Source-linked AI summary
TransMeme: A Multi-Agent Framework for Cross-Cultural Meme Transcreation
Jingyi Zheng, Yule Liu, Zifan Peng, Tianyi Hu, Yuemeng Zhao, Xinhu Zheng, Xinlei He
TL;DR
Cross-cultural meme transcreation must preserve communicative intent, adapt culture-dependent meaning, and maintain text-image coherence across audiences. TransMeme addresses this problem with a coordinated multi-agent plan-and-revise framework, and it consistently outperforms baselines on bidirectional Chinese-English evaluation. Remaining errors are concentrated in humor reconstruction and image-text alignment, while the study evaluates the framework in a controlled single-model-family setting.
Problem
Cross-cultural meme transcreation requires jointly preserving intent, adapting cultural meaning, and maintaining image-text coherence beyond ordinary translation or rewriting.
Method
TransMeme coordinates specialized agents for cultural planning, constrained target-text rewriting with DPO, critic-guided revision, and conditional visual execution.
Results
TransMeme consistently outperforms all baselines across human and LLM evaluation, achieving the best human scores on all four dimensions and the highest LLM Top-1 ranking rate.
Takeaways & Limitations
The results highlight explicit cultural reasoning and coordinated multimodal adaptation as important for effective meme transcreation.
Takeaways & Limitations
The evaluation uses a controlled single-model-family setting, and remaining non-top-ranked cases are dominated by weak humor reconstruction and image-text mismatch.
Abstract
from arXiv · showhide
Internet memes are a pervasive form of multimodal online communication; however, such communication often involves users from diverse linguistic and cultural backgrounds. Therefore, adapting memes across cultures and languages is a central challenge for enabling mutual understanding in online communication. Unlike ordinary translation or standalone text rewriting, cross-cultural meme transcreation must jointly preserve communicative intent, adapt culture-dependent meaning for the target audience, and maintain coherence between text and image. In this work, we first provide an explicit task analysis of cross-cultural meme transcreation and identify three core challenges: culture-specific knowledge understanding, intent and tone preservation, and multimodal consistency. Based on this analysis, we propose a multi-agent framework with specialized agents that are coordinated to address these challenges through cultural adaptation, target text rewriting, revision, and conditional visual adjustment. The framework strengthens target text adaptation with coordinated feedback to handle difficult cases that require deeper cultural or visual intervention. We evaluate the framework on bidirectional Chinese-English meme transcreation using both human evaluation and LLM-as-a-Judge. Our method consistently outperforms all baselines across both evaluation settings. In human evaluation, it achieves the best performance on all four dimensions and delivers a 33.1% average improvement over the strongest baseline, while in LLM-as-a-Judge, it attains the highest Top-1 ranking rate (60% versus 26% for the second-best baseline). Further analysis indicates that each component contributes to the performance. Our error analysis suggests that the remaining bottlenecks lie in humor reconstruction and image-text alignment rather than simple cultural knowledge gaps, pointing to future work on humor transfer.
1 Introduction
Cross-cultural meme transcreation must preserve communicative effects while adapting culture-dependent meaning and coordinating text with visuals. TransMeme addresses these challenges with a coordinated multi-agent framework and outperforms strong baselines in human and LLM-based evaluation.
- Motivation: Cross-cultural meme transfer is difficult because memes combine written text, visual templates, and implicit shared cultural knowledge.Successful transfer reconstructs a communicative effect for a different audience rather than merely restating source content.
- Challenges: The task analysis identifies three central challenges: culture-specific knowledge understanding, intent and tone preservation, and multimodal consistency.These requirements distinguish meme transcreation from ordinary translation and standalone text rewriting.
- Framework: TransMeme combines structured understanding, cultural planning, DPO-enhanced text rewriting, critic-guided revision, and conditional visual execution.Its specialized agents coordinate adaptation and revision for difficult cultural or visual cases.
- Results: 33.1% average improvement over the strongest baseline was achieved in human evaluation, with the best scores on all four dimensions.The overall human-evaluation average was 4.122 versus 3.097 for the strongest baseline.
- Results: 60% Top-1 ranking rate was achieved in LLM-as-a-Judge, compared with 26% for the second-best baseline.The method remained robust across bidirectional Chinese-English meme transcreation.
