MemeCMD: An Automatically Generated Chinese Multi-turn Dialogue Dataset with Contextually Retrieved Memes

Pufeng Huang; Xianhe Tang; Yuheng Wang

REVIEW 3 major objections 5 minor 43 references

MemeCMD: An Automatically Generated Chinese Multi-turn Dialogue Dataset with Contextually Retrieved Memes

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MemeCMD introduces a fully automatic pipeline that generates Chinese multi-turn dialogues with contextually retrieved memes and reports that its retrieval strategies beat a random baseline on all evaluated dimensions.

desk verdict A genuinely new Chinese meme-dialogue dataset pipeline, but the core retrieval equation has a sign error and the evaluation is too thin to support the claims as written. read the letter →

arxiv 2507.00891 v1 pith:E7XHIS6G submitted 2025-07-01 cs.CL cs.AI

Yuheng Wang , Xianhe Tang , Pufeng Huang This is my paper · ORCID

classification cs.CLcs.AI

keywords memeretrievalmultimodaldialoguedatasetChinesemulti-turnautomaticgenerationMLLMannotationdual-agentadaptivethresholdLLM-as-a-Judgeevaluation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper presents a fully automatic pipeline that turns unlabeled Chinese meme images and agent-generated conversations into a multi-turn dialogue dataset in which each turn can carry a contextually matched meme. Its central claim is that MemeCMD, with 34,758 dialogue turns, is the first automatically constructed Chinese multi-turn dialogue dataset with meme-enhanced responses. The authors argue that a retrieval score combining scenario fit, forbidden-scenario penalty, implicit emotion, and speaker motivation, plus an adaptive threshold that spaces out meme use, produces meme insertions that beat a random baseline on both LLM-as-a-Judge and embedding-based consistency evaluations. If correct, the contribution is a scalable and privacy-preserving way to obtain meme-aware conversational data without manual annotation or real chat logs.

What carries the argument

The load-bearing mechanism is the Meme Aligner, a retrieval function that scores every candidate meme as a weighted sum of four cosine similarities between the dialogue state and the meme's MLLM-generated annotations: scenario matching $\alpha$, penalty for forbidden scenarios $\delta$, implicit emotional/semantic fit $\beta$, and motivational fit $\gamma$. It is paired with an adaptive threshold $\theta_{i,t}=\theta_0+\Delta e^{-\lambda k_{i,t}}$ that prevents consecutive meme sends, and with a selector that picks the top-1 meme under greedy choice or samples from the top-$K$ under diversity-aware sampling. This mechanism is what converts raw image captions and dialogue summaries into contextually appropriate meme placements.

What would settle it

Take a random sample of around 200 MemeCMD turns, blind the strategy labels, and have independent Chinese-speaking raters judge whether the inserted meme is contextually appropriate and emotionally fitting. If the random baseline matches or beats the greedy and top-$K$ strategies under human rating, or if raters disagree with the MLLM annotations on a substantial share of the meme library, the central claim of contextually appropriate automatic meme selection would be falsified.

Watch

Extended reading notes

Core claim

MemeCMD's core claim is that contextual meme insertion can be generated automatically rather than collected from real chats or hand-annotated. Starting from 6,023 curated Chinese meme images, a multimodal LLM annotates each image with appropriate scenarios $S^+$, inappropriate scenarios $S^-$, implicit emotion $E$, and user motivation $\Psi$; a dual-agent dialogue system then produces multi-turn conversations from news or role-based cold starts, while a Summary Agent condenses the dialogue into scenario, emotion, and motivation queries. The Meme Aligner scores every meme against these queries with a weighted cosine-similarity sum $T_{i,t,m}=\omega_1\alpha+\omega_2\delta+\omega_3\beta+\omega_4\gamma$, and an adaptive threshold $\theta_{i,t}=\theta_0+\Delta e^{-\lambda k_{i,t}}$ decides whether a meme is sent, with greedy or top-$K$ sampling choosing the final image. Experiments report that both retrieval strategies outperform a random-control baseline on all evaluated datasets and dimensions, supporting the claim of a reproducible, privacy-preserving pipeline for meme-inclusive Chinese dialogue data.

Load-bearing premise

The load-bearing premise is that the multimodal-LLM annotations of each meme—which scenarios fit, which do not, what emotion it carries, and what motivation it serves—are accurate enough to serve as ground truth for retrieval, and the paper reports no human validation of those labels.

