Research map / Large language models / Language, safety and interpretability
A Holistic Approach to Undesired Content Detection in the Real World
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed classifier instead of Perspective API | content moderation classification · Public / S, H, V, HR, SH, S3, H2, V2 | AUPRC: .8709* -> .9703; .6914 -> .7968; .5201 -> .7371; .3902* -> .6191 | Helped |
| Use the proposed classifier instead of Perspective API | content moderation classification · Jigsaw / Identity-hate | AUPRC: .6644 -> .6890 | Helped |
| Use the proposed classifier instead of Perspective API | content moderation classification · Jigsaw / Insult, Obscene, Threat, Toxic | AUPRC: .8814 -> .8548; .9500 -> .8353*; .7492 -> .6144*; .9769 -> .9304* | Mixed |
| Use the proposed classifier instead of Perspective API | content moderation classification · TweetEval / Hate | AUPRC: .5961 -> .6473 | Helped |
| Use the proposed classifier instead of Perspective API | content moderation classification · TweetEval / Offensive | AUPRC: .7919* -> .7024* | Hurt |
| Use the proposed classifier instead of Perspective API | content moderation classification · Stormfront / Hate | AUPRC: .8754 -> .9053 | Helped |
| Use the proposed classifier instead of Perspective API | content moderation classification · Reddit / Sexual | AUPRC: .8961* -> .9417* | Helped |
| Select production training examples with active learning instead of random sampling | content moderation model · production traffic, all eight categories | AUPRC | Helped |
| Select examples with active learning instead of random sampling | content moderation data collection · production traffic, undesired sample yield | yield: 10+ times | Helped |
| Add curated synthetic examples to mitigate counterfactual bias | hate classifier · hateful content | AUPRC: 0.417 -> 0.551 | Helped |
| Add a large amount of noisy zero-shot and few-shot synthetic data | content moderation model · existing labeled training data | model performance | Hurt |
| Add Wasserstein domain adversarial training (ablation) | content moderation model · production validation / PUB | AUPRC: SH: 0.063 -> 0.281; H: .417 -> .491; V: .490 -> .529; HR: .258 -> .369; SH: .063 to 0.281 | Helped |
Closest papers by meaning
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.85)
- Efficient Toxic Content Detection by Bootstrapping and Distilling Large Language Models (AAAI 2024 · similarity 0.85)
- Human Texts Are Outliers: Detecting LLM-generated Texts via Out-of-distribution Detection (NeurIPS 2025 · similarity 0.84)
- BingoGuard: LLM Content Moderation Tools with Risk Levels (ICLR 2025 · similarity 0.84)
- Uncovering and Mitigating the Hidden Chasm: A Study on the Text-Text Domain Gap in Euphemism Identification (AAAI 2024 · similarity 0.84)
- Hateful Meme Detection through Context-Sensitive Prompting and Fine-Grained Labeling (Student Abstract) (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.