Towards Detecting Chinese Harmful Memes with Fine-Grained Explanatory Augmentation

Abstract

The rapid proliferation of social media and generative large language models has increased multimodal harmful content, making harmful meme detection and explanation generation crucial for content moderation. In Chinese social media, meme harmfulness relies on implicit visual–textual interactions in cultural contexts, but existing research lacks a comprehensive understanding of such cultural specificity. This neglect of the social background knowledge and metaphorical expressions inherent in memes results in limited detection performance. To address this challenge, we propose a novel fine-grained explanation-enhanced Chinese harmful meme detection framework (FG-E2HMD), a framework using Multimodal Large Language Models (MLLMs) with a culturally aware explanation generation module to produce structured explanations, which integrate with multimodal features for decision-making. Comprehensive quantitative experiments and qualitative analyses were conducted on ToxiCN MM, the first large-scale dataset dedicated to Chinese harmful meme detection. The experimental results reveal that existing methods still have significant limitations in detecting Chinese harmful memes. Concurrently, our framework improves detection accuracy and decision transparency by incorporating explicit Chinese cultural background knowledge, paving the way for more intelligent, culturally adaptive content moderation systems.

1. Introduction

In the era of digital information, memes have emerged as a pivotal component of internet culture, serving as a multimodal information carrier that integrates images and text. Owing to their humorous and highly communicative characteristics, memes significantly shape public communication and influence the public opinion ecosystem. Nevertheless, the ubiquity of memes and their substantial impact render them a double-edged sword. Beyond fulfilling benign entertainment and social functions, memes are increasingly employed as covert instruments for disseminating hate speech, promoting extremist ideologies, spreading misinformation, and engaging in personal attacks [1,2]. The vast quantity of memes on the internet and social media platforms presents a significant economic burden when manually reviewed. Consequently, there is an urgent need to develop multimodal content understanding technology for the automatic identification of harmful memes. However, the implicit and satirical nature of these harmful memes poses a serious challenge to automatic content moderation systems. In 2020, Facebook hosted the Multimodal Hate Meme Detection competition, marking a significant milestone in research on toxic meme detection [2]. Numerous datasets have been proposed by researchers, and corresponding recognition methods have been developed, yielding favorable recognition results [3,4,5,6]. However, these studies have primarily centered on English memes. Despite some advancements in this area, the body of research concerning Chinese harmful memes remains notably limited.

On the one hand, the harmfulness of Chinese harmful memes is often not reflected in literal meanings or individual image elements but is deeply rooted in complex cultural contexts. Harmful memes are typically characterized as multimodal units, combining image and text, that inflict harm on individuals, organizations, communities, or social groups by targeting specific social entities. Moreover, memes on Chinese platforms often encompass general offenses, sexual innuendos, or promote a disheartening culture. Even though these memes sometimes lack specific targets, they still possess potential toxicity. Subtly promoting negative values, they can contribute to severe consequences such as violent acts and sexual harassment [7]. The creators or spreaders of these memes might employ homophonic jokes, historical references, social events, and internet subcultures as a means to obscure their true intentions. Such intentions necessitate a nuanced understanding of culture to be unearthed [7,8,9]. For instance, the act of juxtaposing a commonplace image with an ostensibly unrelated ancient poem or internet catchphrase to indirectly criticize a specific group or individual would likely go unnoticed by an automatic moderation system devoid of the requisite cultural knowledge [10,11,12,13].

On the other hand, even the most powerful large multimodal models, including Qwen2.5-VL and GPT4o, possess significant general world knowledge. However, without targeted guidance and fine-tuning, these models frequently miss the intricate nuances of specific cultural contexts, particularly within the Chinese internet. As noted by Yang et al. [7], a deep understanding of Chinese multimodal content requires culturally embedded visual literacy, a crucial element that these general models often lack.