Editorial extensions

If this is right

Meme-enriched Chinese dialogue data can be produced at scale without human annotation, since meme labeling, conversation generation, and meme placement are all automatic.
The same annotation-plus-summary retrieval score can be applied to stickers, GIFs, emoji, or other visual reply media once analogous embeddings exist.
The adaptive threshold gives downstream users a direct knob for meme density, so the same pipeline can mimic meme-heavy or meme-sparse chat styles.
The released dataset and library become a reusable testbed for training and evaluating meme-aware response generation and meme retrieval.
Because dialogues are synthesized rather than scraped, the dataset sidesteps the privacy risks of real user chat logs, a stated design goal of the pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the portability of the pipeline matters more than the 34,758-turn count; if the MLLM labels are even weakly reliable, the same recipe can refresh the meme library as new memes appear.
Editorial inference: the paper's own acknowledgment that MLLMs are weak at emotional/implicit meme understanding suggests the embedding-based consistency score is a lower-bound sanity check, and a human-preference study on a subsample would locate the true performance gap.
Editorial inference: comparing retrieval quality under MLLM-generated labels against the same retrieval under human-annotated labels would turn annotation noise into a directly measurable quantity.
Editorial inference: the retrieval score could be reused as a reward signal for training a policy that decides when and which meme to send, rather than only for offline dataset construction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things you should know up front. First, the paper does deliver a new artifact: a 6,023-meme library with four-dimensional MLLM annotations, a dual-agent pipeline that generates 34,758 Chinese dialogue turns, and an interpretable retrieval score with adaptive threshold decay. That combination is not in the cited literature, and the pipeline is sensible enough to be reusable if the details were pinned down. Second, there is a load-bearing formal error in the retrieval score. Eq. (20) defines implicit semantic matching as beta = -cos(E_i,t, E_m). A negative cosine means a meme whose emotion annotation matches the context gets a lower raw score. The surrounding prose says the opposite, and the reported results only make sense if the implementation used a positive cosine or if omega_3 is negative. The weights omega_1 through omega_4 are never reported, so as written the method is internally inconsistent. This is not a nitpick; it is the scoring function that allegedly outperforms random. The authors likely made a sign typo, but they need to correct it and confirm the implementation matches the corrected formula.

What the paper does well: the cold-start designs (news-based and role-based) are thoughtful, the summary agent extracting scenario/emotion/motivation is a reasonable idea, and the adaptive threshold decay is a nice mechanism for spacing meme insertions. The evaluation tries to triangulate with both LLM-as-a-judge and an independent Chinese CLIP embedding metric, which shows the authors are aware of the circularity risk. The dataset, if released with annotations, could be a useful resource.

The soft spots are the ones the reader flagged, plus the sign error. The MLLM annotations are never validated against humans, and the same model family (GPT-4) generates the dialogues, annotates the memes, and judges the outputs, which can inflate apparent quality. The embedding-based gains are tiny (0.06 to 0.21 points) and come without significance tests or confidence intervals. The human evaluation is described only in a few anecdotal bullet points with no numbers or protocol. Neither the dataset nor the exact prompts are released, only a code repository, so independent verification is currently impossible. There is also no direct comparison with existing meme/sticker datasets like MOD or StickerInt, which would put the claimed novelty in context.

Who this is for: researchers working on synthetic multimodal dialogue or meme retrieval. It deserves a serious referee, but not acceptance as is. I would send it out with a request for major revision: fix and report the retrieval weights, correct Eq. (20), validate a sample of the annotations with human raters, add statistical tests, and release the dataset and prompts.

Referee Report

3 major / 5 minor

Summary. The paper introduces MemeCMD, an automatically generated Chinese multi-turn dialogue dataset in which memes are inserted into conversations. The authors first build a meme library of 6,023 images, each annotated by an MLLM with four dimensions: suitable scenarios, unsuitable scenarios, implicit emotions/internet meanings, and psychological motivations. They then use a dual-agent GPT-4-based framework to generate multi-turn dialogues from news-based or role-based cold starts. At each turn, a Summary Agent produces a structured representation of the dialogue context, and a Meme Aligner scores candidate memes using a weighted combination of scenario similarity, scenario penalty, implicit semantic matching, and motivation alignment. An adaptive threshold with exponential decay controls meme frequency. The paper evaluates Greedy and Top-K Sampling selection against a Random baseline using LLM-as-a-Judge and a cross-modal embedding-based semantic consistency metric, reporting that retrieval-based strategies outperform random across all settings. The main claimed contributions are the annotated meme library, the automatic pipeline, and the first automatically constructed Chinese multi-turn meme-enhanced dialogue dataset with 34,758 turns.

