ScholarWorks@성균관대학교

초록

The proliferation of social media has facilitated the spread of negative memes, yet research on harmful visual content remains limited, particularly for Korean memes which often lack text. This study introduces a vision-centric approach to detect negative content in Korean memes. To support this research, we created the Korean Meme Dataset (KMD) with detailed annotations. We implemented a vision-centric model and bench-marked its performance against text-only and zero-shot multi-modal baselines. In the binary classification task, a text-based model (KoELECTRA) achieved the highest accuracy of 0.7427. However, in the more complex multi-label task, its performance degraded significantly to a Macro F1-score of 0.3099. In contrast, our proposed vision-centric model, while showing a lower binary classification accuracy of 0.6462, achieved superior performance in the multi-label task with a Macro F1-score of 0.4181 and a weighted F1-score of 0.5385, outperforming all baselines on key metrics. The vision-centric model also recorded the highest precision (0.5169) and recall (0.5256) in the multi-label task. This result provides strong evidence that a vision-centric approach is crucial for classifying memes that heavily rely on visual context. A detailed category-wise analysis revealed that the effectiveness of each model depends on the meme's specific characteristics. This research addresses the key limitation of existing text-focused methods by providing a foundational study on Korean meme detection. It also paves the way for future multimodal approaches that can effectively interpret complex visual cues and integrate them with text analysis. The dataset and model implementation are available at https://github.com/dxlabskkU/KMD.

키워드