Extracting safety information from multi-lingual accident reports using an ontology-based approach. (October 2019)

• Record Type: Journal Article
• Authors: Hughes, Peter; Robinson, Ryan; Figueres-Esteban, Miguel; van Gulijk, Coen
• Is Part Of: Safety science. Volume 118(2019)
• Page Start: 288
• Page End: 297
• Publication Date: 2019-10
• Subjects: Industrial accidents -- Periodicals; Accident Prevention -- Periodicals; Safety -- Periodicals; Travail -- Accidents -- Périodiques; 363.11
• DOI: 10.1016/j.ssci.2019.05.029
• Languages: English
• ISSNs: 0925-7535

Abstract

Highlights: Text from 5065 accident reports written in 3 different languages was analysed. A technique was developed to extract safety information from the reports. The technique extracted information regardless of the language used for the report. An independent assessment scored the accuracy at better than 98%. Abstract: This paper describes an approach to extract meaning from multi-lingual free-text safety incident reports. A sample of 5065 safety incident reports from the Swiss Federal Office of Transport were used in the study. Each report was written in either German, French or Italian natural language. An interactive learning approach between a human and computer software was undertaken to identify key terms in the text that are relevant to discovering meaning. A multi-lingual ontology was created to join meaningful semantic patterns and identify specific classes of safety incident on the railway, including injuries occurring whilst passengers were boarding or alighting from vehicles, falling down stairs, struck by closing doors, or struck by objects such as suitcases. A graph database was used to query the text records via the ontology and identify reports of incidents in each class, regardless of the language used in the report. Fluent speakers of each language – German, French and Italian – reviewed the results to confirm true positive results and detect false positives. The performance of the process varied across languages and incident types, however the overall true positive rate was determined by the fluent speakers to be 98.5%.