Introduction:
Online shopping and E-commerce are challenging new areas for product safety authorities but also provide a new source of product information. The large number of online customers is likely to leave product reviews that include feedback on safety concerns like product hazards, missing information, or suggestions for improvement.

Objectives:
We aimed to monitor the emergence of safety and security risks within relevant product categories using online costumer review data. To achieve this, we utilized software that automatically downloads the selected reviews into a database and then employs natural language processing techniques to analyze the data. The software should be accessible through a web application.

Methods:
Language models such as GPT-3 are utilized as a methodology to identify and summarize safety risks. Problem areas will be identified by employing the BERTopic clustering algorithm to group the identified issues into clusters. This algorithm utilizes advanced machine learning techniques to identify common themes and topics within the review data, which can then be used to group related issues. For this study, the software was applied to one of the big online shopping portals.

Results:
Through a dashboard the main quality, functionality and design problems with the products are highlighted. The time course and distribution of the assessments can be displayed, and specific products and reviews within each problem area can be listed. For “chainsaws”, e.g., the following main safety issues were ranked as “poor” from 1 200 costumer reviews: hot start performance, chain tension, chain tensioner counter screw, thumb screw (and many more).

Conclusion:
Using this methodology and software, online costumer reviews can quickly and efficiently be screened for safety-related issues for any product category of interest. This enables insights into product safety issue in the vast online shopping market. This, in turn, will allow authorities to take proactive measures to address these issues and improve the safety of products.

Keywords: Product hazard detection, E-commerce, Automatic data analysis, Language models

Other info: University of Applied Sciences Upper Austria - Hagenberg - Upper Austria - AustriaAnna Maschek*, Robert Bauer**, Johanna Trauner-Karner**. * University of Applied Sciences Upper Austria, ** Austrian Road Safety Board (KFV).