Pith. sign in

REVIEW 5 major objections 5 minor 40 references

ViSP: A PPO-Driven Framework for Sarcasm Generation with Contrastive Learning

T0 review · 5 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 The paper claims that combining PPO rewards from DIP with contrastive candidate ranking, trained on a new filtered multimodal dataset, makes sarcasm generation outperform all baselines, including LLMs.

desk verdict A useful dataset and broad baselines undermined by a circular sarcasm metric that filters, rewards, selects, and evaluates—needs major revision before the quality claims hold. read the letter →

arxiv 2507.09482 v1 pith:OFXBCK4U submitted 2025-07-13 cs.CL cs.AIcs.HC

Changli Wang , Rui Wu , Fang Yin This is my paper · ORCID
classification cs.CLcs.AIcs.HC
keywords multimodalsarcasmgenerationM2SaGdatasetProximalPolicyOptimizationcontrastivelearningrewardmodelingvision-languagefactualincongruitylargelanguagemodels
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

The paper sets out to show that image-driven sarcasm generation, which has lagged behind sarcasm detection, can be made to work by combining three ingredients: a carefully filtered multimodal dataset, a reward signal from a sarcasm-intensity model, and reinforcement learning plus contrastive ranking. To that end it introduces M2SaG, a 4,970-sample dataset of image, sarcastic text, and explicit sarcasm targets, and ViSP, a ViLT-BART generator trained with PPO and contrastive learning. The authors report that ViSP beats all baselines across five metric sets, including large language models, and that its outputs receive higher mean sarcasm scores (0.898 vs 0.770) and factual incongruity (0.768 vs 0.739) than the human-authored dataset. A sympathetic reader would care because the result suggests a concrete recipe for steering generation toward a subjective quality, and because it documents where current LLMs fall short.

What carries the argument

The load-bearing object is DIP, a pretrained network that scores sarcasm by modeling factual and affective incongruity between image and text. In this paper DIP does triple duty: it filters M2SaG by removing samples scoring below 0.5, it supplies the reward $r_\theta(x,y)$ in the PPO loss, and it is the evaluator behind the final quality comparison. The generation pipeline is carried by ViLT, a vision-and-language transformer that fuses the image with a text prompt built from the sarcasm target, OCR text, image caption, and detected objects; its $[CLS]$ token is concatenated into BART, a sequence-to-sequence transformer that generates top-$k$ candidate texts. PPO updates the generator toward higher DIP rewards while a momentum encoder approximates the supervised fine-tuned policy for the KL penalty, and contrastive learning uses the highest-scoring candidate as anchor and the lower-scoring candidates as negatives. The total loss is $\mathcal{L} = \mathcal{L}_{\mathrm{ce}} + \lambda_{\mathrm{ppo}} \mathcal{L}_{\mathrm{ppo}} + \lambda_{\mathrm{cl}} \mathcal{L}_{\mathrm{cl}}$.

What would settle it

Have human annotators blind to source rate a random sample of M2SaG originals plus ViSP, baseline, and LLM outputs for sarcasm strength, then compare the human ordering to DIP's ordering; if DIP disagrees with human judgment on a substantial share, or if humans do not rate ViSP outputs as more sarcastic, the central claim that ViSP produces higher-quality sarcasm is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a PPO-driven, contrastive-learning-enhanced generation framework can produce sarcastic texts whose sarcastic intent is stronger, as measured by DIP, than both competing systems and the original dataset. The central comparison in Table 2 shows ViSP ahead of every baseline on BLEU-1 to BLEU-4 (24.26, 13.71, 8.92, 5.68), ROUGE, CIDEr, METEOR, and Sent-BERT similarity, with the largest gains on BLEU and CIDEr. The paper also claims to be the first to apply PPO loss to sarcasm generation, and it interprets the weak performance of large language models as evidence that instruction-tuned LLMs lack the cross-modal incongruity handling that sarcasm requires. In the dataset analysis, ViSP outputs receive higher mean DIP sarcasm scores and higher factual incongruity than the original M2SaG texts.

