Using text mining to sub-classify safety concern mentions in online reviews

In this study, we build on previous work in safety monitoring of online reviews by adapting injury severity coding schemes from the Consumer Product Safety Commission (CPSC) and Health Canada manuals, typically used in hospital surveillance. These manuals enable the sub-classification of reviews by safety-event and hazard type. We analysed a large dataset of Amazon.com reviews across 17 product categories, along with a secondary dataset from a second Fortune 50 retailer. We manually tagged each review to identify any mention of safety concerns and developed lists of specific terms, known as ‘smoke’ terms—unique words and phrases that signal safety issues. These terms were then used to detect, subcategorise, and prioritise safety concerns across different product categories. The study revealed that using smoke terms allowed for more effective detection and categorisation of safety concerns than traditional sentiment analysis. This text mining approach offers a more efficient method for discovering safety concerns in online reviews, benefiting organisations and regulatory agencies by helping them prioritise safety issues for deeper analysis and potentially preventing consumer product-related injuries.