Classification
Classification is the most flexible labeling work type for evaluating, comparing, or categorizing data. Experts answer predefined questions about images, videos, text, or audio — without drawing or annotating on the data itself. Assign one or more labels from a fixed set to your data — e.g., “cat”, “blurry”, “positive sentiment”. Unlike annotation-based work types, Classification captures structured responses to questions you define.
Ideal for: High-level categorization, quality assessment, content moderation, and multi-input comparisons.
¶ When to Use
| Task Type | Examples |
|---|---|
| Single-Input Tasks | “Is there a car in this image?” “Is this video blurry or clear?” “What language is being spoken?” “Is this email spam or not?” |
| Multi-Input Tasks | “Which landing page looks better?” (compare 2 images) “Does this audio match the video?” (image + audio) “What is the correct output?” (code + result comparison) |
| Subjective Judgments | Quality assessment, sentiment analysis, content appropriateness, preference ranking |
¶ When to Use Other Work Types
- Need pixel-level accuracy? → Segmentation
- Need object locations? → Object Detection
- Need temporal events? → Event Tagging
- Need entity extraction? → Named Entity Recognition
- Need document fields? → Document Annotation
¶ Supported Data Types
- Image, Video, Audio, Text - Single or multiple inputs per task
- Mixed Media - Combine different data types in one task
¶ Key Features
Custom UI Builder (ATGL) - Design tailored interfaces for your specific needs:
- Display multiple videos side-by-side
- Overlay audio with transcript panels
- Present code snippets with output regions
Consensus Labeling - Have 3-7 experts independently label each task. The system uses majority voting to determine the final result.
- Higher accuracy for subjective tasks
- Reduces individual bias
- Provides confidence scores
¶ Example Scenarios
| Input Type | Question Example |
|---|---|
| Image | “Is this photo clear enough for the website?” |
| Audio | “Is the speaker’s accent native?” |
| Text | “Does this express positive sentiment?” |
| Image Pair | “Which design is more appealing?” |
| Image + Audio | “Does the sound match this scene?” |
¶ Classification Types
- Single-label - One category per item
- Multi-label - Multiple categories per item
- Binary - Yes/No decisions
- Hierarchical - Nested categories
- Comparative - Rank or compare items
¶ Getting Started
- Define your classification categories and questions
- Configure custom UI layouts if needed (optional)
- Enable consensus for multiple expert opinions (optional)
- Prepare your input data
- Submit batches via API
- Track progress and download results
¶ How to Submit Batches
Indirect Demand (CSV File) - Best for large-scale projects with thousands of items.
Direct Demand (Inline Data) - Best for dynamic tasks or real-time processing.
¶ Best Practices
- Clear Questions - Make questions unambiguous and mutually exclusive
- Provide Context - Include relevant metadata for better decisions
- Use Consensus - Enable for subjective or opinion-based tasks
- Custom Layouts - Leverage ATGL for complex multi-input scenarios
- Quality Guidelines - Document edge cases and ambiguous scenarios