Extracting typical incident patterns from text data

Abstract

To prevent industrial incidents, it is important to learn why and how past incidents occurred and escalated. Information regarding accidents is recorded primarily in natural language texts, which are not convenient for analysing incident progression. This paper proposes a method for recognising the typical flow of events in a large set of text reports. Our method transforms each sentence in reports about industrial incidents into a vector (bag-of-words) to facilitate the detection of similar contexts and stories. In this way, we can recognise the typical progression of accidents.