ConstructCIE dataset targets causal extraction from construction accident reports

Construction accident narratives are rich with causal information, but that evidence is often implicit, long-span, and distributed across the text. To tackle this, researchers from academia introduced ConstructCIE, a manually annotated dataset built from OSHA construction accident reports. The dataset uses a hierarchical schema that covers accident types, causal factors, sub-causal factors, and supporting evidence spans, providing a structured framework for training and evaluating causal information extraction systems. The paper, posted on arXiv (2608.06495), benchmarks both supervised sequence taggers and instruction-tuned LLMs in an end-to-end hierarchical extraction setting.

Results show that most evaluated models achieve strong performance on accident-type prediction and can recover broad causal meaning, but they remain limited when it comes to precise span-level extraction. The researchers compared joint hierarchical extraction (JHE) and iterative hierarchical extraction (IHE) strategies, finding JHE generally delivers stronger exact and soft matching, while IHE sometimes achieves higher keyword F1. Error distributions vary by extraction strategy, but evidence-selection and span-boundary errors remain common across the board. The findings indicate that reliable causal information extraction for construction accidents requires stronger domain grounding and more accurate evidence extraction, pointing toward future work in integrating domain knowledge and refining evidence-span modeling.