Load-bearing premise

The load-bearing premise is that DIP's sarcasm scores track what human readers experience as sarcasm, because the same scorer shapes the dataset, provides the training reward, and serves as the quality metric for the final comparison.

Editorial extensions

If this is right

  • Multimodal sarcasm generation can be improved by reward steering from a sarcasm-intensity evaluator, so the bottleneck is less about model scale than about a usable quality signal.
  • M2SaG provides an explicit-target benchmark, so future systems can be compared on the same images, texts, and targets rather than on hashtag-mismatched data.
  • The poor showing of large language models indicates that current LLM prompting does not capture cross-modal incongruity, making specialized multimodal training a more promising route.
  • Top-k candidate generation with a contrastive anchor and negative samples is a transferable recipe for inducing a generator to produce texts that a reward model prefers.
  • The finding that generated texts are both more sarcastic and more incongruent with images than the original dataset suggests the model has learned to amplify distinctive features of sarcastic expression.

Reading between the lines

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

  • The paper's own limitations section lists dependence on the external sarcasm evaluator's quality, unstable PPO training, and noise from naively concatenating OCR text with the sarcasm target, so the recipe's transfer to other datasets is not guaranteed.
  • The same DIP model is used as data filter, reward, and evaluation metric, so the reported quality gain could partly be an artifact of optimizing toward that scorer; a human-judgment study is the natural check.
  • Because the contrastive anchor is chosen by DIP rather than by humans, the model may converge on a narrow high-scoring style; measuring output diversity and running human preference tests would reveal whether variety is lost.
  • The LLM comparison may understate what large models can do, since the paper does not attempt in-context examples or reward-based fine-tuning for the LLMs; a stronger test would give those models the same target prompt and candidate selection before concluding they are inherently weak at sarcasm.
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.

Referee Report

5 major / 5 minor

Summary. The paper introduces M2SaG, a multimodal sarcasm-generation dataset of 4,970 image–text–target triples, and ViSP, a ViLT+BART framework trained with a combination of supervised cross-entropy, a DIP-scored reward objective labeled PPO, and a contrastive loss. The authors report that ViSP outperforms text-only, VLM, and LLM baselines on five metric families, and that the generated texts have higher mean DIP-based Sarcasm Scores (0.898 vs. 0.770) and higher CLIP-based Factual Incongruity (0.768 vs. 0.739) than the original dataset.

Significance. If the empirical claims were sound, M2SaG and ViSP would be a useful contribution to multimodal sarcasm generation. The paper contains a substantial dataset-construction effort, a broad baseline comparison spanning text-only models, VLMs, and LLMs, and a stated commitment to release data and code. The main conditioning result, however, is not currently supported: DIP is used to filter the training data, to select positive and negative candidates during training, to choose the inference output, and as the Sarcasm Score evaluation metric in Fig. 4(a). The reported quality gains are therefore expected consequences of optimizing DIP, not independent evidence of sarcastic quality. In addition, Eq. (4) is not a PPO objective, and the dataset arithmetic is internally inconsistent. These issues affect the central claims of the paper, not merely the presentation.

