Imagine you teach a new employee to sort incident reports at a construction company. They learn the pattern: first the work setup, then what went wrong, then the actual accident, then the injury. Now you transfer that employee to a chemical plant. The jargon changes — 'scaffolding collapse' becomes 'reactor leak' — but the narrative skeleton is the same. This paper asks whether NLP models can pull off the same trick. The committed claim: fine-tuned French language models trained exclusively on construction-sector accident narratives can classify factual units in metallurgy, chemistry-plastics, and an independent company corpus into four process roles (work situation, unfavourable condition, accident event, consequence) without any target-domain retraining — and they do so at 85.6–85.8% balanced accuracy across all three unseen sectors. The corpus is substantial: 42,244 factual units from 6,040 construction-sector narratives, expert-annotated into the four roles. The authors compare frozen pretrained representations (essentially using a language model as a fixed feature extractor) against task-specific fine-tuning and supervised representation-learning strategies. The verdict is clear: task-adapted models consistently outperform frozen ones on cross-domain transfer. The three leading fine-tuning strategies cluster tightly at 85.6–85.8% average balanced accuracy across target sectors, suggesting the performance ceiling is a property of the task, not the method. The architecture choice matters here. This is a text classification pipeline built on French pretrained language models (the specific model names aren't given in the abstract, but CamemBERT and FlauBERT are the obvious candidates in this ecosystem). The key design decision is whether to freeze the encoder or fine-tune it. Freezing treats the model as a fixed embedding table — cheap but brittle to domain shift. Fine-tuning lets the model reshape its internal representations around the four-role taxonomy, which is precisely what gives it the cross-sector transfer power. On validation integrity, this is a clean experimental design by NLP standards. Training and evaluation domains are strictly separated — no target-domain tuning, no domain adaptation tricks. The target corpora come from genuinely different industrial sectors and include an independently collected company corpus, which is a stronger test than most cross-domain NLP papers bother with. The limitation is that all corpora are French occupational accident narratives following a shared regulatory reporting structure, which likely constrains the narrative forms more than the authors discuss. The milestone question is practical: where does 86% balanced accuracy sit on the usability spectrum? For a fully automated coding system replacing human experts, it's not enough — the error budget is still ~14% per factual unit, and errors compound across multi-unit narratives. For an assisted-coding system where experts review machine-suggested labels, it's plausibly useful today. The next meaningful threshold is probably 92–95% balanced accuracy on truly heterogeneous corpora (not just French-regulated sectors), which would reduce expert review overhead enough to make the system economically compelling at scale. The experiment they didn't run — and the one that would matter most — is multilingual transfer. French occupational accident narratives follow a particular regulatory reporting structure. Does the four-role classification scheme and the learned representations transfer to German, English, or Spanish narratives with different regulatory traditions? The authors likely didn't run this because (a) building expert-annotated corpora in other languages is expensive, and (b) the institutional partners who supplied the data are French. But this is the experiment that would elevate the work from a useful French NLP tool to a generalizable framework for occupational safety research.
