Modeling Argument Structure in Bilingual UN Resolutions via Retrieval-Augmented and Self-Consistent Reasoning LLMs

Team: HybridArguer at UZH Shared Task @ ArgMining Workshop 2026 : Reconstructing the Reasoning in United Nations Resolutions

This repository contains documentation on a project that forms part of a broader PhD research effort titled "Identifying the Stance of Argumentative Opinions in Political Discourse", conducted under the HYBRIDS Project within the Horizon Europe framework.

The primary contributor and point of contact for this repository is Siddharth Bhargava (sbhargava@fbk.eu).

---

Overview

The repository contains the scripts to extract argument structures in form of pre-defined JSON scripts from bilingual (French and English) UN resolutions documents.

---

Data

The dataset used in this work is derived from the Shared Task on "Reconstructing the Reasoning in United Nations Resolutions", organized by the University of Zurich (UZH) as part of the 13th Workshop on Argument Mining (2026). The training set comprises 2,695 bilingual (original French with English translations pre-generated by a LLM) documents, while the held-out test set includes approximately 45 resolutions distributed across 90 JSON files with a fixed schema.

The data can be downloaded from the provided Shared Task repository using the following command:

$ python src/download_data.py

Remember to add in your HuggingFace Access Token (with correct priveleges) in the file above for faster data download. This is also a gated repository requiring special permission for access from Zurich NLP team. For further details further to README in Data folder.

Dataset Description

Languages

• English
• French

Granularity

• Paragraph level

Splits

Training Set (parsed_data_en)

• Source: UN-RES dataset Gao et al., 2025
• Size: 2,695 UN resolutions
• Language: French (with machine-generated English translations)
• Annotation: paragraph-level argumentative structure

Test Set (parsed_data_fr)

• Source: UNESCO International Conference on Education (1934–2008)
• Size: 45 parsed documents (each may contain up to three resolutions in JSON)
• Language: French
• Annotation: paragraph-level (held out for evaluation)
• Validation set: none

Tags

• See Data/education_dimensions_updated.csv

---

Data Format

All data are provided in JSON format following a fixed schema.

Example (simplified)

"TEXT_ID": "ICPE-25-1962_RES1-FR_res_54",
  "RECOMMENDATION": 54,
  "TITLE": "LA PLANIFICATION DE L'ÉDUCATION",
  "METADATA": {
    "structure": {
      "doc_title": "ICPE-25-1962_RES1-FR",
      "nb_paras": 58,
      "preambular_para": [], 
      "operative_para": []
      "think": ""      
    }
  },
  "body": {
    "paras": [
      {
        "para_number": 1,
        "para": "La Conférence internationale de l'instruction publique, Convoquée à...",
        "type": null,      
        "tags": [],     
        "matched_paras": [],     
        "think": "",
        "para_en": "The International Conference on Education, convened in ..."
      },
      ...
    ]
  }
}

---

Methodology

Our approach comprises four stages:

1. A reasoning LLM that classifies all paragraphs in a given document collectively as either preambular or operative (see src/doc_llm_generation.py)

2. Embedding-based similarity retrieves tag candidates for each paragraph (see src/tag_candidate_selection.py)

3. Similar to stage 2, candidate source paragraphs are selected under a chronological constraint (see src/para_candidate_selection.py)

4. The reasoning LLM processes each target paragraph individually to assign it tags from its candidate pool and predict one or more predefined relation types with its candidate source paragraphs (see src/para_llm_generation.py)

Outputs from the first and final stages are consolidated into the final structured representation.

---

Experimental Setup

Step 1. Build the Docker Container using the following command

$ docker build --rm -t YOUR_CONTAINER_NAME . 

Thia builds a Docker container containing your project environment. All experiments have been executed on a Linux server with a 48GB NVIDIA Ampere A40 GPU, CUDA version: 12.4 and Python version: 3.11.

For implementation, we define a shell file exe.sh that runs the main.py script.

python main.py

Remember to change mode of the shell file: chmod +x exe.sh.

The command is run in the docker environment as follows:

$ docker run   --gpus='"device=DEVICE_NUMBER"' --runtime=nvidia --rm -ti --shm-size=32gb -v $PWD:/app YOUR_CONTAINER_NAME ./exe.sh 

---

Evaluation

There are three main tasks that are evaluated for each paragraph across all test documents: (1) Paragraph Classification, (2) Multi-label Tag Assignment, and (3) Multi-label Relation Prediction and Classification (the latter can be sub-divided into relation prediction and relation type multi-label classification).

Due to lack of ground truth, we evaluate the performances of each task using LLM-as-a-judge with manual self-evaluation for error analysis.

---

Acknowledgements

This research work has received funding from the European Union's Horizon Europe research and innovation programme under the Marie Skłodowska-Curie Grant Agreement No. 101073351. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them.

Additional Comment:

• Use of any AI coding tool or service has been limited to preliminary skeletal programming and debugging support.

---

Citation