The existing research on the detection and interpretive generation of harmful memes primarily focuses on the targets of the memes and the corresponding harmful content, overlooking the understanding of the metaphorical expressions of harmful memes and the related cultural background knowledge. Consequently, this creates a significant barrier for individuals lacking the requisite background knowledge to comprehend their implicit meaning [14,15,16,17]. Scott et al. [8] demonstrate that metaphor is one of the main expressions of harmful memes. Currently, the vast majority of content moderation systems operate like a black box, only outputting a label of “harmful” or “harmless” without providing explanations for their decisions. This opacity not only undermines the trustworthiness of moderation outcomes but also prevents platform administrators and users from comprehending the essence of harmful content, undoubtedly hindering the construction of a healthy online community ecosystem [18,19,20]. Recent research has begun to address this issue, such as Lin et al. [9] proposing to generate explanations through debates from both harmful and harmless perspectives using a large multimodal model. However, it has not yet systematically integrated culturally specific knowledge into the explanation process. Therefore, the identification of metaphorical expressions in harmful memes, along with the associated cultural and social background knowledge, is essential for generating comprehensive, accurate, and easily understandable explanations [21,22,23,24]. As illustrated in Figure 1a, from the text within the image, including the phrases “generation gap”, “Palace Jade Liquid Wine”, “180 yuan per cup”, and “post-00s”—along with the three wolves in the image displaying divergent facial expressions—it is impossible to deduce the cause of the generation gap. The understanding of the meme’s profound meaning becomes clear only when one integrates the cultural context that the phrase “ Palace Jade Liquid Wine, 180 yuan per cup” originates from the sketch “Strange Encounters at Work” performed at the 1996 CCTV Spring Festival Gala. The image features two wolves on either side with open mouths, suggesting joyful communication, whereas the central wolf, labeled “post-00s”, exhibits a confused demeanor [25,26,27]. Given that individuals born post-2000 most likely have not watched “Strange Encounters at Work”, this meme humorously underscores the linguistic and cultural disparities between generations. As illustrated in Figure 1b, a pig is playing with a cabbage, and the text shows “I’ve settled you down”. Without considering the Chinese cultural background knowledge, this meme might be judged as harmless. However, when viewed within the context of Chinese cultural background knowledge, it becomes apparent that this is a meme with gender discrimination against men. Because this image expresses a Chinese proverb “All cabbage had been twiddled by pig”, which is a satire on some male–female relationships where the male’s overall quality is perceived to be inferior to the female’s. As shown in Figure 1c, the image shows five green puppy plush toys, and the text shows “which vegetable dog are you”, which would be judged as harmless without Chinese cultural background knowledge considered, while with Chinese cultural background considered, “vegetable dog” is a Chinese ironic phrase to mock some people who are incompetent. Thus, recognizing the metaphorical techniques in memes and incorporating pertinent cultural knowledge is imperative for assessing the potential harm of Chinese memes.

In order to address the aforementioned challenges, we focus on Chinese harmful meme detection, aiming to build an automatic moderation system that is both accurate and interpretable. Our main contributions are as follows:

• We propose a novel explanation-enhanced detection framework, which is designed around a two-stage “first explanation, then judgment” mechanism. Initially, the framework is directed to generate a structured reasoning process that is both human-readable and of high quality (i.e., an explanation). Subsequently, this explanation is leveraged to enhance the final classification decision. This methodology significantly bolsters the robustness and accuracy of meme detection.
• In this study, we take the lead in systematically integrating Chinese cultural background knowledge into our harmful meme detection model. This integration is both explicit and structural, ensuring a comprehensive understanding and generation of Chinese cultural background knowledge within the model. We employ a meticulously crafted “culture-aware prompt” to effectively activate and guide the MLLM to utilize the extensive knowledge acquired during its pre-training phase. This approach enables the MLLM to decode the profound cultural connotations embedded within memes.
• We conducted extensive experiments on ToxiCN MM [10], the largest dataset of Chinese harmful memes with fine-grained annotations and compared the performance with various advanced methods. The experimental results fully demonstrated the effectiveness of our proposed method.

Overall, this work presents an MLLM-based fine-grained explanation-enhanced framework for Chinese harmful meme detection. Different from previous work, our proposed method prompts the model to output a structured description of “cultural background-metaphor-harm-target of attack” based on Chinese context through carefully designed prompt templates. By leveraging the powerful image understanding and text generation capabilities of Multimodal Large Language Models, the generated descriptions serve as strong auxiliary information, combined with multimodal image–text alignment features to enhance the overall performance of the model. The proposed method is rigorously evaluated on the ToxiCN MM test, set against various state-of-the-art models to verify its effectiveness. The remainder of this paper is structured as follows. Section 2 reviews the related works. Section 3 outlines our proposed method, including three key components and culture-aware prompt templates. Section 4 presents and analyzes the experimental results. Section 5 discusses several cases and corresponding limitations and considerations. Finally, Section 6 concludes this work and highlights future directions.

2. Related Works

2.1. Multimodal Harmful Meme Detection

Harmful meme detection, as an important branch of multimodal content safety, has garnered significant attention in recent years [28,29,30]. Initial research efforts concentrated on the development of sophisticated fusion modules to integrate image and text features. With the success of the Transformer architecture [12], subsequent studies have employed Visual Transformer (ViT) [13] for visual understanding and BERT [11] for text understanding. These studies then leveraged the cross-attention mechanism to capture cross-modal correlations. For instance, the HMGUARD [1] framework enhances detection capabilities by analyzing propaganda techniques and visual arts. Ji et al. [14] argue that mapping the extracted text and image features to a unified feature space may introduce significant semantic discrepancies, potentially compromising the final performance of the model. Therefore, they propose to first generate descriptive text and key-attribute information of image data, thus converting multimodal data into unimodal pure text data. Following this conversion, a prompt-based pre-trained language model processes the textual data, facilitating the detection of multimodal harmful memes [31,32,33,34]. However, the majority of these works have been trained and validated on English datasets, and their applicability to Chinese and other linguistic and cultural contexts still requires verification. The ToxiCN MM dataset is the first fine-grained annotated Chinese harmful meme dataset, which primarily encompasses the following dimensions: harmfulness assessment, a binary classification to determine whether the meme is harmful; identification of harmful types, a multi-classification involving targeted harmful, sexual innuendo, general offense, and dispirited culture; and modality combination feature analysis, which examines whether the harmful content is conveyed through text alone, images alone, or a combination of both. This framework employs a straightforward yet effective visual prompt tuning paradigm to optimize the mutual information between images and texts. It achieves this by incorporating a cross-modal contrastive alignment objective. Although ToxiCN MM and CHMEMES are both Chinese harmful meme datasets, they have not explored the research of generating explanations related to harmfulness combined with the Chinese cultural background knowledge.

