Using Large Language Models to Analyze Argument Structures in UN Resolutions: An Innovative Approach from the HYBRIDS Project

Core Idea: The HybridArguer team under the HYBRIDS project has developed a four-stage LLM reasoning workflow that can automatically extract argument structures from bilingual (English and French) UN resolution texts. This solves the problem that traditional methods struggle to capture deep-level rhetorical and argumentative patterns, providing a new technical path for political discourse analysis. This method was validated in the shared task of the 2026 ArgMining Workshop.

In the interdisciplinary field of political science and computational linguistics, understanding the argument structures of international diplomatic texts (such as UN resolutions) is a core challenge. Traditional text analysis methods struggle to capture their deep-level rhetorical and argumentative patterns. The HYBRIDS project is funded by the European Horizon Programme (Grant Agreement No.101073351). The HybridArguer team (including Dr. Siddharth Bhargava from the Italian FBK Research Institute) focuses on the topic of "identifying the stance of argumentative views in political discourse" and validated its results in the shared task of the 2026 ArgMining Workshop.

Dataset: The training set contains 2695 bilingual (English and French) UN resolutions (UN-RES dataset, Gao et al., 2025); the test set includes 45 resolutions from the UNESCO International Conference on Education (1934-2008). The data is stored in JSON format, including document identifiers, recommendation numbers, metadata (structure, paragraph indices), and body paragraphs (annotation types, labels, translations, etc.).

Task Definition: The shared task requires completing three subtasks: 1. Classify paragraphs as "preambulatory" or "operative"; 2. Assign multiple semantic labels to paragraphs; 3. Predict the argumentative relationships and their types between paragraphs.

The HybridArguer team proposed a four-stage modular architecture:

1. Document-level Paragraph Classification: Use a reasoning LLM to classify paragraphs (preambulatory/operative) as a whole, capturing paragraph coherence;
2. Label Candidate Retrieval: Retrieve label candidates based on embedding vector similarity;
3. Source Paragraph Candidate Selection: Select premise candidates based on similarity plus chronological constraints (only preceding paragraphs);
4. Paragraph-level Fine-grained Reasoning: The LLM integrates previous outputs to assign labels to paragraphs and predict the types of argumentative relationships with source paragraphs.

Technical Implementation: Adopt Docker containerized deployment. The experimental environment is a Linux server with 48GB NVIDIA Ampere A40 GPU, CUDA 12.4, and Python 3.11 to ensure reproducibility.

Evaluation Method: Due to the lack of manually annotated ground truth, a heuristic feature-based evaluation method is used, providing a feasible approach for system comparison in the shared task.

Code Structure: Includes data download scripts, four core processing modules (document-level LLM generation, label candidate selection, paragraph candidate selection, paragraph-level LLM generation), and a main execution script. The modular design facilitates reproducibility.

Research Significance: Academically, it demonstrates a new path combining LLM reasoning capabilities with traditional computational argumentation; in application, it supports policy analysts and diplomats to efficiently understand the argumentative patterns of international resolutions. It reflects the European Horizon Programme's emphasis on interdisciplinary (computer science, linguistics, political science) research (the views expressed are those of the authors and do not reflect the position of the European Union).

Future Outlook: Expand to texts such as parliamentary debates, court judgments, and international treaties; use multilingual LLMs to support political discourse analysis in more languages.