📄 TypeCoT at UZH Shared Task 2026: Reconstructing Argumentative Structure in UN Resolutions using Type-Informed Chain-of-Thought

Summary

TypeCoT is a submission to the UZH Shared Task at ArgMining 2026, which targets the implicit argumentative structure of UN and UNESCO resolutions. Participants could only use open-weight models with at most 8B parameters, and the whole pipeline runs on Qwen-2.5-7B-Instruct.

For Subtask 1, a 4-bit LoRA adapter trained with Unsloth labels each paragraph as preambular or operative. The 141 multi-label tags are assigned zero-shot, prompting one of 15 dimensions at a time to reduce hallucination. For Subtask 2, the predicted paragraph types act as explicit constraints in a two-stage Chain-of-Thought. The model first summarizes paragraphs, then proposes backward links and relation labels while obeying type rules derived from development data. A multi-pass recovery procedure handles documents that exceed the 32,768-token context limit.

The system was trained on UN Human Rights Council resolutions and tested on historical UNESCO ICE recommendations. It covered all 2,959 test paragraphs and placed 8th overall. The work shows what a small open model can do on legal discourse and also exposes its weaknesses: overly dense link graphs and missed rare relation types.

Key points

• Data: the training material consisted of 2,695 resolutions drawn from the UN-RES corpus (Gao et al., 2025), while evaluation was carried out on 45 UNESCO ICE documents (1934–2008), a deliberate shift in both domain and time period.
• Subtask 1b tagging: 141 codes in 15 dimensions are handled with separate per-dimension prompts, each with an "NA" option. This took 44,385 inference calls in total, and 2,949 of 2,959 paragraphs received tags.
• Type rules for Subtask 2: development data shows that operative→operative links dominate (61.3%) and preambular→operative links never occur. All true links point backward. The four relation labels are contradictive, supporting, complemental and modifying.
• Context recovery: long documents (over 45 paragraphs) are handled in three stages. Regex salvage of truncated outputs recovers about 85% of paragraphs, and the rest are retried in chunks of 15 paragraphs with condensed one-line summaries.
• Results and resources: the system ranked 6th on the F1 leaderboard, 8th on the LLM-Judge leaderboard and 8th overall; the paper reports only ranks, not absolute scores. The authors note that the constraints suppress modifying and contradictive relations, and that no ablation against an unconstrained baseline was run. Code is on GitHub (ChandanKumar683/ArgMining), and outputs and reasoning traces are in a Hugging Face collection (Chandan683/argmining).