2.2. Explainable Multimodal Learning

Memes frequently employ a plethora of rhetorical devices, such as metaphors, which often elude the current image caption generation methods designed for precise depiction of image scenes, due to their inability to interpret the inherent deep meanings [35,36,37]. As research on multimodal memes progressively deepens, researchers have further proposed the field of meme interpretation generation. Hee et al. [16] have developed HatReD, a human-curated dataset for generating explanations of multimodal hateful memes, in an endeavor to augment comprehension of the implications embedded within such memes. Their research has revealed a pronounced disparity between the performance of current State-of-The-Art (SOTA) multimodal pre-trained models and that of human experts when encountering memes from novel domains. This study constitutes a fundamental basis for future investigations into the field of harmful meme explanation generation. Hwang et al. [17] constructed a dataset called MemeCap. Each meme in this dataset was annotated with text information, image description, metaphorical expression, and human-generated meme interpretation. Lin et al. [9] simulated a multimodal debate between pro and con perspectives, enabling their model to generate arguments from both harmful and non-harmful viewpoints, and make a final judgment based on these arguments [38,39,40,41]. Our work takes a step forward on the basis of this idea, not only pursuing the generation of explanations but also emphasizing the cultural depth and structured nature of the content and using it as a key information source to improve decision-making accuracy.

2.3. Chinese Content Security and Cultural Particularity

Chinese information processing is challenging due to its unique linguistic and cultural characteristics. For advancing research in this area, a high-quality dataset is essential. The ToxiCN MM dataset represents a groundbreaking effort. As the first large-scale, fine-grained annotated Chinese harmful meme dataset, it provides a valuable cornerstone for researching multimodal harmful meme detection in the Chinese domain. Yang et al. [7] explicitly proposed the concept of culture-embedded visual literacy in their work and systematically explored various challenges in Chinese multimodal content detection, which emphasized the necessity of incorporating cultural knowledge. Zhu et al. [18] proposed the MiniGPT-4 model, which consists of a frozen ViT and Q-Former vision encoder, a linear projection layer, and frozen Vicuna. The MiniGPT-4 only needs to train a linear projection layer to align the visual features and Vicuna. Zhang et al. [19] proposed a multimodal model named LLaMA-Adapter, which adaptively injects image features into LLaMA based on a gated zero-initialized attention mechanism while preserving the knowledge obtained from its pre-training stage, achieving good results in downstream tasks. Gao et al. [20] introduced LLaMA-Adapter V2, an enhanced version of LLaMA-Adapter. This model is fine-tuned using highly parameter-efficient visual instructions to achieve superior recognition performance in downstream tasks. Remarkably, it accomplishes this with fewer fine-tuning parameters compared with its predecessor, demonstrating improved efficiency and effectiveness. Subsequently, Liu et al. [21] introduced the LLaVA model, which employs GPT-4 to generate instruction fine-tuning training data. During the training phase, visual tokens are incorporated for instruction fine-tuning, utilizing GPT-4’s responses to align with the user’s intent. Ye et al. [22] unveiled a comprehensive multimodal model, mPLUG-Owl, exhibiting superior visual-language comprehension capabilities. Notably, it can interpret visual element-centric jokes. In our study, the proposed FG-E2HMD significantly enhances the efficacy of large multimodal models in detecting Chinese harmful memes. By integrating a culturally aware prompt, we further activate the extensive knowledge of MLLM acquired during pre-training, enabling a more accurate decoding of meme metaphors and profound cultural nuances.

3. Materials and Methods

The core idea of our method is to imitate the cognitive process of human experts in reviewing harmful content. When an expert is faced with a suspicious meme, they do not immediately make an intuitive judgment but go through an analytical process of decomposition, reasoning, and integration: first, observe the images and text; then, mobilize their background knowledge (including culture, history, and current events, etc.) for analysis, thereby forming a complete chain-of-explanation; and finally, based on this logically rigorous chain-of-explanation, give a final professional judgment. Our framework is designed to mimic this intricate procedure, encapsulated by the core principle: “Explain first, then predict”. Instead of directly producing the final output from the raw input, we introduce an explicit, structured explanation generation step as an intermediate bridge. This explanation facilitates human comprehension and operates as a refined and amplified signal, steering the model towards more dependable decisions. To this end, we design an overall model architecture as shown in Figure 2, comprising three pivotal modules: (1) a multimodal feature alignment module, (2) a culture-aware explanation generation module, and (3) an explanation-enhanced decision-making module.

