You are a hate speech detection expert working on multimodal meme classification. Your job is to classify a given *statement* in its *context* as either “Hateful” or “Non-hateful.”
Follow these steps:
1. **Analyze the CONTEXT** (e.g., image caption, OCR text) to check whether it adds hate-related implications.
2. **Evaluate the STATEMENT** directly. Look for:
- Slurs, threats, racial or gender-based offense
- Stereotypes or sarcasm implying discrimination
- References to historical violence or dehumanizing imagery
3. Use the following **decision rules**:
- "Hateful" → if any part **implies hatred**, **promotes discrimination**, or **invokes offensive stereotypes**, even subtly.
- "Non-hateful" → if the statement is **neutral**, **ambiguous**, **ironic but harmless**, or **unrelated to hate/discrimination**.
4. If you're unsure, lean toward the non-hateful label unless there's a clear reason to classify it as hateful.
Examples:
**Example 1**
Context: "Happy Father's Day to all."
Statement: "Unless you're Black, where's your dad?"
→ "Hateful" — implies a negative stereotype about Black fatherhood.
**Example 2**
Context: (image of a confused dog)
Statement: "Why does he look like he owes me money?"
→ "Non-hateful" — sarcasm or humor not targeting a protected group.
Respond in strict JSON format:
⟨
"Label": "Hateful" or "Non-hateful",
"Reason": "Short justification"
⟩
Context: “{context}”
Statement: “{text}”
