PolicyVerse Research Project

Partners: College of Engineering and Computer Science (CECS), College of Business and Management (CBM), VinUniversity, Vietnam, UCREL, Lancaster University, UK, and Cardiff University, UK

Government policies are often difficult to interpret because they span multiple ministries, evolve over time, and are published across languages and formats. This project proposes a two-phase programme to develop a multi-agent LLM framework for Collaborative Multilingual Reasoning (CMR), an automated system that detects overlaps, highlights potential inconsistencies, and maps dependencies across policy documents. The system supports analysts and institutions without heavy reliance on legal experts, limiting human involvement to light validation.

Funding: $100,000

Duration: 24 months

Two-Phase Research Structure

The project follows a structured progression from high-resource grounding to low-resource transfer.

Phase 1, High-Resource Development (English and Portuguese)

• Build the multi-agent reasoning engine
• Develop multilingual retrieval and grounding
• Construct Policy World Models (PWMs)
• Evaluate contradiction detection and dependency mapping

Phase 2, Low-Resource Adaptation (Vietnamese and Welsh)

• Integrate authentic Vietnamese and Welsh policy documents
• Extract policy logic, rules, and dependencies using PWMs
• Demonstrate robustness across typologically diverse languages
• Validate cross-lingual generalisation beyond high-resource settings

Innovation: Collaborative Multilingual Reasoning (CMR)

Policy analysis is framed as a structured interaction between domain-specialised LLM agents. Rather than producing a single interpretation, the system surfaces contradictions and hidden dependencies through agent dialogue, generating structured reports for expert review.

• Specialised agent dialogue across policy domains
• Multilingual RAG grounded in real policy documents
• Transparent contradiction detection and audit trails

Policy World Models (PWMs)

PWMs provide structured, language-agnostic representations of policy rules, dependencies, and exceptions. Built through multilingual semantic parsing and rule induction, they enable interpretable policy logic graphs.

• Simulate policy outcomes using encoded rules
• Support context-aware inference beyond retrieval
• Generalise policy logic across languages

Collaborative Expertise

• Mo El-Haj (PI, VinUniversity): Multilingual NLP and low-resource modelling
• Leandro Marcolino (Co-PI, VinUniversity): Portuguese policy data and evaluation
• Paul Rayson (Lancaster): Corpus design and evaluation frameworks
• Dawn Knight (Cardiff): Welsh language modelling and validation
• Nguyen Thi Mai Lan (VinUniversity): Vietnamese policy and regulatory expertise

Expected Deliverables

• Multilingual policy dataset (English, Portuguese, Vietnamese, Welsh)
• Multi-agent LLM framework for policy reasoning
• Structured policy knowledge extraction using PWMs
• GlobalPolicyQA benchmark for multilingual evaluation