Evaluating Transformer Models for Sentiment Analysis of User Reviews in Personal Safety Apps

The use of personal safety mobile applications, such as women's safety apps, SOS alert applications, and location sharing applications, is on the rise to decrease the growing rates of social issues, including harassment, assault, and other associated threats. Although reviews in the Play Store provide useful information about user experiences, the immense quantity of feedback makes it hard for developers and policy-makers to systematically detect patterns and actionable concerns. Sentiment analysis may provide more scalable results in the case of Play Store reviews, but there exists a limited amount of research that uses sentiment analysis in applications related to safety issues. This article uses Transformer-based architectures (BERT, RoBERTa, DistilBERT, and ALBERT) to classify 6,527 user reviews of 15 personal safety apps from the app store into positive, neutral, or negative sentiment. According to the evaluation of our experiments, BERT showed the highest performance with an accuracy of 0.8789, followed by DistilBERT(0.8720) and RoBERTa (0.8697) right behind it in performance. DistilBERT and ALBERT were quicker to provide inference but less stable in evaluating neutral sentiment. The results of our thematic analysis indicated that constructive feedback emphasized trustworthiness, usefulness, and simplicity of operation, while critical feedback mentioned technical glitches, app instability, and GPS problems. These findings show that Transformer-based sentiment analysis can provide more tractable information to developers and policy-makers than star ratings, potentially improving the trust and reliability of safety apps. his.