CETA: Context-Enhanced and Target-Aware Hateful Meme Inference Method

Kaichun Wang, Junyu Lu, Bingjie Yu, Liang Yang, Hongfei Lin

Abstract: Hateful memes spread quickly online and harm society, necessitating effective detection methods. Detecting these memes is challenging due to the need for comprehensive reasoning. Though great efforts have been made, existing detection methods overlook the specific target of hateful meme, resulting in inadequate meme comprehension and hindering performance. In this paper, we propose Context-Enhanced and Target-Aware Hateful Meme Inference Method (CETA), which detects hateful memes while simultaneously pays attention to their targets. Specifically, CETA employs a prompting template to integrate contextual information and guide the model in reasoning the meme’s target and assessing its hatefulness based on the context. Experimental results show that our approach outperforms the state-of-the-art baselines on two hateful meme datasets, achieving up to 1.82% improvement in accuracy, significantly enhancing model inference and detection performance.

External IDs:dblp:conf/nlpcc/WangLYYL24