2 Task Analysis
Cross-cultural meme transcreation is defined as producing a target-culture meme that preserves communicative intent, adapts culture-dependent meaning, and maintains image-text coherence. Its challenges arise because cultural references, pragmatic effects, and visual framing interact.
- Task Requirements: The task produces a target-culture meme whose text may be rewritten and whose image may be preserved or edited.It combines cross-lingual transfer, textual reformulation, visual adaptation, and cultural adaptation without reducing to any one related task.
- Challenges: Cultural adaptation requires deciding whether source references should be retained, lightly reformulated, or replaced with target-culture counterparts.Literal transfer can preserve wording while failing to remain understandable, natural, or appropriate for the target audience.
- Task Requirements: Intent preservation requires retaining humor, sarcasm, ridicule, irony, emotional stance, and other pragmatic functions.Semantic similarity alone may leave the text close in meaning while losing its punch, tone, or humor.
- Challenges: Multimodal consistency requires judging whether the original image remains suitable, should be edited, or should be replaced.The same text can succeed with one template and fail with another, so text-only adaptation is insufficient.
- Challenge Interaction: The three challenges are tightly connected because cultural references, rhetorical effects, and visual expression can jointly determine a meme’s meaning.Therefore, the task requires explicit understanding, deliberate planning, careful rewriting, and a final coherence check.
3 Method
TransMeme uses five specialized agents and an execution layer to route, culturally adapt, rewrite, critique, revise, and render cross-cultural memes. Its design follows the task’s three challenge categories and coordinates intermediate decisions through a shared harness.
- Framework Overview: The framework contains the Interpreter, Coordinator, Adapter, Rewriter, Critic, and an Execution Layer for final visual realization.The design directly follows the three challenge categories identified in the task analysis.
- Understanding and Control: The Interpreter recovers semantic and pragmatic content, including communicative effect, humor mechanism, emotional stance, background knowledge, and image-text dependency.It writes source-meme understanding into the Transcreation Harness for downstream control.
- Understanding and Control: The Coordinator routes simple cases through lightweight processing and difficult cases through full transcreation, then manages revision decisions from critical feedback.Routing cues include literal traps, text length, reaction-meme patterns, and image-text coupling strength.
- Cultural Adaptation Planning: The Adapter creates cultural and visual adaptation plans, assigning literal, minimal-rephrase, or cultural-map tiers before generation.It estimates the target audience’s comprehension gap and determines whether cultural elements should be preserved, reformulated, or replaced.
- Text Realization: The Rewriter generates target-side meme text under explicit adaptation constraints and uses DPO trained on ranked colloquial and slang rewrites.The constraints include adaptation tiers, cultural mappings, rhetorical goals, and the planned visual setting.
- Revision and Execution: The Critic evaluates intent preservation, cultural adaptation, and image-text coherence, triggering targeted revision when inconsistencies arise.The Transcreation Harness stores intermediate outputs, routing decisions, mappings, generated text, and critic feedback before final rendering.
4.1 Experimental Settings
The evaluation uses bidirectional Chinese-English meme transcreation with human judgments on a subset and LLM-based evaluation on the full set. The LLM evaluator is validated against human judgments before large-scale comparison.
- Dataset and baselines: 1,000 memes are evaluated bidirectionally, including 500 ZH→EN and 500 EN→ZH instances, against three baselines.The baselines include direct one-shot transcreation and structured single-agent transcreation.
- Human evaluation: 200 instances, evenly split across directions, receive human ratings for intent preservation, cultural adaptation, multimodal coherence, and expression quality.Five annotators independently rate anonymized outputs using the original meme and randomized system ordering.
- LLM evaluation: The LLM evaluator receives the original meme and all final rendered outputs, producing the same four scalar scores and an overall ranking.It is applied to the full evaluation set after validation on the human-annotated subset.
- Evaluator validation: 89.0% Top-1 agreement with aggregated human judgments supports using the LLM evaluator as a scalable comparison proxy, while human evaluation remains primary.Scalar-score Spearman correlations range from 0.693 to 0.802, with weakest alignment on Cultural Adaptation.
4.2 Main Results
TransMeme consistently outperforms all baselines in human and full-set LLM evaluations. Its advantage is stable across evaluation dimensions and both Chinese-English transfer directions.
- Human evaluation: 4.122 overall average on human evaluation makes M best across all four dimensions, versus 3.097 for strongest baseline B2.M scores 3.950, 4.097, 4.111, and 4.328 on Intent Preservation, Cultural Adaptation, Multimodal Coherence, and Expression Quality, respectively.