3.1. Multimodal Feature Alignment

This module is designed to extract both intermodal and intramodal information from the meme, converting raw multimodal data into deep semantic features. Drawing inspiration from Hee et al. [6], we employ a leading pre-trained model to secure high-quality representations. Given that our focus is on detecting Chinese harmful memes, we utilize the Qwen2.5-VL as our vision–language encoder. Specifically, we feed the meme image, denoted as I, and its text T into the frozen pre-trained multimodal large model to generate a unified multimodal representation $F_{mm}$:

$$F_{mm}={\mathrm{E}\mathrm{n}\mathrm{c}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}}_{MLLM}\left(\left[I,T\right]\right)$$

$F_{mm}$ will serve as the input basis for all subsequent modules.

3.2. Culture-Aware Explanation Generation

This is the core component of our proposed framework, designed to accept the multimodal representation $F_{mm}$ and a piece of culture-aware prompt template, subsequently generating a high-quality, structured, and culturally enriched explanatory text $T_{exp}$. We achieve this by fine-tuning a powerful large multimodal model (Qwen2.5-VL in our method) with instructions. Inspired by [10], we meticulously crafted a series of refined instruction templates that employed a chain-of-thought-like approach, directing the MLLMs to execute systematic, step-by-step reasoning and generate outputs adhering to a predefined structure [23]. This not only enhances the structure of the output but also guarantees that crucial aspects of the explanations are not overlooked. The template example is shown in

Table 1. Example of an instruction template.

Prompt Template:
You are a multimodal content security expert with a deep understanding of Chinese culture and the online ecosystem. Please analyze the given meme and provide a structured analysis based on the following four aspects:   Cultural Background Analysis: Does this image or text involve specific cultural elements of China (such as historical events, characters, internet memes, proverbs, social phenomena, etc.)? Please specify its origin and meaning.   Metaphor and Subtext: How are the image and text content combined, and what are they implying, suggesting, or insinuating? What is the underlying subtext or metaphor?   Harm Assessment: Based on the above analysis, please determine whether the meme is harmful or harmless.   Target of Attack: If the meme is harmful, please indicate which specific individual, group, or organization is being targeted, attacked, satirized, or smeared.   Summarize the above generated content into 5 sentences, with no more than 50 words in each sentence.

.

By fine-tuning the model through structured and well-designed instructional prompts, the model learns to systematically decompose and analyze a meme. This approach provides a more detailed understanding than a general impression, which is particularly beneficial for decision-makers seeking a concrete basis for judgment.

3.3. Explanation-Enhanced Decisions

With the structured explanation text $T_{exp}$ obtained, this module is responsible for utilizing it to assist and enhance the final classification decision. As shown in the overall architecture of the model in Figure 2, the generated explanatory text $T_{exp}$ is first input into the text encoder RoBERTa to obtain its deep semantic representation $F_{exp}$. $F_{exp}$ is the essence of the model’s inference and refinement of the original information, focusing on the key harmful information underlying the meme. We integrate the original multimodal features $F_{mm}$ with the generated explanatory features $F_{exp}$ using a simple vector concatenation to obtain the final decision-making features $F_{final}=\left[F_{mm}; F_{exp}\right]$, in which $F_{final}\in {\mathit{ℝ}}^{ h \mathit{✕} 2 }$ and $h$ represents the dimension of the hidden layer of the model. Finally, the fused $F_{final}$ is fed into a simple trainable classifier, which first performs a linear transformation on $F_{final}$, and then generates detection probability through the softmax function to judge whether the meme is harmful. Since $F_{final}$ contains both original information and explanatory information after logical reasoning, the decision-making process is thus more robust, transparent, and accurate.

4. Experimental Results and Analysis

4.1. Datasets and Baselines

All the experiments of this work are conducted on ToxiCN MM, which is the first large-scale dataset for Chinese harmful meme detection, containing 12,000 samples with fine-grained annotations of meme types. We follow its official data split, which divides the dataset into training set (80%), validation set (10%), and test set (10%), and mainly conduct a binary classification task of harmful or non-harmful. In addition to the FG-E2HMD proposed in this paper, we have evaluated the performance of various baseline models. These baseline models include both unimodal and multimodal types. For unimodal model baselines, we select DeepSeek-V3, RoBERTa, and GPT4 as pure text modality baselines to directly classify the text of the memes and Image-Region, ResNet, and ViT as pure image modality baselines to directly classify the extracted image features of the memes. For multimodal model baselines, we choose GPT4, CLIP+MKE, Visual BERT COCO, Hate-CLIPper, MOMENTA, PromptHate, zero-shot Qwen2.5-VL (here Qwen2.5-VL is not fine-tuned but is directly asked the question “Is this meme toxic?” and judged based on its response), and the SOTA model related to explanation generation, debate-based model [9], then fine-tune it on ToxiCN MM. These are the main comparison objects of our method. Furthermore, for the baseline models we compared, if there is the same baseline as that in ToxiCN MM, the results in ToxiCN MM are directly adopted.