Significance. If the results hold, the paper offers a scalable and privacy-preserving pipeline for constructing meme-incorporated dialogue data, which is a genuinely useful resource for multimodal conversational AI. The manual curation of the meme collection, the release of code, and the use of an independent Chinese CLIP embedding metric as a secondary evaluation are concrete strengths that partially mitigate concerns about the GPT-4-based generation and annotation loop. The central novelty—an automatically generated Chinese multi-turn dialogue dataset with contextually retrieved memes—is clearly stated and, if properly validated, would be a contribution to the community. However, the evaluation evidence for the central claim is currently weaker than the prose suggests, and one equation in the retrieval formulation appears internally inconsistent with the described behavior. These issues are addressable but need to be fixed before the claims can be accepted.

major comments (3)

[Section 4.2.1, Eq. (20)] The implicit semantic matching score is defined as β = -cos(E_{i,t}, E_m). Since cosine similarity is maximized when the two vectors are aligned, the negative sign makes β largest when the summary emotion and the meme emotion are anti-correlated. This directly contradicts the surrounding text, which states that this component should 'prioritize memes that ... resonate at a deeper semantic and emotional level.' The weights ω1–ω4 in Eq. (22) are never reported, so the reader cannot tell whether the implementation uses a negative ω3 to invert the sign (in which case Eq. (20) misstates the method) or uses the equation as written (in which case the described scoring function penalizes the intended matches). Please correct the sign, report the full weight configuration, and confirm which form was used in the experiments. Because Eq. (22) is the core retrieval score behind Table 1 and Figure 4, this is a load-bearing inconsistency.
[Section 5, Table 1 and Figure 4] The embedding-based semantic consistency gaps over the Random baseline are 0.06–0.21 points on a 0–100 scale. No significance tests, confidence intervals, or evaluation-set sizes are reported, so the claim that retrieval strategies 'consistently outperform' the random baseline, and the Figure 4 caption's use of 'significantly higher,' are not statistically supported. In addition, the primary LLM-as-a-Judge metric uses chatgpt-4o-latest, which is from the same model family (GPT-4) used to generate the dialogues and annotate the memes, creating a possible same-model evaluation loop. The human evaluation mentioned in Section 5.4 is only a qualitative statement with no protocol, sample size, or numerical results. Please add statistical testing and confidence intervals, provide details on the human evaluation or remove the claim, and report the independence of the judge from the generation process.
[Section 3, Eq. (1)] The meme library annotations are produced entirely by the MLLM mapping f_MLLM, with no human validation, inter-annotator agreement, or manual audit reported. These annotations are the ground truth against which the retrieval score in Eq. (22) operates; if the labels are noisy or biased, the 'contextually appropriate' dataset claim is weakened. Please include a validation study on a random sample of the 6,023 memes, such as human agreement on appropriateness, emotion, and motivation labels, or at least a qualitative audit reporting error rates and disagreement patterns.

minor comments (5)

[Eq. (21)] The denominator in the motivation alignment score uses Ψ_u, which is undefined; it should presumably be Ψ_{i,t}, the summary agent's motivation vector.
[Eq. (18)] The denominator in the scenario matching score is malformed: it reads '∥S_{i,t}|∥S+_m∥' instead of ∥S_{i,t}∥ ∥S+_m∥. Please fix the norm notation.
[Section 5.3] The retrieval weights ω1–ω4 in Eq. (22) and the Top-K value for sampling are not reported; only θ0, Δ, and λ are given. All hyperparameters should be listed to make the experiments reproducible.
[Abstract and Section 1] The paper states that code is available but does not state where the dataset itself can be downloaded. Since the dataset is a primary contribution, a release location or an explicit statement about data availability should be provided.
[Figure 3] The axis labels in Figure 3 contain rendering artifacts such as '/glyph1197ews-based' and should be regenerated to read 'News-based 6 Turns' and so on.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the dataset pipeline and retrieval evaluation do not reduce to their inputs by construction.

