Cross-Sector Generalization of Accident-Process Role Classification in French Occupational Accident Narratives

Forum topic · 小凯 · 2026-09-22

Summary

This arXiv paper (2609.22081) by Aho Yapi, Pierre Latouche, Arnaud Guillin, and Yan Bailly evaluates how well accident-process role classifiers trained on French occupational accident narratives generalize to unseen industrial sectors. The authors built an expert-annotated corpus labeling factual units into four roles: work situation (A0), explicitly reported unfavorable conditions (A1), accident event or deviation (B), and reported consequences (C). Classifiers were developed solely on 42,244 factual units extracted from 6,040 construction-sector narratives, then tested without retraining or target-domain fine-tuning on metallurgy, chemicals-plastics, and independently collected corporate corpora. The study compared frozen pretrained representations, task-specific fine-tuning, and supervised representation learning. Task-specific adaptation consistently outperformed frozen-representation transfer: across repeated runs, three leading adaptation strategies achieved average balanced accuracy between 85.6% and 85.8% on the three target corpora. The findings support deployable, transferable auxiliary coding systems that can consistently structure heterogeneous occupational accident narratives for expert review and cross-sector prevention analysis.

Paper Overview

Field: NLP Authors: Aho Yapi, Pierre Latouche, Arnaud Guillin, Yan Bailly Published: 2026-09-18 arXiv: 2609.22081

Abstract

Occupational accident narratives contain valuable information about work situations, unfavorable conditions, accident events, and their consequences. Automatically structuring these narratives can facilitate large-scale accident analysis and support occupational risk prevention. However, the terminology and writing styles used to describe accidents vary considerably across sectors and organizations, raising questions about whether automated coding systems can generalize beyond their training domain.

This paper evaluates the cross-sector generalization of accident-process role classification in French occupational accident narratives. The authors construct an expert-annotated corpus in which factual units are classified into four roles:

A0 — work situation
A1 — explicitly reported unfavorable conditions
B — accident event or deviation
C — reported consequences

Methodology

Role classifiers were developed and selected using only 42,244 factual units extracted from 6,040 construction-sector narratives.
Evaluation was performed on unseen corpora from the metallurgy and chemicals-plastics sectors, plus an independently collected corporate corpus, without any retraining or target-domain fine-tuning.
Three approaches were compared: frozen pretrained representations, task-specific fine-tuning, and supervised representation learning.

Key Findings

Task-specific adaptation consistently outperformed frozen-representation cross-domain transfer.
Across repeated training runs, the three leading adaptation strategies achieved average balanced accuracy between 85.6% and 85.8% on the three target corpora.
These results support the development of transferable auxiliary coding systems that can consistently structure heterogeneous occupational accident narratives for expert review and cross-sector prevention analysis.

--- *Auto-collected on 2026-09-22*