Abstract
Predatory conversation detection on social media can proactively prevent the netizens, including youngsters and children, from getting exploited by sexual predators. Earlier studies have majorly employed machine learning approaches such as Support Vector Machine (SVM) for detecting such conversations. Since deep learning frameworks have shown significant improvements in various text classification tasks, therefore, in this paper, we propose a deep learning-based classifier for detecting predatory conversations. Furthermore, instead of designing the system from the beginning, transfer learning has been proposed where the potential of the pre-trained BERT (Bidirectional Encoder Representations from Transformers) model is utilized to solve the predator detection problem. BERT is mostly used to encode the textual information of a document into its context-aware mathematical representation. The inclusion of this pre-trained model solves two major problems, i.e. feature extraction and Out of Vocabulary (OOV) terms. The proposed system comprises two components: a pre-trained BERT model and a feed-forward neural network. To design the classification system with a pretrained BERT model, two approaches (feature-based and fine-tuning) have been used. Based on these approaches two solutions are proposed, namely, BERT_frozen and BERT_tuned where the latter approach is seen performing better than the existing classifiers in terms of F-1 and F-0.5- scores.