Organizations investigated for accidents explicitly document causes, contributing factors, and lessons learned, yet the same factors often reappear in later accidents. This problem asks whether such organizational learning decay is detectable and measurable directly from longitudinal corpora of accident investigation report texts: linking causes, contributing factors, and lessons documented in earlier reports to their recurrence (or absence) in later reports from the same organizations, sectors, or systems; quantifying the rate at which documented knowledge of previously identified factors decays; and establishing whether textual traces can distinguish retained from lost lessons. Adjacent literatures document cause recurrence by classification codes, model forgetting by simulation, and measure lesson retention by interviews and questionnaires, while text-mining studies of accident reports stop short of cross-temporal, lesson-level linkage — leaving the corpus-based measurement methodology and its empirical validation open.

Background

Organizational learning refers to the processes by which organizations convert experience into changes in routines, systems, and processes; in safety-critical industries, a central instance is learning from incidents (LFI): after an accident or incident, an investigation produces a report that assigns causes and contributing factors (conditions or actions that helped produce the unwanted outcome without being sufficient alone) and states lessons learned (documented conclusions intended to prevent recurrence). A recurring finding across high-hazard sectors is that documented causes recur: in the chemical process industries, analyses of major accidents repeatedly identify the same categories of causes reappearing over years (Belke, Loss Prevention Bulletin 159, 2001), and simulation work formalizes this as a stock of 'Lessons Learned' that is continuously depleted by organizational forgetting (memory loss through turnover, restructuring, and fading attention), so that relative safety knowledge grows only when the rate of incident learning exceeds the rate of memory loss (Cooke and Rohleder, System Dynamics Review 22, 2006).

Empirically, whether lessons are retained or forgotten over time has so far been measured by surveys and interviews: a Dutch academic hospital's COVID-19 crisis lessons were tracked over three years through a 2020 staff survey followed by in-depth interviews, which found lessons on infection prevention and staff deployment largely retained while lessons on decision-making and interdepartmental collaboration were largely forgotten (Broekema, Louisa and Schmidt, BMC Health Services Research 26, 2026); longitudinally, the quality of hospitals' learning from 4,667 investigated incidents was assessed over six years by inspector scoring rather than by analyzing the reports themselves (de Kam et al., Health Policy 124, 2020); and questionnaire instruments such as the Learning from Incidents Questionnaire measure perceived learning-process quality through self-report (Littlejohn et al., Safety Science 99, 2017).

In parallel, natural language processing (NLP) and text mining have been applied to the report texts themselves: extraction of safety risk factors from construction accident reports (Xu et al., Safety Science 138, 2021) leaves the relationships among factors unmodeled, listing the causal-and-effect relationship among extracted factors and the factor-and-severity relationship as domain-bound future analyses of construction accident reports; in aviation, topic modeling — where latent Dirichlet allocation (LDA) discovers co-occurring word clusters ('topics') in a document collection without annotations — was applied to occurrence-report corpora to infer cause and effect co-occurrences, but its precision proved too low for inferring relations between causes and between causes and effects, and no linkage of report content across time was attempted (Jonk et al., arXiv:2301.05663, 2023); temporal topic modeling of aviation safety reports tracks how the mixture of themes evolves over the years but not the fate of specific documented lessons (Robinson, Safety Science 116, 2019).

From the software-incident side, a study of learning from failure states verbatim that 'How learnings change over time as more failures are experienced is an open question' and calls for models of how learnings and proposed changes are managed over time, including how intervention effects on recurrence could be tested (Sillito and Pope, arXiv:2402.09538, 2024); a review of cybersecurity incident response likewise characterizes retention of learning as dependent on retention activities, storage affordances, and internalized causal maps (Lakshmi, Naseer, Maynard and Ahmad, arXiv:2107.02941, 2021). Thus the phenomenon (causes recur; lessons fade) is established by codes, models, and interviews, and the texts are mined for factors and themes — but no line of work measures learning decay from the investigation-report corpora themselves.