full rationale

The paper's derivation chain is an empirical pipeline rather than a formal derivation, and the central claims are not equivalent to their inputs by construction. The meme annotations (Eq. 1), dialogue summaries (Eq. 16), and retrieval score (Eq. 22) are defined from separate MLLM and embedding components, and no parameter is fitted to the evaluation outcome or renamed as a prediction. There are no load-bearing self-citations, and the secondary Chinese-CLIP semantic-consistency metric provides an external, generation-independent signal, so the evaluation is not a pure tautology. Flagged as correctness/transparency concerns rather than circularity: Section 5.3's LLM-as-a-Judge uses chatgpt-4o-latest, the same model family as the GPT-4 generator and annotator (Sections 2.1 and 3), which can inflate apparent quality; Eq. (20) defines implicit semantic matching with a negative cosine, contradicting the prose unless an unstated negative weight omega_3 is used, and the weights omega_1 through omega_4 in Eq. (22) are never reported; Section 3 reports no human validation of MLLM annotations. These are load-bearing specification and validity issues, but not cases where a result reduces to its own inputs by definition, so the circularity score remains 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central load-bearing assumptions are unvalidated MLLM annotations, the validity of embedding cosine similarity for meme appropriateness, and the realism of GPT-4-generated dialogues. The retrieval weights in Eq. (22) are free parameters with no reported values, and the threshold parameters are hand-chosen. No genuinely new entities such as particles or forces are introduced; the adaptive threshold is a mechanism, not an invented entity.

free parameters (5)

Retrieval weights omega_1 to omega_4 = not specified
Equation 22 uses a weighted sum of scenario, penalty, implicit, and motivation scores; the weights are neither reported nor fitted, and the retrieval score depends on them.
Base threshold theta_0 = 0.7
Set in Section 5.3; controls whether a meme is inserted at a turn.
Penalty term Delta = 0.2
Set in Section 5.3; adds a penalty to the adaptive threshold after a recent meme.
Decay rate lambda = 1
Set in Section 5.3; controls the exponential decay of the adaptive threshold over turns.
Top-K size for sampling = 3
Used in Diversity-aware Sampling in Section 5.3; no justification is given for choosing K=3.

assumptions (4)

domain assumption MLLM annotations of appropriate scenarios, inappropriate scenarios, emotions, and motivations are accurate and consistent for all 6,023 memes.
Section 3, Eq. (1) defines the annotation mapping f_MLLM, but no human validation or inter-annotator agreement is reported.
domain assumption Cosine similarity in the MiniCPM embedding space is a valid measure of meme-context appropriateness.
Section 4.2 defines all retrieval scores as cosine similarities between embeddings, without a learned aligner or calibration to human judgments.
domain assumption Dual-agent GPT-4 dialogues are realistic and representative of natural Chinese online conversations.
Section 4.1 generates dialogues automatically; no human naturalness study is reported.
domain assumption LLM-as-a-judge scores approximate human preferences for meme appropriateness.
Section 5.1 relies on the GPT-4 judge without providing evidence that its scores correlate with human judgments for this specific meme-insertion task.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MemeCMD: An Automatically Generated Chinese Multi-turn Dialogue Dataset with Contextually Retrieved Memes." pith.science (2026). https://pith.science/paper/E7XHIS6G

@misc{pith2026250700891,
  author       = {Pith},
  title        = {Pith review of: MemeCMD: An Automatically Generated Chinese Multi-turn Dialogue Dataset with Contextually Retrieved Memes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7XHIS6G}},
  note         = {Machine review of arXiv:2507.00891}
}

read the original abstract

Memes are widely used in online social interactions, providing vivid, intuitive, and often humorous means to express intentions and emotions. Existing dialogue datasets are predominantly limited to either manually annotated or pure-text conversations, lacking the expressiveness and contextual nuance that multimodal interactions provide.To address these challenges, we introduce MemeCMD, an automatically generated Chinese Multi-turn Dialogue dataset with contextually retrieved memes. Our dataset combines a large-scale, MLLM-annotated meme library with dialogues auto-generated by dual agents across diverse scenarios. We introduce a retrieval framework and adaptive threshold to ensure contextually relevant, naturally spaced meme usage. Experiments demonstrate the effectiveness of our approach in generating contextually appropriate and diverse meme-incorporated dialogues, offering a scalable and privacy-preserving resource for advancing multimodal conversational AI.