major comments (5)
  1. [Dataset; Analysis of Generated Sarcastic Texts (Fig. 4(a)); Discussion and Conclusion] The central quality claim is circular. DIP scores are used (1) to discard training samples with scores below 0.5 in the Dataset section, (2) as the reward r_theta in Eq. (4), (3) to select the positive anchor and negatives in the Contrastive Learning Module, (4) to select the final output at inference, and (5) as the Sarcasm Score in Fig. 4(a). The reported improvement from 0.770 to 0.898 is therefore a direct artifact of optimizing the same evaluator, not evidence that ViSP produces objectively better sarcasm. The paper's own limitation statement in Discussion and Conclusion acknowledges that performance relies on the external sarcasm evaluator and that poor evaluators may introduce bias, but no validation of DIP against human judgments or an independent sarcasm metric is provided. A human evaluation or at least an independently trained sarcasm classifier is needed before the claim 'higher-quality sarcastic content' can be supported.
  2. [Contrastive Learning Module, Eq. (2)] The positive sample g_i^+ is defined as 'identical to the anchor' g_i, so sim(g_i, g_i^+) is always 1 and the numerator exp(1/tau) is a constant. The loss in Eq. (2) therefore contains no learnable alignment term; it only pushes the negative samples away from the anchor. The text's claim that contrastive learning 'encourages the model to favor outputs with higher reward scores' is not supported by the equation as written.
  3. [PPO Reinforcement Module, Eq. (4)] Eq. (4) is not a Proximal Policy Optimization objective. It lacks the importance-sampling ratio, the clipped surrogate objective, and advantage estimation; it is a KL-regularized expected-reward objective of the kind used in RLHF. Consequently, the paper's claim to be 'the first to introduce reinforcement learning with PPO loss into the domain of sarcasm generation' is not substantiated, and the title's 'PPO-Driven' is misleading without a corrected formulation.
  4. [Dataset] The dataset arithmetic is internally inconsistent. The text states that filtering MSTI yields 2,678 samples and filtering MORE+ yields 2,279 samples, and that the combined result is 4,970 samples, but 2,678 + 2,279 = 4,957. The sentence 'Combined with other data' is vague and does not explain the source of the missing 13 samples. This must be corrected before the dataset can be audited.
  5. [Main Results, Table 2; Ablations, Table 3] No error bars, repeated runs, or significance tests are reported for any of the automatic metrics. Given that generation is stochastic and that the final output is selected as the top DIP-scored candidate among k candidates, the claim that ViSP 'consistently outperforms all baselines' needs at least variance estimates over multiple seeds or a paired significance test.
minor comments (5)
  1. [Related Works heading] The heading 'Realted Works' is misspelled and should be 'Related Works'.
  2. [Table 3 caption and Table 3(b)] The caption says 'BLUE-1' where BLEU-1 is intended, and Table 3(b) has the header 'Loss fuction' which should be 'Loss function'.
  3. [Experiment Settings] The sentence 'The temperature parameter tau in Lppo is set to 0.07' appears to refer to the contrastive loss in Eq. (2), since Eq. (4) contains no temperature parameter; please clarify which loss the hyperparameter applies to.
  4. [Comparison Models] The LLM baselines (LLaVA, DeepSeek, Qwen, Hunyuan) are identified only by family name; no prompt templates, decoding hyperparameters, or exact model versions are given, which makes the comparison difficult to reproduce.
  5. [Table 1] The column headers 'Avg. length |V|' are ambiguous: it is not clear whether |V| denotes vocabulary size or sequence length for the sarcastic text and target columns; consider separating the two quantities.

Circularity Check

4 steps flagged · score 7.0 of 10

The paper's central sarcasm-quality claim is not independently supported: DIP is used as dataset filter, training reward, inference selector, and evaluation metric, so the reported quality gains are expected by construction.

  1. self definitional [Dataset: Exclusion; Introduction (Fig. 1)]
    "we employ the pretrained DIP (Wen, Jia, and Yang 2023) to evaluate each sample, removing those with scores below 0.5, which results in a subset of 2,678 samples. ... We evaluate the remaining samples with DIP, filtering out those scoring below 0.5, which results in 2,279 samples. ... our M2SaG dataset exhibits a higher mean sarcasm score of 0.7700 and a lower standard deviation of 0.1817, indicating the presence of stronger and more consistent sarcastic content."

    The dataset is constructed by retaining only samples whose DIP score exceeds 0.5. The reported mean sarcasm score of 0.7700 is therefore imposed by the inclusion rule, not discovered. This filter also sets the reference distribution for the later comparison 'generated 0.898 vs. original 0.770', so the comparison is between a DIP-optimized output and a DIP-truncated reference population.