- Statistical significance: All gains over the baselines are statistically significant across every dimension and the overall average, with all p<0.001 after Bonferroni correction.The tests use two-sided Wilcoxon signed-rank tests.
- Directional robustness: M remains best in both ZH→EN and EN→ZH, with method ordering stable across directions.M, B1, and B2 perform better in ZH→EN, whereas B3 shows the opposite pattern.
- Qualitative comparison: M’s qualitative gains include more natural humor reconstruction, stronger cultural substitution, and better image-text coherence than the baselines.Baselines more often produce literal or awkward text, insufficient cultural adaptation, or image-text mismatches.
4.3 Component Analysis
Component analyses examine the rewriter, critic-revision loop, and cultural planning module. Results show strong contributions from DPO-enhanced rewriting and targeted revision, with cultural planning used asymmetrically across directions.
- Rewriter analysis: 4.177 overall text-quality average makes M best, exceeding A1 without DPO at 2.838 and larger-model replacement A2 at a lower level.M leads A1 and A2 on Naturalness, Cultural quality, and Humor.
- Critic-revision loop: 84.0% of triggered cases prefer revised outputs, raising the overall average from 2.453 to 3.709.Revision also improves Intent Preservation from 2.352 to 3.632, Cultural Adaptation from 2.272 to 3.704, and Expression Quality from 2.736 to 3.792.
- Cultural planning: 181 of 1,000 samples trigger cultural planning, including 155 ZH→EN and 26 EN→ZH cases.The asymmetry indicates more frequent explicit cultural planning when adapting Chinese memes for English audiences.
- Cultural planning: Cultural planning is essential for cases requiring non-literal cultural reinterpretation or visual substitution.The module handles adaptations such as culturally dependent humor and visual changes.
4.4 Error Analysis
Among cases where TransMeme does not rank first, weak humor reconstruction is the dominant error, followed by image-text mismatch; cultural knowledge gaps are comparatively rare.
- Error distribution: 400 of 1,000 evaluation samples do not rank TransMeme first.Manual analysis identifies weak humor reconstruction, image-text mismatch, and cultural knowledge gaps as the three main error types.
- Error distribution: 67.8% of non-top-ranked cases involve weak humor reconstruction, making it the dominant failure mode.These outputs generally preserve source intent but produce humor that is less witty, natural, or punchy in the target culture.
- Error distribution: 27.0% of non-top-ranked cases involve image-text mismatch, where plausible rewritten text does not fully align with the visual template or discourse framing.Cultural knowledge gaps account for 5.3% of these cases.
4.5 Discussion on Model-Family Scope
The experiments aim to validate TransMeme’s framework design through controlled comparisons of system organization rather than exhaustively benchmarking model families.
- The study evaluates framework design, not every possible model-family combination.
- A controlled single-family setting tests whether role decomposition, cultural planning, and iterative revision improve on a strong end-to-end alternative.The strongest baseline already uses a strong single-agent setup.
- Broader cross-family validation remains future work.
5 Related Work
Prior meme research emphasizes understanding, generation, and multimodal translation, but does not directly address cross-cultural transcreation as coordinated multimodal adaptation. The closest benchmark establishes the task without fully analyzing its challenges or providing a tailored method.
- Prior work mainly covers meme understanding, meme generation, and multimodal translation.
- Meme understanding research treats memes primarily as multimodal recognition and interpretation problems.Applications include harmful-content detection, emotion analysis, captioning, metaphor interpretation, and contextual grounding.
- These research lines do not directly address transcreation requiring intent preservation, humor reconstruction, and text-image coherence.
- Zhao et al. introduced a benchmarked cross-cultural meme transcreation task but did not characterize its core challenges or propose a tailored solution.
6 Conclusion
The paper frames cross-cultural meme transcreation as multimodal adaptation requiring intent preservation, cultural meaning adaptation, and text-image coherence. TransMeme addresses these requirements with plan-and-revise agents, and experiments show consistent gains over strong baselines.
- Cross-cultural meme transcreation requires preserving intent, adapting culture-dependent meaning, and maintaining text-image coherence.
- TransMeme uses explicit cultural planning, iterative revision, and conditional visual execution in a multi-agent framework.
- Experiments on Chinese-English meme transcreation consistently outperform strong baselines, with analyses confirming contributions from the main components.