Figures

Figures reproduced from arXiv: 2507.00891 by the authors.

**Figure 2.** Keyword Distributions Across Semantic Dimensions [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The distribution of Retrieval Scores in the Meme [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Cross-Modal Embedding-Based Semantic Consistency Evaluation Results demonstrate that both Greedy selection and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Comparative Effectiveness of Multi-Strategy Meme [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 32 canonical work pages

[1]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

arXiv 2023
[2]

Bo Adler, Niket Agarwal, Ashwath Aithal, Dong H Anh, Pallab Bhattacharya, Annika Brundyn, Jared Casper, Bryan Catanzaro, Sharon Clay, Jonathan Cohen, et al. 2024. Nemotron-4 340b technical report. arXiv preprint arXiv:2406.11704 (2024)

arXiv 2024
[3]

Bahar Aydemir, Deblina Bhattacharjee, Tong Zhang, Mathieu Salzmann, and Sabine Süsstrunk. 2024. Data Augmentation via Latent Diffusion for Saliency Prediction. In European Conference on Computer Vision . Springer, 360–377

work page 2024
[4]

Francesco Barbieri, Miguel Ballesteros, Francesco Ronzano, and Horacio Sag- gion. 2018. Multimodal Emoji Prediction. In NAACL-HLT (2). Association for Computational Linguistics, 679–686

work page 2018
[5]

Heng Er Metilda Chee, Jiayin Wang, Zhiqiang Guo, Weizhi Ma, and Min Zhang

work page
[6]

Claude Coulombe. 2018. Text Data Augmentation Made Simple By Leveraging NLP Cloud APIs. ArXiv abs/1812.04718 (2018). https://api.semanticscholar.org/ CorpusID:54481373

work page Pith review arXiv 2018
[7]

Haixing Dai, Zhengliang Liu, Wenxiong Liao, Xiaoke Huang, Yihan Cao, Zihao Wu, Lin Zhao, Shaochen Xu, Fang Zeng, Wei Liu, et al. 2025. Auggpt: Leveraging chatgpt for text data augmentation. IEEE Transactions on Big Data (2025)

work page 2025
[8]

Zhengcong Fei, Zekang Li, Jinchao Zhang, Yang Feng, and Jie Zhou. 2021. Towards expressive communication with internet memes: A new multimodal conversation dataset and benchmark. arXiv preprint arXiv:2109.01839 (2021)

work page Pith review arXiv 2021

Show all 43 references

[9]

Shen Gao, Xiuying Chen, Chang Liu, Li Liu, Dongyan Zhao, and Rui Yan. 2020. Learning to respond with stickers: A framework of unifying multi-modality in multi-turn dialog. In Proceedings of the Web Conference 2020 . 1138–1148

2020
[10]

GetActivity. 2025. EmojiPackage: A high-quality QQ and WeChat emoji set. https://github.com/getActivity/EmojiPackage. Accessed: 2025-07-01

2025
[11]

Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al . 2024. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594 (2024)

2024 arXiv
[12]

Hang Guo, Yong Guo, Yaohua Zha, Yulun Zhang, Wenbo Li, Tao Dai, Shu-Tao Xia, and Yawei Li. 2025. Mambairv2: Attentive state space restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference . 28124–28133

2025
[13]

Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zhen Leng Thai, Kaihuo Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chaochao Jia, Guoyang Zeng, Daha...

2024 arXiv
[14]

Akshi Kumar. 2025. Natural Language Processing for Affective, Psychological, and Content Analysis. 159–221. doi:10.1002/9781394297290.ch7

2025 doi
[15]

Guanzhou Lan, Qianli Ma, Yuqi Yang, Zhigang Wang, Dong Wang, Yuan Yuan, and Bin Zhao. 2024. Towards Flexible and Efficient Diffusion Low Light Enhancer. arXiv preprint arXiv:2410.12346 (2024)

2024
[16]

Bin Liang, Bingbing Wang, Zhixin Bai, Qiwei Lang, Mingwei Sun, Kaiheng Hou, Lanjun Zhou, Ruifeng Xu, and Kam-Fai Wong. 2024. Reply with Sticker: New Dataset and Model for Sticker Retrieval. arXiv preprint arXiv:2403.05427 (2024)

2024
[17]

Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, and Kaipeng Zhang. 2024. ConvBench: A Multi-Turn Conversation Evaluation Benchmark with Hierarchical Ablation Capability for Large Vision-Language Models. In Advanc...

