Automated Coding of Injury and Illness Data
The Survey of Occupational Injuries and Illnesses (SOII) collects data from sampled establishments on OSHA forms 300 and 301. We use the information provided on these forms to generate detailed statistics on the characteristics of cases involving injury or illness.
In order to generate these statistics, survey staff must convert the text entries in the OSHA forms to standard codes used by BLS, as indicated in the table below:
The set of all fields, taken together, is considered the case "narrative." Prior to survey year 2014, BLS exclusively relied on humans to code cases. In 2014, BLS began using machine learning to code a subset of cases. To use machine learning we first select a learning algorithm and then train it on large quantities of previously coded SOII narratives. During this process the algorithm calculates how strongly various features, such as words, pairs of words, and other items are associated with the codes that can be assigned. After training, we use the algorithm to estimate the best codes for each uncoded narrative and assign those codes if the model’s confidence exceeds a predetermined threshold. For 2014-2017 BLS used regularized multinomial logistic regression. In 2018, BLS switched to deep neural networks with character-level convolutional embeddings and Long-Short-Term-Memory recurrent layers (source code is available here). In 2019, BLS began autocoding secondary source for the first time.
BLS use of autocoding has expanded significantly over time. In 2014, only 5 percent of codes and only occupation codes were assigned by machine learning. By 2019 automatic coding had been expanded to include all six primary coding tasks (occupation, nature, part, source, secondary source and event) with the model assigning approximately 85% of these codes.
For additional technical information on our techniques, please contact OSHS_Autocoding@bls.gov.
Last Modified Date: September 21, 2020