  2. fitted input called prediction [Methodology: PPO Reinforcement Module (Eq. 4); Experiments: Analysis of Generated Sarcastic Texts, Fig. 4(a)]
    "We adopt DIP as a reward model to optimize the generation policy. Specifically, DIP serves as a sarcasm-aware evaluator that provides feedback on the quality of generated texts in terms of their sarcastic intent. ... rθ(x, y) denotes the reward score from DIP ... We compute the sarcasm scores of generated texts and compare them with annotated sarcastic texts in Fig. 4(a). Compared to the original data, the generated texts exhibit a higher mean sarcasm score of 0.898 ..."

    The same DIP score is used as the reward rθ in the PPO objective (Eq. 4) and then reused as the 'Sarcasm Score' evaluation metric in Fig. 4(a). Maximizing a reward function guarantees higher values of that function, so the 0.898 vs. 0.770 comparison measures the optimization target rather than an independent property such as human-judged sarcasm. The paper's own limitation note concedes that 'model performance relies on the quality of the external sarcasm evaluator', but no human or independent validation of DIP is provided.

2 more flagged steps
  1. fitted input called prediction [Methodology: Overview; Experiments: Analysis of Generated Sarcastic Texts, Fig. 4(a)]
    "At inference, the model generates top-k candidates and selects the highest-scoring one as the final output. ... We compute the sarcasm scores of generated texts and compare them with annotated sarcastic texts in Fig. 4(a). Compared to the original data, the generated texts exhibit a higher mean sarcasm score of 0.898 ..."

    At inference, the final output is selected as the candidate with the highest DIP score. Evaluating that selected output with the same DIP-based Sarcasm Score is selecting on the evaluation variable: the maximum of k samples will exceed the average score of a single unselected or human-written text even if generation quality is unchanged. Thus the claimed quality advantage is in part a selection artifact, not evidence of improved sarcastic competence.

  2. fitted input called prediction [Experiments: Analysis of Generated Sarcastic Texts, Fig. 4(b); Methodology: PPO Reinforcement Module]
    "Following DIP (Wen, Jia, and Yang 2023), we compute the image-text similarity score S_inter using CLIP, and define factual incongruity as 1−S_inter. As shown in Fig. 4(b), the generated sarcastic texts exhibit a higher factual incongruity, with a mean of 0.763 compared to 0.739 for the original data, indicating a stronger degree of sarcasm."

    DIP is explicitly a 'Dual Incongruity Perceiving' network, and factual incongruity is one of the two incongruity components it models. Since DIP is the reward in Eq. 4, the generator is directly rewarded for increasing this quantity. Reporting higher factual incongruity in Fig. 4(b) as evidence of quality is therefore measuring the reward variable again, not an independent human judgment or an external sarcasm standard.

full rationale

The circularity is concentrated in the paper's sarcasm-quality claim, not in the dataset's external annotations or in the BLEU, ROUGE, CIDER, METEOR, and Sent-BERT comparisons, which are independent of DIP. DIP is nonetheless used in four places that make the '0.898 vs. 0.770' and '0.768 vs. 0.739' claims near-tautological: as a dataset filter, as the PPO reward rθ in Eq. 4, as the contrastive-learning anchor selector, and as the inference-time top-k selector. The final output is therefore selected and trained to maximize exactly the variable that Fig. 4 then reports as evidence of quality. The paper's limitation note admits that 'model performance relies on the quality of the external sarcasm evaluator', but no human or independent validation of DIP is supplied, so the reported gains over the original dataset are expected by construction. The factual-incongruity metric is also not independent, since DIP is explicitly a dual-incongruity perceiver and factual incongruity is one of the two components it models. Because the standard n-gram and semantic metrics in Table 2 are not DIP-derived, the paper is not wholly circular; the self-definition applies to the sarcasm-intensity distribution analysis and the 'higher quality' conclusion drawn from it. No load-bearing self-citation chain is present; the DIP citation is to legitimate external prior work, so the issue is one of metric reuse rather than citation. An additional auditing inconsistency (2,678 + 2,279 = 4,957, not the stated 4,970) does not change the circularity verdict but reduces confidence in the dataset accounting.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on DIP as a reliable sarcasm scorer, on a contrastive loss whose positive is identical to its anchor, and on a policy objective that is labeled PPO but lacks PPO's defining clipping. These assumptions are not independently validated in the paper.