Problem Statement

Establish whether organizational learning decay is detectable and measurable from longitudinal corpora of accident investigation report texts. The question spans three dimensions, all of which a resolution must address; a contribution limited to a single dimension is a partial result rather than a resolution: (1) linking causes, contributing factors, and lessons explicitly documented in earlier investigation reports to their recurrence or non-recurrence in later reports concerning the same organizations, sectors, or systems, handling changes in vocabulary, taxonomy, and reporting practice over the corpus time span; (2) empirical decay estimates reported as two distinct quantities rather than one — the rate at which documented knowledge of previously identified factors fades from later reports, and the rate at which that knowledge fails to prevent those factors from reappearing as documented causes in later reports — each with uncertainty and with attention to confounds such as reporting volume, exposure, and structural reorganizations; and (3) whether textual traces can distinguish retained from lost lessons. The question is open in both directions. Its scope across sectors, languages, document types, and analytic methods is not restricted to the scope of any single cited source; no single sector, report genre, or technique is prescribed.

References

  1. Sillito, J., Pope, M. (2024). Learning From Lessons Learned: Preliminary Findings From a Study of Learning From Failure. arXiv:2402.09538. https://arxiv.org/abs/2402.09538
  2. Jonk, P., de Vries, V., Wever, R., Sidiropoulos, G., Kanoulas, E. (2023). Natural Language Processing of Aviation Occurrence Reports for Safety Management. arXiv:2301.05663. https://arxiv.org/abs/2301.05663
  3. Xu, N., Ma, L., Liu, Q., Wang, L., Deng, Y. (2021). An improved text mining approach to extract safety risk factors from construction accident reports. Safety Science 138:105216. doi:10.1016/j.ssci.2021.105216. https://doi.org/10.1016/j.ssci.2021.105216
  4. Cooke, D. L., Rohleder, T. R. (2006). Learning from incidents: from normal accidents to high reliability. System Dynamics Review 22(3):213-239. doi:10.1002/sdr.338. https://doi.org/10.1002/sdr.338
  5. Belke, J. C. (2001). Recurring causes of recent chemical accidents. Loss Prevention Bulletin 159:6-10. doi:10.1205/026095701750285490. https://doi.org/10.1205/026095701750285490
  6. Broekema, W., Louisa, H., Schmidt, E. (2026). Why organizations forget or retain lessons learned from a crisis over time: tracking COVID-19 lessons in an academic hospital in the Netherlands. BMC Health Services Research 26(1):766. doi:10.1186/s12913-026-14429-x. https://doi.org/10.1186/s12913-026-14429-x
  7. de Kam, D., Kok, J., Grit, K., Leistikow, I., Vlemminx, M., Bal, R. (2020). How incident reporting systems can stimulate social and participative learning: A mixed-methods study. Health Policy 124(8):834-841. doi:10.1016/j.healthpol.2020.05.018. https://doi.org/10.1016/j.healthpol.2020.05.018
  8. Littlejohn, A., Margaryan, A., Vojt, G., Lukic, D. (2017). Learning from Incidents Questionnaire (LFIQ): The validation of an instrument designed to measure the quality of learning from incidents in organisations. Safety Science 99:80-93. doi:10.1016/j.ssci.2017.02.005. https://doi.org/10.1016/j.ssci.2017.02.005
  9. Robinson, S. D. (2019). Temporal topic modeling applied to aviation safety reports: A subject matter expert review. Safety Science 116:275-286. doi:10.1016/j.ssci.2019.03.014. https://doi.org/10.1016/j.ssci.2019.03.014
  10. Wu, X., Sun, P. (2024). Dynamic Analysis and Temporal Governance of Safety Risks: Evidence from Underground Construction Accident Reports. Sustainability 16(19):8531. doi:10.3390/su16198531. https://doi.org/10.3390/su16198531
  11. Lakshmi, R., Naseer, H., Maynard, S., Ahmad, A. (2021). Sensemaking in Cybersecurity Incident Response: The Interplay of Organizations, Technology and Individuals. arXiv:2107.02941. https://arxiv.org/abs/2107.02941