2024
[18]

Shengzhe Liu, Xin Zhang, and Jufeng Yang. 2022. SER30K: A large-scale dataset for sticker emotion recognition. In Proceedings of the 30th ACM International Conference on Multimedia. 33–41

2022
[19]

Yihang Luo, Shangchen Zhou, Yushi Lan, Xingang Pan, and Chen Change Loy
[20]

Ni Luh Nilasari, I Nengah Sudipa, and Ni Wayan Sukarini. 2018. Sticker Emoticons Used in LINE Messenger; A Semantic Study. J. Humanis 22 (2018), 585–591

2018
[21]

In Proceedings of the Computer Vision and Pattern Recognition Conference

3DEnhancer: Consistent Multi-View Diffusion for 3D Enhancement. In Proceedings of the Computer Vision and Pattern Recognition Conference . 16430– 16440
[22]

Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada Mihalcea. 2019. MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics...

2019 doi
[23]

Pictory.ai. 2024. What Do You Meme: How Memes Enhance Digital Commu- nication and Culture. https://pictory.ai/blog/memes-digital-culture. Accessed: 2025-07-01

2024
[24]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

2022
[25]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

2021
[26]

Chhavi Sharma, Deepesh Bhageria, William Scott, Srinivas PYKL, Amitava Das, Tanmoy Chakraborty, Viswanath Pulabaigari, and Björn Gambäck. 2020. SemEval- 2020 Task 8: Memotion Analysis- the Visuo-Lingual Metaphor!. In Proceedings of the Fourteenth Workshop on Semantic Evaluatio...

2020 doi
[27]

Karsten Roth, Zeynep Akata, Dima Damen, Ivana Balazevic, and Olivier J Hénaff
[28]

In Proceedings of the Computer Vision and Pattern Recognition Conference

Context-aware multimodal pretraining. In Proceedings of the Computer Vision and Pattern Recognition Conference . 4267–4279
[29]

Bingbing Wang, Yiming Du, Bin Liang, Zhixin Bai, Min Yang, Baojun Wang, Kam- Fai Wong, and Ruifeng Xu. 2025. A New Formula for Sticker Retrieval: Reply with Stickers in Multi-Modal and Multi-Session Conversation. In AAAI Conference on Artificial Intelligence. https://api.seman...

2025
[30]

Xu Tan, Jiawei Chen, Haohe Liu, Jian Cong, Chen Zhang, Yanqing Liu, Xi Wang, Yichong Leng, Yuanhao Yi, Lei He, et al. 2024. Naturalspeech: End-to-end text-to- speech synthesis with human-level quality. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 6 (2024)...

2024
[31]

LLM Red Team. 2025. emo-visual-data: Emotion-centric Visual Dataset. https: //github.com/LLM-Red-Team/emo-visual-data. Accessed: 2025-07-01

2025
[32]

Xiaoxue Zang, Lijuan Liu, Maria Wang, Yang Song, Hao Zhang, and Jindong Chen. 2021. PhotoChat: A Human-Human Dialogue Dataset With Photo Shar- ing Behavior For Joint Image-Text Modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics...

2021
[33]

Xiaoyang Wang, Chen Li, Jianqiao Zhao, and Dong Yu. 2021. Naturalconv: A chinese dialogue dataset towards multi-turn topic-driven conversation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 14006–14014

2021
[34]

Zhen Wang, Dongyuan Li, Guang Li, Ziqing Zhang, and Renhe Jiang. 2024. Multimodal Low-light Image Enhancement with Depth Information. In ACM Multimedia 2024. https://openreview.net/forum?id=Tsz6Kra6fX

2024
[35]

Sijie Zhao, Yixiao Ge, Zhongang Qi, Lin Song, Xiaohan Ding, Zehua Xie, and Ying Shan. 2023. Sticker820k: Empowering interactive retrieval with stickers. arXiv preprint arXiv:2306.06870 (2023)

2023 arXiv
[36]

Qixuan Zhang, Zhifeng Wang, Dylan Zhang, Wenjia Niu, Sabrina Caldwell, Tom Gedeon, Yang Liu, and Zhen Qin. 2024. Visual Prompting in LLMs for Enhancing Emotion Recognition. In Conference on Empirical Methods in Natural Language Processing. https://api.semanticscholar.org/Corpu...

