Modern NLP for Proactive Harmful Content Moderation
summary
Compares GPT‑4, Llama/Mistral and other LLMs to encoder transformers (mT5, mT0, mBART), finding out‑of‑the‑box LLM classification often performs poorly while fine‑tuned encoders beat them with far less compute; LLMs generate high‑quality explanations and useful counter‑speech but multilingual detoxification shows big drops outside high‑resource languages, with no single model working well for all languages; builds parallel detox corpora and target‑aware counter‑speech decoders, recommends accuracy below 80% for off‑the‑shelf LLMs and hundreds of examples per language for fine‑tuning.
AI-written summary · about
chapters (13)
- 0:00NLP for harmful content moderation
- 1:13Researcher background
- 3:19Hate speech definitions
- 5:06Alignment gaps in definitions
- 7:58Datasets and toxicity models
- 9:00LLMs vs encoder models
- 10:21Toxifying text with LLMs
- 11:46Multilingual toxic rewriting
- 12:51LLMs for toxic span explanations
- 14:13Cultural toxic patterns
- 15:17Counter speech generation
- 17:26Takeaways and future research
- 18:33Q&A
related talks
40mThe Future of Language Models
30mML Automoderation System for LLM-Augmented Online Platforms
43mBeyond the Hype: A Realistic Look at Large Language Models
53mLLM Quality Optimization Bootcamp
30mWhere there's a will, there's a way: The mad genius of LLMs as a classifiers
29mLocal LLM fine-tuning: a practical example
24mHarnessing the Power of LLMs in Advanced Analytics
40mPanel discussion "NLP"