Content Category Taxonomies
What you'll learn: How content moderation systems classify harmful outputs into standardized categories to enable consistent filtering and policy enforcement.
The Problem of Unstructured Harm
When you built input and output filters (lessons 1430-1431), you learned where to check content. But what exactly are you checking for? Without a clear taxonomy, every engineer would define "harmful" differently, leading to inconsistent protection.
Content category taxonomies are standardized classification schemes that break down harmful content into discrete, well-defined buckets. Think of them as the "periodic table" of harmful content—each category has specific boundaries, examples, and severity levels.
The Six Core Categories
Most production systems recognize these foundational categories:
Hate Speech: Content targeting protected characteristics (race, religion, gender, etc.) with dehumanizing or violent language. Example: slurs, calls for exclusion.
Violence: Graphic depictions, glorification, or instructions for physical harm. Includes terrorism, weapons, and gore.
Sexual Content: Explicit sexual material, especially involving minors or non-consent. Ranges from suggestive to explicit.
Self-Harm: Content promoting suicide, eating disorders, or self-injury. Includes methods and encouragement.
Harassment: Targeted abuse, doxxing, stalking, or sustained intimidation of individuals.
Policy Violations: Platform-specific rules like spam, misinformation, copyright infringement, or illegal activities.
Why Taxonomies Matter
Your filtering logic (from lesson 1431) relies on these categories. When a moderation API returns {"category": "violence", "severity": 0.89}, you know exactly what rule to apply. Taxonomies enable you to set category-specific thresholds, log structured metrics, and explain decisions to users.
Key Takeaway: Content category taxonomies transform vague "safety" requirements into concrete, actionable classification systems that your filtering pipelines can enforce consistently across all inputs and outputs.