Data‑Centric AI Practices for Content Moderation at NetEase Yidun

With the rapid development of AI, NetEase Yidun leverages Data‑Centric AI to improve the efficiency of harmful content interception, emphasizing the need to understand data scenarios, define and manage data, and maximize its value for algorithmic innovation.

The concept of Data‑Centric AI, originally proposed by Andrew Ng, highlights that data contributes the most to AI performance; however, both academia and industry often under‑invest in data, focusing instead on models and solutions.

In the content moderation domain, massive internet data streams contain a tiny fraction of violations, leading to extreme long‑tail distributions, highly similar non‑violating features, tiny target objects, and open‑domain categories that constantly evolve.

Yidun’s data workflow includes:

Data definition and characterization to reduce model and labeling difficulty.

Data collection and generation to support cold‑start and continuous expansion.

Multi‑level labeling with refined documentation, model‑assisted pre‑labeling, and selective human annotation.

Data augmentation, selection, and cleaning to ensure high‑quality training samples.

Iterative data cycles that integrate semi‑supervised, self‑supervised, and weakly supervised methods for stability, domain adaptation, and fine‑grained supervision.

Testing and validation align data and metrics with real‑world deployment goals, ensuring that offline performance translates to online effectiveness.

The Q&A section discusses effective integration of model pre‑labeling with human annotation and strategies for handling ambiguous class boundaries, including hierarchical model design and retrieval‑based augmentation.

Overall, Yidun’s Data‑Centric AI framework aims to minimize resource consumption while maximizing data value across the entire AI lifecycle for content safety.