An AI and NLP approach for detecting grooming behavior

Abstract

"Grooming children on social media is a dangerous side effect of modern internet era. AI models, specially NLP have the potential to play a critical role in detecting grooming behavior. Even though, there have been studies in the past to build a grooming detection system, there is limited research on building such systems us- ing modern NLP techniques. In this paper, we propose a modern sexual grooming detection system using state-of-the-art NLP models and techniques that can detect and alert users to potentially dangerous online interactions between groomers and their targets. Our detection system is a ConversationClassifier which is able to clas- sify conversations, whether they are grooming or not. With over 19,000 grooming sentences collected from PervertedJustice grooming conversations, we created an annotated dataset exhibiting the grooming characteristics. Conversational data was also collected from both PervertedJustice and PAN12 dataset. With the sentence- level annotated dataset, we trained a SentenceClassifier model based on RoBERTa & DeBERTa to be able to accurately predict if a sentence has grooming character- istics or not. The ConversationClassifier was built on top of the SentenceClassifier with LSTM & GRU to capture the sequential features in the conversation. Further- more, a self-attention mechanism was added so that the model can focus on relevant sentences. Our models achieved promising results. In case of the SentenceClassifier, it displayed an accuracy of 93% for RoBERTa and 94% for DeBERTa. We paired the RoBERTa based SentenceClassifier with LSTM which yielded an accuracy of 97% and DeBERTa based SentenceClassifier with GRU which yielded an accuracy of 95%."