2024
[37]

Yiqun Zhang, Fanheng Kong, Peidong Wang, Shuang Sun, Lingshuai Wang, Shi Feng, Daling Wang, Yifei Zhang, and Kaisong Song. 2024. STICKERCONV: Generating Multimodal Empathetic Responses from Scratch. In Annual Meeting of the Association for Computational Linguistics . https://a...

2024
[38]

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems 36 (2023), 46595–46623

2023
[39]

Zhengyi Zhao, Shubo Zhang, Yuxi Zhang, Yanxi Zhao, Yifan Zhang, Zezhong Wang, Huimin Wang, Yutian Zhao, Bin Liang, Yefeng Zheng, et al. 2025. MemeRea- Con: Probing Contextual Meme Understanding in Large Vision-Language Models. arXiv preprint arXiv:2505.17433 (2025)

2025 arXiv
[40]

Zhaoolee. 2025. ChineseBQB. https://github.com/zhaoolee/ChineseBQB. Ac- cessed: 2025-07-01

2025
[42]

Yixuan Zhou, Xiaoyu Qin, Zeyu Jin, Shuoyi Zhou, Shun Lei, Songtao Zhou, Zhiyong Wu, and Jia Jia. 2024. VoxInstruct: Expressive Human Instruction-to- Speech Generation with Unified Multilingual Codec Language Modelling. In Proceedings of the 32nd ACM International Conference on...

2024
[43]

Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large lan- guage models. arXiv preprint arXiv:2304.10592 (2023)

2023 arXiv
[2025]

In Proceedings of the ACM on Web Conference 2025

PerSRV: Personalized Sticker Retrieval with Vision-Language Model. In Proceedings of the ACM on Web Conference 2025 . 293–303

2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

arXiv 2023

[2] [2]

Bo Adler, Niket Agarwal, Ashwath Aithal, Dong H Anh, Pallab Bhattacharya, Annika Brundyn, Jared Casper, Bryan Catanzaro, Sharon Clay, Jonathan Cohen, et al. 2024. Nemotron-4 340b technical report. arXiv preprint arXiv:2406.11704 (2024)

arXiv 2024

[3] [3]

Bahar Aydemir, Deblina Bhattacharjee, Tong Zhang, Mathieu Salzmann, and Sabine Süsstrunk. 2024. Data Augmentation via Latent Diffusion for Saliency Prediction. In European Conference on Computer Vision . Springer, 360–377

work page 2024

[4] [4]

Francesco Barbieri, Miguel Ballesteros, Francesco Ronzano, and Horacio Sag- gion. 2018. Multimodal Emoji Prediction. In NAACL-HLT (2). Association for Computational Linguistics, 679–686

work page 2018

[5] [5]

Heng Er Metilda Chee, Jiayin Wang, Zhiqiang Guo, Weizhi Ma, and Min Zhang

work page

[6] [6]

Claude Coulombe. 2018. Text Data Augmentation Made Simple By Leveraging NLP Cloud APIs. ArXiv abs/1812.04718 (2018). https://api.semanticscholar.org/ CorpusID:54481373

work page Pith review arXiv 2018

[7] [7]

Haixing Dai, Zhengliang Liu, Wenxiong Liao, Xiaoke Huang, Yihan Cao, Zihao Wu, Lin Zhao, Shaochen Xu, Fang Zeng, Wei Liu, et al. 2025. Auggpt: Leveraging chatgpt for text data augmentation. IEEE Transactions on Big Data (2025)

work page 2025

[8] [8]

Zhengcong Fei, Zekang Li, Jinchao Zhang, Yang Feng, and Jie Zhou. 2021. Towards expressive communication with internet memes: A new multimodal conversation dataset and benchmark. arXiv preprint arXiv:2109.01839 (2021)

work page Pith review arXiv 2021