Reproduce “Covering Cracks in Content Moderation: Delexicalized Distant Supervision for Illicit Drug Jargon Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minkyoo Song, Eugene Jang, Jaehan Kim, Seungwon Shin
JEDIS, a framework for context-aware illicit drug jargon detection, is introduced, utilizing delexicalized distant supervision to train models without extensive human-labeled data. The framework achieves F1-scores of 0.6419 on Reddit Drug and 0.6320 on Silk Road Forum datasets, outperforming prior state-of-the-art systems and detecting significantly more unique jargon terms.
Key insights
- Delexicalized distant supervision effectively trains robust drug jargon detectors by forcing models to learn from context rather than specific words, enhancing resilience to evasion tactics and euphemisms.
- Ensembling a context-only module with a word attribute module, both domain-pretrained, leverages both contextual and lexical information for more precise and comprehensive jargon detection.
- Careful design of negative sampling strategies and selection of optimal contextual window sizes during distant supervision are crucial for balancing precision and recall in identifying evolving and ambiguous drug jargon.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.