4.2. Evaluation Metrics

We employ precision, recall, macro F1-score, and the F1 score of toxic memes as our evaluation metrics. Since our work mainly targets the detection of Chinese harmful memes, the MLLM in the overall framework is the frozen Qwen2.5-VL, which primarily functions as the encoder in the visual-language alignment module and the explanation generator in the culture-aware explanation generation module. For the encoder within the explanation-enhanced decision module, we opt for RoBERTa, given its superior encoding capabilities for Chinese texts. Notably, the specific version we use is chinese-roberta-wwm-ext-base.

4.3. Parameter Setting

The experimental parameters are set as follows: the batch size is 32, the learning rate is 1 × 10⁻⁵, and the number of training epochs is 10. We use AdamW as our optimization strategy, with the decay rate β₁ and β₂ set to 0.9 and 0.95, respectively. The weight decay is set to 0.02 to perform regularization control on the model parameters. The Qwen2.5-VL used is the 7B version. All experiments were conducted on two NVIDIA RTX A6000 GPUs. For details, see Appendix A.1 and

Table A2. Hyperparameters for experiments.

Parameter Category	Setting
Optimizer	AdamW
Batch Size	32
Learning Rate	1 × 10−5
Training Epochs	10
β1	0.9
β2	0.95
Weight Decay	0.02
Model	Qwen2.5-VL-7B
	RoBERTa (chinese-roberta-wwm-ext-base)
GPUs	2 × NVIDIA RTX A6000

.

4.4. Quantitative Analysis

We evaluated our proposed FG-E2HMD and all baseline models on the ToxiCN MM test set, and the detailed results are shown in

Table 2. Performance comparison of different models on the ToxiCN MM test set. The best and second-best results of each metric are highlighted in bold and underlined, respectively.

Model		P	R	F1	F1harm
Text-only	DeepSeek-V3	74.63	75.36	74.99	58.75
	RoBERTa	75.52	77.54	76.36	66.48
	GPT4	74.52	65.59	68.01	51.78
Image-only	Image-Region	65.96	66.06	66.01	52.95
	ResNet	66.61	66.92	66.76	53.76
	ViT	68.97	68.61	68.78	57.24
Multimodal	GPT4	74.67	68.64	70.11	55.77
	CLIP+MKE	79.76	80.79	80.23	72.35
	VisualBERT COCO	72.21	69.36	70.76	57.21
	Hate-CLIPper	73.56	68.52	70.95	60.36
	MOMENTA	74.05	69.88	71.90	62.45
	PromptHate	75.83	72.36	74.05	63.15
	Qwen2.5-VL(zero-shot)	70.25	73.36	71.77	58.76
	Debate-based model	79.93	81.16	79.91	70.53
	FG-E2HMD	83.53	83.26	83.39	74.43

. As can be seen from the experimental results, multimodal models generally outperform unimodal models, confirming the efficacy of integrating image and text features for identifying harmful memes. Although the zero-shot performance of the general multimodal large model Qwen2.5-VL is commendable, it falls short of models that have been specifically fine-tuned, underscoring the importance of domain adaptation. Notably, the two interpretation-based methods, the debate-based model and our proposed FG-E2HMD, markedly outperformed other baselines except for CLIP+MKE. This suggests that generating detailed explanations of rich cultural backgrounds enhances model performance in detecting Chinese harmful memes. Most importantly, our proposed FG-E2HMD has reached the SOTA performance in all metrics, with an F1 score of 83.39, which has increased by 3.94% compared with the previous best CLIP+MKE model. In addition, the performance of our proposed FG-E2HMD is also significantly superior to another interpretation-based method, the debate-based model. This fully demonstrates the superiority of our culture-aware and explanation-enhancement mechanisms over general explanation generation techniques.

4.5. Ablation Study

To verify the effectiveness and necessity of each innovative module in FG-E2HMD, we have conducted a series of ablation experiments, the results of which are shown in Table 3. The results of ablation experiments indicate that