free parameters (5)
  • DIP score threshold = 0.5
    Used to filter the dataset; samples below threshold removed. Choice affects dataset composition and downstream metrics.
  • Contrastive loss weight lambda_cl = 0.5
    Set to 0.5 after ablation; influences final loss balance.
  • Candidate count k = 5
    Number of top-k candidates generated during training and inference.
  • Temperature tau = 0.07
    Temperature in the contrastive loss.
  • PPO loss weight schedule lambda_ppo = linear schedule, exact values not specified
    Scheduled linearly during training; exact progression is not reported, making the optimization setup partially underspecified.
assumptions (3)
  • domain assumption DIP provides a valid and unbiased measure of sarcasm intensity.
    The paper uses DIP scores to filter the dataset, as reward in Eq. 4, and as evaluation metric in Fig. 4(a), without independent validation against human judgments.
  • ad hoc to paper The contrastive loss with the positive sample identical to the anchor is a meaningful learning signal.
    In Eq. 2, g+ is identical to g, making the numerator constant; the loss reduces to pushing negative samples apart, which does not teach the model to match ground-truth text.
  • ad hoc to paper The reward optimization in Eq. 4 is a valid PPO objective.
    Eq. 4 omits clipping and advantage estimation, so it is not PPO; the reference policy is a moving EMA rather than a fixed SFT policy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViSP: A PPO-Driven Framework for Sarcasm Generation with Contrastive Learning." pith.science (2026). https://pith.science/paper/OFXBCK4U

@misc{pith2026250709482,
  author       = {Pith},
  title        = {Pith review of: ViSP: A PPO-Driven Framework for Sarcasm Generation with Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OFXBCK4U}},
  note         = {Machine review of arXiv:2507.09482}
}
read the original abstract

Human emotions are complex, with sarcasm being a subtle and distinctive form. Despite progress in sarcasm research, sarcasm generation remains underexplored, primarily due to the overreliance on textual modalities and the neglect of visual cues, as well as the mismatch between image content and sarcastic intent in existing datasets. In this paper, we introduce M2SaG, a multimodal sarcasm generation dataset with 4,970 samples, each containing an image, a sarcastic text, and a sarcasm target. To benchmark M2SaG, we propose ViSP, a generation framework that integrates Proximal Policy Optimization (PPO) and contrastive learning. PPO utilizes reward scores from DIP to steer the generation of sarcastic texts, while contrastive learning encourages the model to favor outputs with higher reward scores. These strategies improve overall generation quality and produce texts with more pronounced sarcastic intent. We evaluate ViSP across five metric sets and find it surpasses all baselines, including large language models, underscoring their limitations in sarcasm generation. Furthermore, we analyze the distributions of Sarcasm Scores and Factual Incongruity for both M2SaG and the texts generated by ViSP. The generated texts exhibit higher mean Sarcasm Scores (0.898 vs. 0.770) and Factual Incongruity (0.768 vs. 0.739), demonstrating that ViSP produces higher-quality sarcastic content than the original dataset. % The dataset and code will be publicly available. Our dataset and code will be released at \textit{https://github.com/wclapply/ViSP}.

Figures

Figures reproduced from arXiv: 2507.09482 by the authors.

