generate candidate critical questions, and second, reranking them using a fine-tuned ModernBERT classifier to select the most useful ones. This combined approach consistently outperformed prompting-only baselines and achieved a test score of 0.569, ranking 5th overall in the ArgMining 2025 shared task.

Repository Contents

src/              # Core Python code for CQ generation and reranking
prompts/          # Prompt templates used for instructing LLMs
generated_cqs/    # Raw critical questions generated by each LLM
reranked_cqs/     # Top-3 useful CQs per intervention after reranking with ModernBERT
dataset/          # Sample, validation, and test sets provided by the shared task 
submission/       # Final prediction files submitted to the ArgMining 2025 shared task
README.md         # This readme file summarizing our submission and approach