French Meme Risk Annotation Guideline

1. General Principles

• Annotate the meme's message, not your personal agreement or disagreement with it.
• Use the image, OCR text, and description together. A meme's meaning often depends on the combination.
• Assess the three risk dimensions independently. A meme can be harmful, hateful, and propagandistic at the same time.
• Satire does not automatically make content safe. Satire can criticize a harmful message, but it can also spread or normalize it.
• If the signal is real but ambiguous, use potential_risk and uncertain where appropriate.
• Use outside cultural or political knowledge only when needed to understand the meme. Do not invent hidden meaning without evidence.

2. Recommended Workflow

1. Look at the image carefully.
2. Read the OCR text, if available.
3. Read the description.
4. Decide the overall risk_presence.
5. Label is_harmful, is_hateful, and is_propagandistic separately.
6. Assign severity, satire_humor, context_required, and confidence.
7. Check that the labels are logically consistent before submitting.

3. Label Definitions

3.1 risk_presence

• no_risk: no meaningful harmful, hateful, or propagandistic signal is present.
• potential_risk: there is a borderline, ambiguous, indirect, or context-dependent risk signal.
• clear_risk: the meme clearly contains harmful, hateful, and/or propagandistic content.

3.2 is_harmful

Choose yes when the meme could contribute to psychological, social, or physical harm. This includes glorifying violence, spreading dangerous misinformation, humiliating or degrading targets, or encouraging damaging attitudes or behavior.

• yes: the harmful signal is clear.
• no: no meaningful harmful signal is present.
• uncertain: harm is possible, but evidence is weak or context is missing.

3.3 is_hateful

Choose yes when the meme attacks, demeans, stereotypes, excludes, or dehumanizes a person or group based on a protected characteristic such as ethnicity, nationality, religion, gender, sexual orientation, or disability.

• yes: the target and hateful framing are clear.
• no: no hateful attack on a protected group or identity is present.
• uncertain: there may be a protected-group target or hateful implication, but it is not clear enough.

3.4 is_propagandistic

Choose yes when the meme uses manipulation techniques to influence opinion rather than merely express a view. Common signs include:

• fear appeals
• scapegoating
• loaded or emotionally manipulative language
• false dilemmas
• causal oversimplification
• false equivalence

• yes: manipulation is clear.
• no: the meme may be opinionated but not clearly propagandistic.
• uncertain: manipulation may be present, but evidence is borderline.

3.5 severity

• none: use only when risk_presence = no_risk.
• low: weak or limited risk signal, little direct harm, or mild manipulative framing.
• medium: clear risk signal with meaningful harmful, hateful, or manipulative impact.
• high: strong or extreme risk signal, including explicit dehumanization, incitement, or severe manipulation.

3.6 satire_humor

• clear_satire: the meme is clearly satirical, ironic, or mocking, and the satirical intent is easy to recognize.
• possible_satire: humor or irony may be present, but the intent is mixed or unclear.
• not_satire: the meme is straightforward, or any humor is not central to the message.

3.7 context_required

• essential: the meme cannot be judged fairly without external cultural, political, or regional context.
• helpful: context improves the judgment, but a reasonable annotation is still possible without it.
• not_needed: the meme is understandable from the task materials alone.

3.8 confidence

• very_high: you are almost certain.
• high: you are confident, with only minor uncertainty.
• medium: the label is plausible, but there is real ambiguity.
• low: the task is difficult and the final label is uncertain.

4. Easy Decision Rules

5. Example Patterns

• A political meme using fear and blaming one group for all national problems: likely clear_risk, is_propagandistic = yes, and possibly is_hateful = yes if the group is protected.
• A meme calling a protected group "animals" or "parasites": likely clear_risk, is_hateful = yes, and usually is_harmful = yes.
• A sarcastic meme mocking hateful rhetoric in order to criticize it: often clear_satire; risk may be no_risk or potential_risk depending on how clear the criticism is.
• A meme with crude humor or insults aimed at a politician, but without protected-group hate or clear manipulation: may be no_risk or potential_risk, depending on how harmful the content is.

6. Common Mistakes to Avoid

• Do not label content as hateful just because it is rude or aggressive. Protected-group targeting matters.
• Do not assume satire always cancels risk.
• Do not ignore the visual part of the meme. Images often change the meaning of the text.
• Do not overuse clear_risk when the case depends heavily on missing context.
• Do not use no_risk if one of the sub-labels is yes or uncertain.

7. Final Check Before Submit