Mining Human Error Incident Patterns with Text Classification

Abstract

Reports on human error incidents contain crucial information to understand why and how incidents happened. There are huge numbers of documents that report industrial accidents or incidents. Instead of reading by humans, we can use document classification technique to find valuable knowledge hidden in the incident reports. Using the technique of document classification, we can detect group of frequent incident. First, the computer observes a set of words (so-called bag-of-word) which appear in each report. Second, the computer calculates similarities among reports. The report similarity is evaluated as agreement ratio of frequency of term appearances. Third, the computer builds the tree of report similarity: this tree holds group of similar reports on a branch. We find the typical patterns of incidents as branches of the tree. Fourth, the computer now calculates similarities of words, which are evaluated as ratio of word co-occurrence.