Low-Frequency Aware Unsupervised Detection of Dark Jargon Phrases on Social Platforms

With the development of the Internet, the number of people communicating on social platforms has soared, which means that it is crucial for platform moderators to review and remove illegal content to create a clean network environment for users. However, identifying such content becomes complex due to the use of dark jargons. These jargons are seemingly innocent or newly coined words and phrases, such as “coke” for cocaine or “vanilla sky” for synthetic cathinone, to convey illegal meanings, aiming to evade detection by moderators. Existing methods primarily focus on detecting dark jargons at the word level, yielding commendable results. However, given the prevalence of phrase-level dark jargons in the context, relying solely on word-level detection can introduce ambiguity. For example, “black” is not a dark jargon, but “black bart” is a dark jargon. As a result, there is a growing interest in developing techniques specifically targeting phrase-level dark jargon detection. Unfortunately, such efforts are relatively limited, potentially resulting in the oversight of numerous low-frequency dark jargon phrases. To tackle this challenge, we propose the Low-Frequency Aware Dark Jargon Phrases Detection (DJPD) model. Our approach centers around finding a noun phrasal attention map pattern based on Transformer that enhances the perception of low-frequency phrases, enabling the selection of candidate dark jargon phrases. Subsequently, the candidate dark jargon phrases’ sentence-level context is analyzed to detect dark jargon phrases. Remarkably, our model achieves a significant 84.66% improvement in F1-score compared to the current state-of-the-art method for dark jargon phrase detection in the corpus.