Recent Publication from team

We all know that construction contracts have nuanced semantics and a very complex nested logic. This is where traditional rule-based and ML/NLP methods fall short. To date, their review remains time-consuming and error-prone.

In our new paper, we try to address this significant challenge. We present and test a technical solution that includes
(1)  a Nested Contract Knowledge Graph (NCKG) for fine-grained, clause-level risk modelling,
(2)  a Graph Retrieval-Augmented Generation (GraphRAG) to enrich LLM reasoning with structured knowledge, and
(3)  a semi-automated pipeline that links contract clauses to risks like payment terms, liabilities, and ambiguities.
Two state-of-the-art LLMs were tested:
a) GPT-4 (OpenAI), and
b) LLaMA-3 (Meta).
We compared the results from testing:
- LLM-only (No external context; direct prompting on raw contract clause),
- Vector database-enhanced (Contract clauses retrieved from a vector database using cosine similarity (i.e., classic RAG based on semantic similarity), and
- NCKG-enhanced (our approach): Retrieval of relevant entities, events, constraints from a Nested Contract Knowledge Graph.

143 clauses from standard forms (FIDIC Red and Silver Books, NEC) from international construction projects.

335 total triples of which 179 were nested triples, modelling complex logic (e.g., time constraints, conditional clauses) and covering Covered six risk categories: Assignment, Payment, Temporal, Financial, DSC (Differing Site Conditions), Liability.

The experiments demonstrated that our approach (NCKG + GraphRAG) significantly outperformed both standard prompting and traditional RAG techniques for automated construction contract risk review.

hashtag#Kudos to our visiting PhD Student Chunmo Zheng from Zhejiang University who led this work with contribution from all collaborators: Saika Wong, Xing Su, Yinqiu (Rachel) T., Ahsan Nawaz (Ph.D).

As a background: this work was shortlisted as one of the top papers in the European Council on Computing in Construction's EC3 2024 conference. Despite this, it took four extensive rounds of reviews to get it published requiring the inclusion of further techs, datasets, functionalities and experimental testing to reach this format. Thanks to all reviewers and the editorial team for their feedback and guidance.

50-day free access to the paper via: https://lnkd.in/ew8PqVca

hashtag#ConstructionContracts hashtag#DigitalConstruction hashtag#ConstructionTech
hashtag#ContractManagement hashtag#RiskManagement hashtag#LargeLanguageModels hashtag#LLM
hashtag#KnowledgeGraphs hashtag#GraphRAG hashtag#AutomationInConstruction hashtag#ConstructionLaw hashtag#LegalTech hashtag#SemanticAI Newcastle University

https://www.sciencedirect.com/science/article/pii/S0926580525002195?dgcid=coauthor

https://www.linkedin.com/feed/update/urn:li:activity:7317474055035506690/

Last modified: Mon, 14 Apr 2025 13:50:16 BST