MetaOff-Meme: A Metaphor-Enriched Benchmark for Meme Offensiveness Detection

Meme has becomea popular medium of expression among internet users, utilizing its multimodal nature to vividly and effectively convey emotions and intentions. However, some memes carry offensive undertones. The abuse of these offensive memes not only causes harm to individuals, but also poses profound negative impacts on social stability, public health, and the healthy development of the online environment. Meanwhile, with the emergence of large models, there is growing attention on whether the content generated by these models aligns with human ethics and values. Since large models may inadvertently generate offensive content, detecting and preventing the potential misuse of offensive memes is of critical importance. However, to evade regulation by online platforms, offensive content in memes is often expressed and disseminated in more implicit ways. These implicit expressions frequently rely on extensive use of metaphors to convey offensiveness.
Current models rely solely on multimodal sentiment analysis methods to identify offensive sentiments in memes, while neglecting the complex metaphorical features inherent within memes, which significantly limits their performance. Therefore, accurately understanding the metaphorical features underlying memes is key to effective offensiveness detection. To improve the performance of offensiveness detection for memes enriched with metaphorical content, we construct MetaOff-Meme consisting of 3,033 image-text pairs with manual annotations.
(1) Metaphor Annotation. MetaOff-Meme defines the syntactic structure of metaphors as: 'A' is as 'relationship' as 'B' and use this structure to expand the metaphors and metaphorical relationships within memes. In this syntactic structure, A and B represent the source domain and target domain of the metaphor, respectively, while the relationship denotes the shared comparative attribute between them.
(2) Offensiveness Annotation. Offensive memes are often presented in the form of humor or self-deprecation, appearing harmless on the surface but potentially embedding biases, discrimination, or malicious intent. The annotation process for meme offensiveness is shown in Figure 2(b). We first determine whether a meme is offensive. Subsequently, we perform fine-grained annotations on offensive memes in terms of offensive topic, level, and scope.

During the collection of offensive memes, we retained non-offensive memes to ensure sample diversity. According to Figure 3(a), offensive memes account for 61% of the total dataset, while non-offensive memes make up the remaining 39%. Further metaphorical analysis reveals that 82% of the samples contain metaphors. Offensive memes with metaphors make up 53%, those without metaphors account for 29%, and non-metaphorical memes comprise only 18%. This distribution indicates that the use of metaphors is prevalent in offensive memes.
For offensive topics, as shown in Figure 3(b), gender-related topics account for the highest proportion, which may reflect the relatively high prevalence of such memes on social media platforms. For offensive level, as shown in Figure 3(c), the distribution of memes across different levels of offensiveness is relatively balanced, with non-offensive memes accounting for the largest proportion at 39%. Among the remaining categories, slight, moderate, and severe levels exhibit a balanced distribution, suggesting that social media users tend to employ memes with varying degrees of offensiveness to convey harmful content.
According to Figure 3(d), 73% of offensive memes are directed at specific groups, while the remaining 27% target individuals. This trend may be related to the formation of group dynamics and collective identities on social media platforms, as group-targeted memes are more likely to provoke broader discourse and interactions.

We propose four tasks based on MetaOff-Meme to evaluate the performance of meme offensiveness detection. The detailed descriptions of the tasks are as follows:
(1) Offensiveness Detection : a binary classification task to determine whether a meme is offensive.
(2) Offensive Topic Classification: a task aims to classify offensive memes into fine-grained categories, supporting a deeper understanding of their diversity.
(3) Offensive Level Evaluation: a task to evaluate the offensiveness level of memes, which helps measure their severity and more precisely assess their potential impact.
(4) Offensive Scope Analysis: a task to analyze the target scope of offensive memes, uncovering the groups or individuals targeted in their dissemination.
We select 5 SOTA open-source VLMs as experimental models to explore the current applications of multimodal large models in meme offensiveness detection. The models include Qwen-VL-Chat-7B, LLaVA-v1.5-7B, MiniGPT-v2-7B, InternVL2-8B, and MiniCPM-V2.6-7B.
We evaluate all data under a zero-shot setting. This setting enables the assessment of the generalization ability and cross-task performance of metaphors in meme offensiveness detection. To evaluate the performance of VLMs, we design three comparative experiments. The overall experimental process is shown in Figure 4. The evaluation benchmarks include the following:
(1) Baseline : We evaluate the VLMs using only the original meme features to assess the fundamental performance of VLMs in meme offensiveness detection.
(2) Metaphor Thought Chain (MCT) : In this setup, VLMs are required to first analyze metaphorical information within the meme before performing offensiveness detection. This experiment explores whether VLMs can correctly understand metaphorical expressions, and how this understanding influences offensiveness detection.
(3) Metaphor Information Augmentation (MIA) : Based on the original meme features, we further incorporate manually annotated metaphorical content to assist VLMs in offensiveness detection.
Through the above experiments, we not only evaluate the actual performance of current VLMs in meme offensiveness detection, but also analyze their ability to understand metaphors in memes, laying a foundation for deeper meme understanding by VLMs.

In the offensiveness detection task, MiniGPT-v2 achieves the best performance after introducing manually annotated metaphorical content. LLaVA-v1.5 shows a significant improvement under the MCT setting, primarily due to its tendency to adopt non-offensive descriptions when analyzing metaphors, which helps correct a large number of memes that were previously misclassified as offensive. Qwen-VL-Chat performs worse than the baseline in both MCT and MIA settings. We observe that Qwen-VL-Chat tends to classify most memes as offensive, and this bias is further amplified after metaphorical content is introduced. Overall, MIA improves their performance in offensiveness detection task.
In the other three tasks, all models perform better under the MIA setting compared to the baseline. However, their performance under the MCT setting is relatively unstable. Our experiments reveal that VLMs primarily focus on the visual content of memes when analyzing metaphors, failing to effectively incorporate textual information for metaphorical reasoning. This introduces noise into the detection process and consequently affects the recognition performance.
From the average scores of the overall results, we observe that manually annotated metaphorical content can effectively enhance model performance in meme offensiveness detection. However, due to the poor performance of Qwen-VL-Chat in the offensiveness detection task, the overall average improvement is relatively limited. Under the MCT setting, we find that most models fail to accurately understand the metaphorical expressions embedded in memes, resulting in incorrect metaphorical content that misleads the final prediction and degrades performance compared to the baseline. This indicates that current models lack the capability to comprehend metaphorical expressions.