Rank the gray area
Generate a severity score with structured rule matching and rationale, so close cases stay distinguishable.
EMNLP 2026 · INDUSTRY TRACK
Rethinking
Content Moderation
via Severity-Aware Ranking
A production-scale moderation system that moves beyond rigid yes/no labels to understand the gray area — with human-aligned severity, transparent reasoning, and efficient deployment.
A new moderation axisbinary → severityfine-grained · contextual · actionable
01Human-centric
02Verifiable alignment
03Cost-efficient at scale
Built for the gray area.
01 / THE PROBLEMThe risk spectrum is
The risk spectrum is
continuous.
Traditional moderation collapses nuanced content into a binary decision. That makes benign-but-sensitive samples look like violations, while implicit violations slip through the boundary.
KUROKO treats moderation as a ranked decision over a continuous severity spectrum — matching how people actually review content and how platforms enforce policy.
“From a rigid boundary
to a useful signal.”
to a useful signal.”
Figure 1 · Severity-aware ranking02 / THE SYSTEMOne system.
One system.
Three deliberate moves.
KUROKO connects data, alignment, and deployment into one coarse-to-fine moderation loop.
Figure 2 · KUROKO pipelineVerify the reasoning
SACRM turns policy tiers into programmatic rewards that are inspectable, reproducible, and actionable.
Route only what matters
A lightweight recall stage filters traffic before fine-grained inspection, making production-scale inference practical.
03 / THE EVIDENCEDesigned for
Designed for
real moderation loops.
DATA ENGINE↗
Structured data
from policy logic
Expert audit logic becomes training signal through a multi-stage, verifiable data pipeline.
CASE STUDY↗
Context changes
the right action
The same surface cue can lead to a different enforcement decision when context and severity change.
04 / THE PAPERRead the
Read the
full story.
Rethinking Content Moderation via Severity-Aware Ranking
Read on ACL Anthology ↗Download paper PDF ↗View GitHub repository ↗
FG-CMB datasetRelease forthcoming · currently private
05 / CITATION
Cite this work.
Yang Ou, Jielei Zhang, Yu Dai, Zian Yang, Run Yang, Xiang Li, Weihang Wang, Xueqian Zhao, Peiyi Li, and Longwen Gao. 2026. Rethinking Content Moderation via Severity-Aware Ranking. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 1738–1753, Budapest, Hungary. Association for Computational Linguistics.
DOI: 10.18653/2026.emnlp-industry.100
@inproceedings{ou-etal-2026-rethinking,
title = "Rethinking Content Moderation via Severity-Aware Ranking",
author = "Ou, Yang and
Zhang, Jielei and
Dai, Yu and
Yang, Zian and
Yang, Run and
Li, Xiang and
Wang, Weihang and
Zhao, Xueqian and
Li, Peiyi and
Gao, Longwen",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.100/",
doi = "10.18653/2026.emnlp-industry.100",
pages = "1738--1753",
ISBN = "979-8-89176-460-6"
}