A modular pipeline for converting natural language OTC derivatives into Common-Domain-Model (CDM) representations using LLM and RAG.

IJCAI'25 Paper : https://arxiv.org/abs/2506.01063 AI4F Symposium'25 Paper : https://arxiv.org/abs/2510.23990

This repository enables an end-to-end automation flow for OTC derivatives processing:

• Natural Language contract description → CDM representation

It also supports synthetic data generation and contract template creation.

✅ Recommended: Use Conda with Python 3.11.10

# Step 1: Clone the repository
git clone https://github.com/smart-derivative-contracts/automating-otc-derivatives.git
cd automating-otc-derivatives

# Step 2: Create and activate conda environment
conda create -n <your_env_name> python=3.11.10
conda activate <your_env_name>

# Step 3: Install dependencies
# If any package fails, remove it from requirements.txt and try again
pip install -r requirements.txt

• src/natural_language_to_cdm/
  Contains experiments and models for converting contract descriptions to CDM format

• generate_text_from_cdm.py
  Generates synthetic natural language descriptions from CDM data

• create_contract_templates.py
  Generates six contract templates based on the CDM schema

The figure below illustrates the overall pipeline (CDMizer) used for populating CDM templates using RAG-enhanced prompting and LLM inference:

CDMizer Workflow: Recursive traversal, governed by a depth threshold (d), selects substructures (e.g., assignedIdentifier) where the deepest subtree has a depth ≤ d. This ensures manageable task sizes for efficiency and accuracy. Context-aware prompts, incorporating object definitions, traversal paths, and RAG-retrieved examples, guide the LLM in populating fields, which are then validated. Recursive traversal ensures the entire structure is systematically completed.

• Python version must be 3.11.10 for compatibility.
• If certain packages in requirements.txt fail to install, comment them out and retry.
• The repository is under active development.