AI-DRIVEN STRUCTURING AND SEMANTIC MATCHING OF CONSTRUCTION COST DATA FOR EFFICIENCY AND CO₂ IMPACT ASSESSMENT
The buildings and construction sector is one of the largest contributors to global greenhouse gas emissions. It accounts for around 34% of global energy-related CO₂ emissions, which come from both the energy used to operate buildings and the emissions produced when construction materials such as cement, steel and concrete are made. Despite the increasing use of embodied-carbon assessment methodologies in the architecture, engineering and construction (AEC) sector, carbon accounting tools are still not integrated into the cost estimation and procurement processes that directly affect material selection and initial design choices. This study presents the novel, AI-driven 'EEBOQ' framework, which was developed in the context of the Latvian construction market while taking into account broader European and international carbon accounting practices. The proposed framework integrates a hybrid semantic processing pipeline that combines Large Language Models (LLMs), ontology-based classification mechanisms and vector-based similarity retrieval techniques in order to autonomously interpret heterogeneous, spreadsheet-based procurement documentation. The system aligns the free-text estimate positions with standardised embodied-carbon reference datasets, such as the ICE Database, Environmental Product Declarations (EPDs) and EN 15978-compliant life cycle assessment structures. To improve the reliability of practical CO₂ estimation, the framework introduces a Bayesian Feedback Correction Engine (BFCE). This is designed to reduce discrepancies iteratively between generalised look-up table emission factors and observed, project-specific embodied carbon data. The feedback mechanism continuously recalibrates environmental coefficients using primary data provided by suppliers, transport information related to logistics, records of material substitution, and validated environmental product declarations. Experimental validation on a corpus of real-world construction projects demonstrated that the semantic-matching module achieved top-1 matching accuracy of 89.3% and top-3 accuracy of 97.1%. Furthermore, the proposed Bayesian correction mechanism reduced the median absolute percentage error in embodied carbon estimation from 28.5% using a conventional static look-up table to 8.3% after three iterative feedback cycles. The results obtained indicate that the proposed architecture establishes a scalable, reproducible pathway towards real-time, evidence-based embodied carbon accounting that is directly integrated with operational construction cost management and procurement processes.