• “Structured output” provides stable gains. Removing the requirement for structured output, there is a slight decline in model performance (2.54%/1.81%). This indicates that forcing the model to think and express according to the logical chain of “cultural background-metaphor-harmfulness-target of attack” helps form a more organized and focused reasoning process, thereby bringing about stable performance improvements.
• “Cultural awareness” is the core. After removing the specially designed culture-aware prompt and switching to a generic prompt, there was a significant drop in model performance (3.39%/2.78%). This indicates that explicitly incorporating the cultural background information that a meme might imply into the model’s thinking and analysis process greatly benefits the performance of the model in the task of detecting harmful memes.
• “Explanation-enhanced decision-making” is of paramount importance. Removing the explanation-enhanced decision-making module results in the largest performance drop (9.65%/9.23%). This indicates that simply generating an explanation is insufficient. Most importantly, there is a need to effectively reintegrate high-quality explanation features into the decision-making process in order to optimize the model’s final performance.
• Structured representation and the culture-aware prompt have complementary gains. Removing the culture-aware prompt alone results in a 3.39 decrease in F1; when simultaneously removing the structured representation (line 6), the F1 further decreases by 10.3, a decrease significantly greater than the sum of the two individual decreases (3.39 + 2.54 = 5.93). This “superposition” decrease indicates that the structured output provides more controllable semantic slots for cultural prompts, while cultural prompts in turn enhance the adaptability of structured representation to the Chinese language context, forming a positive coupling between the two.
• Complete framework achieves robust detection through “Divide and Conquer”. When all the three modules are removed (last row), the model degrades into a normal black-box classifier, and F1 drops to 72.33, 13.18% points lesser than the complete framework. This verifies the design hypothesis of FG-E2HMD:
• Structured representation is responsible for the controllable output at the “form” level.
• The culture-aware prompt is responsible for the alignment of cultural context at the “content” level.
• Explanation enhancement is responsible for the “logic” layer of verifiable reasoning.
• After superimposing the three models, the model achieves the current optimal 83.39 F1 in the task of Chinese harmful meme detection and has significant fault tolerance to the local failure of individual modules.

Table 3. Evaluation results of the ablation experiments. The bold indicates optimal results, and the numbers before and after the “/” in the performance drop column correspond to the degree of decline in F1 and F1_harm, respectively.

Method	F1	F1harm	Performance Drop	Interpretation
Full FG-E2HMD	83.39	74.43	—	A complete framework containing all modules.
W/O structured representation	81.27	73.08	2.54%/1.81%	Do not force structured output and generate free-form text explanations.
W/O culture-aware prompt	80.56	72.36	3.39%/2.78%	Use the general prompt (directly prompt the model to “explain why it is harmful”).
W/O explanation enhancement	75.34	67.56	9.65%/9.23%	Remove the “explanation-enhancement decision-making module” and classify directly using $F_{mm}$.
W/O culture-aware prompt and explanation enhancement	77.65	68.93	6.88%/7.39%	Jointly remove the culture-aware prompt and explanation enhancement.
W/O structured representation and culture-aware prompt	74.82	67.02	10.3%/9.9%	Jointly remove the culture-aware prompt and structured representation.
W/O explanation enhancement and structured representation	76.03	68.11	8.83%/8.50%	Jointly remove explanation enhancement and structured reprensentation.
W/O all three modules	72.33	64.81	13.18%/12.92%	Remove all three modules.

Table 3. Evaluation results of the ablation experiments. The bold indicates optimal results, and the numbers before and after the “/” in the performance drop column correspond to the degree of decline in F1 and F1_harm, respectively.

Method	F1	F1harm	Performance Drop	Interpretation
Full FG-E2HMD	83.39	74.43	—	A complete framework containing all modules.
W/O structured representation	81.27	73.08	2.54%/1.81%	Do not force structured output and generate free-form text explanations.
W/O culture-aware prompt	80.56	72.36	3.39%/2.78%	Use the general prompt (directly prompt the model to “explain why it is harmful”).
W/O explanation enhancement	75.34	67.56	9.65%/9.23%	Remove the “explanation-enhancement decision-making module” and classify directly using $F_{mm}$.
W/O culture-aware prompt and explanation enhancement	77.65	68.93	6.88%/7.39%	Jointly remove the culture-aware prompt and explanation enhancement.
W/O structured representation and culture-aware prompt	74.82	67.02	10.3%/9.9%	Jointly remove the culture-aware prompt and structured representation.
W/O explanation enhancement and structured representation	76.03	68.11	8.83%/8.50%	Jointly remove explanation enhancement and structured reprensentation.
W/O all three modules	72.33	64.81	13.18%/12.92%	Remove all three modules.

5. Discussion

To further illustrate the basic principle of FG-E2HMD, as shown in

Table 4. Explanation of the case study. The green and red highlighted parts, respectively, represent the correct and incorrect information related to meme hazard detection. In the actual implementation process, all descriptions are in Chinese. ✓ indicates the model makes a correct prediction, while × indicates the model makes a wrong prediction.

