Prompteam at UZH Shared Task 2026: RAG-Augmented Classification and Cosine-Filtered Relation Prediction for UN Resolutions

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Prompteam developed a system for the UZH ArgMining 2026 Shared Task, focusing on reconstructing argumentative structure within UN/UNESCO resolutions. The system addresses two primary challenges: classifying paragraph types and assigning thematic tags from a 141-label taxonomy, and predicting directed argumentative relations between paragraphs. Their pipeline integrates a quantised Qwen2.5-7B-Instruct model with retrieval-augmented generation (RAG), utilizing FAISS-indexed dense embeddings for few-shot prompting and pre-filtering tag candidates. For relation prediction, a sliding-window cosine pre-filter was employed to efficiently reduce the quadratic pair space. This parallelisable and fault-tolerant pipeline, featuring atomic checkpointing, successfully processed 2,959 paragraphs across three concurrent Kaggle T4 sessions, overcoming 12-hour GPU limits. The system secured 2nd place overall on the shared task leaderboard.

Key takeaway

For Machine Learning Engineers developing systems for complex document analysis, this work demonstrates effective strategies for large-scale text processing. You should consider integrating quantised LLMs like Qwen2.5-7B-Instruct with RAG for enhanced few-shot classification, especially when dealing with extensive taxonomies. Furthermore, applying a sliding-window cosine pre-filter can significantly reduce computational costs for relation prediction tasks, enabling efficient processing of thousands of paragraphs even under strict GPU time limits.

Key insights

Prompteam's system combines RAG with a quantised LLM and cosine filtering for efficient argumentative structure reconstruction in UN resolutions.

Principles

• Quantised LLMs can be effective for complex NLP tasks.
• RAG enhances few-shot prompting for classification.
• Cosine similarity filtering reduces computational complexity.

Method

The system classifies paragraph types and assigns tags using a quantised Qwen2.5-7B-Instruct model with RAG. Relation prediction uses a sliding-window cosine pre-filter.

In practice

• Use quantised LLMs for resource-constrained environments.
• Implement RAG for improved few-shot classification.
• Apply cosine pre-filtering to reduce quadratic search space.

Topics

• Argumentative Structure
• UN Resolutions
• Retrieval-Augmented Generation
• Large Language Models
• Model Quantization
• Cosine Similarity

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Few-Shot Biomedical Relation Extraction with Large Language Models: A Viable Alternative to Supervised Learning? — LLMs offer a viable few-shot alternative for biomedical relation extraction, outperforming prior few-shot methods, especially for rare relations.
• An Efficient Approach for Answering Not Readily Attainable Questions for RAG-based Applications — This paper proposes an efficient approach for RAG systems to answer questions not readily found in retrieved data.
• LLM-Assisted Reranking to Operationalize Nuanced Objectives in Recommender Systems — LLM-assisted reranking can amplify extreme content, but prompt regularization can mitigate this while maintaining personalization.
• Your UnEmbedding Matrix is Secretly a Feature Lens for Text Embeddings — LLM unembedding matrices encode an "edge spectrum" subspace biasing embeddings towards frequent, uninformative tokens, which EmbedFilter removes.
• How Confident Is the First Token? An Uncertainty-Calibrated Prompt Optimization Framework for Large Language Model Classification and Understanding — Calibrating LLM first-token uncertainty with label priors enables efficient, adaptive prompt optimization for classification.
• LLM-based Detection of Manipulative Political Narratives — A framework combines prompt-based filtering with unsupervised clustering to detect and structure manipulative political narratives.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.