ArPoMeme: An Annotated Arabic Multimodal Dataset for Political Ideology and Polarization

The dataset's core methodological claim is that organic community self-identification of political ideology produces more reliable labels than researcher annotation. This inverts the typical workflow and assumes Facebook users' own ideological framing is more valid ground truth than expert coding.

This work sits alongside the psychiatric diagnosis coding study from the same batch, which validated that domain-specific embeddings outperform generic NLP on non-English clinical text. Both papers address the same structural problem: major language models remain undertrained on non-English, domain-specific corpora, leaving downstream systems vulnerable to misclassification. ArPoMeme extends that insight to multimodal political discourse, where cultural context and regional polarization patterns are inseparable from the task itself. The annotation methodology also echoes concerns raised in the persona and sycophancy work, which showed that externally imposed labels (even well-intentioned ones) can misrepresent how systems actually behave in context.

If downstream toxicity detection or stance prediction systems trained on ArPoMeme outperform those trained on externally annotated Arabic datasets by more than 5 percentage points on held-out Facebook data, the self-identification approach validates itself. If performance gains disappear on other platforms or on synthetic test sets, the labels may be overfitted to Facebook's specific user base rather than capturing generalizable ideological patterns.