Antagonistic online communities, such as the misogynist incel subculture, often develop a distinct linguistic style that is sometimes difficult for outsiders to interpret. A central challenge for law enforcement is identifying when common words are repurposed with new meanings. Such polysemous jargon words complicate the detection of threats and radicalization cues, as unfamiliar word senses can be hard to recognize. We propose a prompt-based framework that leverages Large Language Models (LLMs) to generate definitions for polysemous words and classify word usage accordingly. Additionally, we introduce an LLM-as-a-judge scoring method to evaluate the framework and perform evaluation on data from an online incel platform. Our results show that the proposed prompt-based, corpus-level approach outperforms existing fine-tuned, sentence-level generation methods.