Generative LLM for adaptive taxonomy-informed multi-label classification in construction safety

Abstract

Accurate hierarchical multi-label classification of construction accident narratives is vital, yet traditional methods struggle with the Occupational Injury and Illness Classification System (OIICS), which contains over 2000 labels across four attributes with four-level hierarchies and is subject to taxonomic updates. This paper proposes the Adaptive Generative Construction Classification (AGCC), a taxonomy-informed generative framework that reformulates coding as structured sequence-to-sequence generation. Each attribute is predicted at a variable depth using leaf codes, while taxonomy-informed serialization and semantic-aware label verbalization promote hierarchical consistency. AGCC incorporates Low-Rank Adaptation for parameter-efficient fine-tuning. On OIICS 3.0, AGCC achieves 74.36% Weighted-F1 (Flat) and 84.66% Weighted-F1 (Hierarchical), outperforming full fine-tuning of FLAN-T5-Large by 4.41 and 2.74 percentage points, respectively. Macro-F1 improves by 5.78 and 5.53 percentage points, respectively, while 3.2% of parameters are trainable. AGCC transfers across backbones and demonstrates robustness to taxonomy shifts, offering an accurate and scalable solution for construction safety coding under evolving taxonomies.