Event Log Extraction Using Natural Language Processing and Text Embeddings

Process mining in the construction industry relies heavily on high-quality event logs, yet existing approaches primarily use structured data sources, leaving unstructured textual reports largely unexploited. This paper proposes a novel two-stage methodology to extract and preprocess event logs from unstructured, natural language on-site construction reports. Initially, domain-customized natural language processing techniques combined with IBM’s multilingual granite text embeddings transform noisy, heterogeneous task descriptions into dense vector representations. Semantic clustering using cosine similarity then aggregates linguistically diverse but conceptually identical activities into unified clusters. Subsequently, a temporal consolidation module leverages the longitudinal occurrence of activities across monthly reports, incorporating spatial metadata to infer continuous task durations and transitions. Evaluation on 2,424 task instances demonstrates effective reduction of redundancy and robust grouping into approximately $102-111$ semantically coherent clusters. This approach addresses a critical knowledge gap in construction process mining by enabling incorporation of rich textual data, thus enhancing event log completeness and quality for downstream analysis. This advancement enables enhanced workflow tracking and delay detection in construction projects, broadening the scope of process mining beyond traditional reliance on structured data sources. Our work lays foundational steps for integrating unstructured textual data into construction process analytics, offering practical insights for improved project monitoring and decision-making.