Technical Language Processing – unlocking information in safety data

In the safety domain we look back to inform the way forward. Our approach to identifying and prioritising risks, the controls we select, and the measures we use to assess the effectiveness of these controls is informed by what has happened in the past. In practice this has meant more reliance on the memories of our people and their life experiences than on the data captured in activity, event and investigation databases, spreadsheets, and reports. This historical data, particularly unstructured text descriptions, is rich in information but is difficult to find, search, extract and transform into useful information efficiently. To complicate matters further, event type and cause identification, labelling and classification is done by one human and seldom reviewed or checked by others. As a result, the meaning of labels is often ambiguous, and this leads to differences in interpretation when data is read by other than the original author. When dealing with data exchange, laborious manual alignment is often required. Terms such as failure mode, failure mechanism, root cause, failure cause, are often used and interpreted differently within and across organisations.