Abstract
With the rapid increase of smartphones and the growing interest in their applications, e.g., Google Play Apps, it becomes necessary to analyze users’ reviews whether they are expressed as ratings or comments. This is because recent studies reported that users’ reviews could provide us with useful clues and valuable features that can help in understanding the broad opinion about some applications in term of security awareness. Several techniques have been developed for this crucial task and significant progress have been achieved such as Semantic and Sentiment Analysis, Topic Modelling, and Clustering. The majority of the existing methods are mainly based on representing reviews’ words in a Bag-Of-Words vector space with String-matched approaches without considering the common polysemy and synonymy problems of words. This is true due to the fact that users who make use of these applications are often from a diverse background and thus, different vocabulary. This paper proposes a new approach to classifying security opinions about applications from users’ reviews while considering special features of synonymous and polysemous words. To achieve this task, the proposed model makes use of word embedding, topic modelling, Bi-LSTM, and n-grams approach. For the proposed model, a new dataset is built that contains reviews about 18 popular applications. The application’s selection was primarily governed by making the dataset diverse in its domain. The experiment results showed that the proposed ensemble model which combines the prediction of the extracted features, which in turn captures synonymy, polysemy, and dependency of words-is significantly useful, and it achieves better results with an accuracy approaching 90% compared to the use of each technique separately. The model could contribute in preventing mobile users from unsafe applications.