Harmful Meme	(a) You Skr Ugly Thing	(b) Scientists are Working to Cure Them	(c) Girls Wash Clothes: Underwear, Dark Clothes, Light Clothes. Boys Wash Clothes.
Ground Truth	General Offensive, Harmful Text	Targeted Harmful, Text–Image Fusion	Targeted Harmful, Text–Image Fusion
Text description	The sentence translates to “You ugly thing”, with “skr” being an internet slang term, overall expressing insult and dissatisfaction towards someone.	Scientists are working to cure diseases that affect women.	This is a reminder about laundry classification, suggesting that girls and boys should wash their clothes separately, and underwear, dark, and light clothes should also be washed separately.
Image description	An exaggeratedly expressive and emotionally agitated man, with his fingers pointing forward, seemed to be shouting.	Two girls in wedding dresses are kissing each other.	Two comic strips, one showing a woman tidying up clothes for washing, the other showing a man casually throwing clothes into the wash, with a cat sitting below.
Explanation from FG-E2HMD	[Cultural Background Analysis] The picture text combines the “skr” meme in Chinese internet culture, carrying a mocking and sarcastic meaning. [Metaphor and Subtext] Expressions paired with text implied a disparaging attitude, denigrating someone’s appearance or behavior. [Harmfulness Determination] The meme is highly offensive and potentially harmful through its abusive language. [Target of Attack] Its main target is individuals referred to as “ugly things”, the specifics of which are unclear. [Conclusion] Based on a comprehensive assessment, this meme is considered harmful content and should be disseminated with caution.	[Cultural Background Analysis] The image combines traditional wedding elements, and the text mentions “scientists are working hard to cure them,” which implies a humorous interpretation of the phenomenon of same-sex weddings. [Metaphor and Subtext] The stereotype and bias towards sexual minorities are metaphorically represented through the kissing scene and the words “scientists are working hard to cure them”. [Harmfulness Determination] The content is clearly insulting and offensive and belongs to harmful content. [Target of Attack] The main attack target is the sexual minority group (LGBTQ + group), which strengthens discrimination against them. [Conclusion] This meme belongs to harmful content. It utilizes stereotypes of sexual minorities, easily inciting insults and harm against the LGBTQ+ community, which is not conducive to creating an inclusive social environment.	[Cultural Background Analysis] The image utilizes gender stereotypes in Chinese internet culture, emphasizing the meticulousness of girls and the roughness of boys. [Metaphor and Subtext] The comparison of washing clothes is a metaphor for the traditional social perception of the differences in behavior between men and women. [Harmfulness etermination] The content is inoffensive and is meant primarily to entertain the audience and evoke resonance. [Target of Attack] No attacks or insults were made against specific groups. [Conclusion] It is a harmless product of online culture and is suitable for easy sharing.
Prediction	GPT4: ✓ Qwen2.5-VL(zero-shot): ✓ Debate-based Model: ✓ FG-E2HMD: ✓	GPT4: × Qwen2.5-VL(zero-shot): × Debate-based Model: × FG-E2HMD: ✓	GPT4: × Qwen2.5-VL(zero-shot): × Debate-based Model: × FG-E2HMD: ×

, we provide three cases for study. From the textual caption and the image caption information of case (a), it can be seen that both our work and the compared baselines have captured the keyword “ugly thing” and simultaneously captured the exaggerated expression and emotionally excited demeanor of the man in the picture; thus, both have correctly classified case (a). However, the explanation generated by our model interprets “skr” as mockery and sarcasm. But from the perspective of Chinese internet culture, the popularity of “skr” originates from a Chinese rap show, which means to praise the contestant, generally referring to being very good. This may be due to the hallucination of the underlying MLLM. The original meaning of “skr” in case (a) should be a homophonic pun in Chinese internet culture, pronounced similar to “shi ge”, which means “is a”; thus, incorporating the text into the meme, it forms a complete sentence “You are an ugly thing”. This is the true understanding of the deep meaning of this meme.

5.1. Case Study

To gain more insights on the challenging issues related to Chinese harmful meme detection, we manually examine most of the misclassified samples of the compared baselines and find that the errors are mainly of two types:

• The model may occasionally misclassify a meme due to the benign information it contains.
• The model frequently demonstrates a deficiency in comprehending the intricate cultural context inherent in Chinese memes, leading to erroneous judgments.

Error classification category I: As illustrated in case (b), when the model is provided with text or image information, both the compared baselines and our proposed model judge the input information as harmless. However, when the model is provided with both text and image information simultaneously, only our proposed model accurately determines that the meme is harmful. This is because our proposed model recognizes that the image combines wedding elements from traditional culture. In the image, two women are kissing, and our model integrates Chinese cultural background knowledge that same-sex marriage is not permitted and, combined with the text part “scientists are working hard to cure them”, metaphorically implies stereotypes or prejudice against the LGBTQ+ community. The underlying implication is that homosexuality is perceived as an abnormality that needs to be “cured”, which is ironic and can potentially provoke insults and harm to sexual minorities. This is precisely the benefit brought by the “structured output” which leads to a more organized and focused reasoning process, as well as the “cultural-aware prompt” which explicitly incorporates the cultural background information behind the meme into the process of assessing their harmfulness, further demonstrating the effectiveness of our proposed methods.

