Bowties are a common hazard analysis tool used within aerospace and other safety critical industries, where they have the advantage of clearly depicting relationships between causes, hazards, consequences, and barriers (hazard controls). For a given bowtie, it is useful to understand how these barriers are performing, which is where occurrence reports provide an insight—describing incidents and how effective the barriers were. This paper presents a Natural Language Processing method—Aspect Sentiment Triplet Extraction, adapted to extract barriers from free-text occurrence reports alongside their effectivity and sentiment—a distinct advantage over pure entity extraction methods. The development of a bespoke labelled dataset to train a Bidirectional Encoder Representations from Transformers (BERT) model required extensive safety expertise, where the resulting model output was then qualitatively assessed. The model was ultimately applied to occurrence reports pertaining to lithium battery issues, proving useful in demonstrating key areas where the associated bowtie barriers were operating effectively, ineffectively, or in situations where no barriers were seemingly present. This was especially evident when the output was processed using an unsupervised k-means algorithm to display generalized themes. The method forms a useful tool for safety professionals who are concerned with extracting bowtie barrier information from high quantities of occurrence or incident data while addressing a critical gap in the body of knowledge by going beyond present entity extraction methods.
Jon Ricketts (Sat,) studied this question.