Error classification category II: As illustrated in case (c), all models, including our proposed one and all compared baselines, have given incorrect classification results. The model represents the information within the image through two illustrative scenes: one depicting a woman washing clothes after sorting them, and the other portraying a man washing clothes without prior classification. All of these are positive information. The text in the picture describes “This is a reminder about sorting laundry. It is recommended that girls and boys wash their clothes separately. Underwear, dark, and light clothes should be washed separately”, which is also benign information. While our proposed model incorporates cultural background knowledge to examine the use of gender stereotypes in Chinese internet culture—notably the portrayal of girls as meticulous and boys as rough— it still assesses the meme as benign. The rationale provided is that the content is harmless, primarily designed to entertain and evoke resonance with the audience. This discrepancy might arise from the fact that both the compared baselines and our proposed model were subjected to a considerable number of prompts and amount of training data underscoring gender equality during pre-training, potentially leading to incorrect assessments. Nevertheless, when the meme is evaluated by human experts, it becomes evident that the meme targets the stereotype of gender discrimination. Consequently, it is imperative to investigate more effective post-training techniques and alignment methods with human values to balance the model’s performance and the challenge of over-emphasis on gender equality.

The aforementioned case study further demonstrates that Chinese harmful meme detection remains a challenging multimodal semantic understanding task for existing LLMs and MLLMs. Beyond the comprehensive integration of visual and textual information, effectively and accurately incorporating Chinese cultural context knowledge continues to pose a critical challenge in this domain.

5.2. Limitations and Ethical Considerations

5.2.1. Dependence on Underlying MLLM Capabilities

Our proposed framework is constrained by the MLLM acting as the “explanation generator”. If the base model itself lacks certain niche or lesser-known knowledge, it will be unable to generate high-quality explanations, and in severe cases, might produce hallucinations that are unrelated to the meme content, which could potentially exacerbate biases or stereotypes against certain social groups. Therefore, more robust methods like retrieval-augmented generation (RAG) can be considered to ensure that the generated explanations are more relevant and of higher quality.

5.2.2. Adaptation to New Memes

The generation and dissemination of internet memes are proliferating at an unprecedented pace, with novel slang terms, trending events, and expressive formats emerging continuously on a daily basis. This phenomenon may lead to a “knowledge lag” issue, wherein the model may fail to remain fully updated with newly emerged, previously unseen memes that were absent from its pre-training dataset. Furthermore, despite advancements in automated harmful meme detection, inherent biases may still persist within the model’s operational framework. Consequently, ongoing evaluation is imperative to ensure that our system maintains optimal performance and ethical integrity in real-world applications.

5.2.3. Potential Misuse Risks

A model capable of generating deep rationales can, in theory, also be potentially leveraged by malicious users to craft more convincing, harder-to-detect harmful content. However, the benefits brought by deep rationales far outweigh the possible risks. By providing deep rationales to content moderators, they can be helped to more effectively identify and flag potentially harmful content and thus foster a more harmonious, inclusive, and free internet environment.

6. Conclusions

This study systematically addresses the dual challenges of cultural comprehension disparities and interpretability deficiencies in Chinese harmful meme detection by proposing an innovative explanation-enhanced detection framework leveraging large multimodal models. Rather than treating the model as an end-to-end black box, we innovatively introduce a culture-aware explanation generation module that strategically guides the model to adopt an “explain-before-judging” paradigm. Through extensive experiments on the first large-scale Chinese harmful meme dataset ToxiCN MM, we empirically validate the efficacy of our proposed FG-E2HMD framework. The FG-E2HMD we proposed achieved the best results on the ToxiCN MM test set, with 83.39 F1 score and 74.43 F1_harm score, respectively. The results demonstrably establish that the explanation-enhanced decision-making approach, particularly when augmented with cultural context integration, constitutes a robust solution for Chinese harmful meme detection. We attribute the method’s success to the explicit explanation generation mechanism, which compels the model to transition from superficial pattern recognition to sophisticated logical reasoning. This mechanism, similar to the chain-of-thought, enables the model to decompose complex problems and focus on the key semantics in the task, thereby avoiding intuitive misjudgments arising from reliance on superficial data features alone.

Future Work

In terms of future work, we will design more effective methods to address the task of Chinese harmful meme detection. This will be developed from the following aspects:

• Automatic knowledge updating mechanism: we will investigate methods to dynamically integrate external knowledge bases such as online encyclopedias, modern language dictionaries, and real-time news feeds into the model to further improve its ability to learn continuously and adapt to evolving meme trends.
• Cross-cultural multilingual meme detection: the proposed framework will be extended to other languages and cultural contexts (e.g., Japanese, Korean) to explore the development of a flexible, configurable content moderation system capable of operating effectively in multicultural environments.
• Zero-shot and few-shot learning: we will examine how the model can leverage its strong inference capabilities to achieve efficient, explanation-enhanced harmful content detection under data-scarce or zero-shot conditions, thereby broadening its applicability to emerging and low-resource risk scenarios.
• Lightweight interpretation model solution: We will design a knowledge distillation framework to transfer the knowledge from large interpretive models to small models and meanwhile, introduce model pruning and quantization techniques. We aim to reduce the number of parameters and computational complexity and explore a simplified version of the attention mechanism while retaining key explanatory capabilities and reducing computational overhead. In addition, we plan to design hierarchical explanation strategies and generate explanations in stages based on the harmfulness level of the memes.