Figure 1
Figure 1. Comparison between MuSG and M2SaG datasets: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of ViSP. the Generation Module, the Supervised Learning Module, the PPO Reinforcement Module, and the Contrastive Learn￾ing Module. Specifically, given a sample (I, T, ST), where I denotes the image, T the sarcastic text, and ST the sar￾casm target, the OCR text (OT) is first extracted from I us￾ing EasyOCR. We concatenate ST and OT into a sequence T ′ , which, along with I, is input to ViLT… view at source ↗
Figure 3
Figure 3. Contrastive Learning Module. We designate the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Evaluation of generated sarcastic texts using Sarcasm Scores and Factual Incongruity. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Black, S.; Gao, L.; Wang, P.; Leahy, C.; and Biderman, S. 2021. Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. If you use this software, please cite it using these metadata, 58(2)

    work page 2021

  4. [4]

    Cai, Y.; Cai, H.; and Wan, X. 2019. Multi-modal sarcasm detection in twitter with hierarchical fusion model. In Proceedings of the 57th annual meeting of the association for computational linguistics, 2506--2515

    work page 2019

  5. [5]

    Castro, S.; Hazarika, D.; P \'e rez-Rosas, V.; Zimmermann, R.; Mihalcea, R.; and Poria, S. 2019. Towards Multimodal Sarcasm Detection (An \_Obviously\_ Perfect Paper). In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Florence, Italy: Association for Computational Linguistics

    work page 2019

  6. [6]

    Desai, P.; Chakraborty, T.; and Akhtar, M. S. 2022. Nice perfume. how long did you marinate in it? multimodal sarcasm explanation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 10563--10571

    work page 2022

  7. [7]

    Target-Augmented Shared Fusion-based Multimodal Sarcasm Explanation Generation

    Goel, P.; Chauhan, D. S.; and Akhtar, M. S. 2025. Target-Augmented Shared Fusion-based Multimodal Sarcasm Explanation Generation. arXiv preprint arXiv:2502.07391

    work page Pith review arXiv 2025

  8. [8]

    Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    arXiv 2025

Show all 40 references
  1. [9]

    He, K.; Fan, H.; Wu, Y.; Xie, S.; and Girshick, R. 2020. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9729--9738

    2020

  2. [10]

    Jia, M.; Xie, C.; and Jing, L. 2024. Debiasing multimodal sarcasm detection with contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 18354--18362

    2024

  3. [11]

    Jing, L.; Song, X.; Ouyang, K.; Jia, M.; and Nie, L. 2023. Multi-source Semantic Graph-based Multimodal Sarcasm Explanation Generation. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 11349--11361. Association for Computational Linguistics

    2023

  4. [12]

    Khanam, R.; and Hussain, M. 2024. YOLOv11: An Overview of the Key Architectural Enhancements. arXiv:2410.17725

    2024 arXiv

  5. [13]

    Kim, W.; Son, B.; and Kim, I. 2021. ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision. In Meila, M.; and Zhang, T., eds., Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research,...

    2021

  6. [14]

    Kumar, A. 2022. The Illustrated Image Captioning using transformers. ankur3107.github.io

    2022

  7. [15]

    S.; and Chakraborty, T

    Kumar, S.; Kulkarni, A.; Akhtar, M. S.; and Chakraborty, T. 2022. When did you become so smart, oh wise one?! Sarcasm Explanation in Multi-modal Multi-party Dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    2022

  8. [16]

    S.; and Chakraborty, T

    Kumar, S.; Mondal, I.; Akhtar, M. S.; and Chakraborty, T. 2023. Explaining (sarcastic) utterances to enhance affect understanding in multimodal dialogues. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 12986--12994

    2023

  9. [17]

    Lewis, M.; Liu, Y.; Goyal, N.; Ghazvininejad, M.; Mohamed, A.; Levy, O.; Stoyanov, V.; and Zettlemoyer, L. 2020. BART : Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. In Jurafsky, D.; Chai, J.; Schluter, N.; and Tet...

    2020

  10. [18]

    Li, J.; Li, D.; Xiong, C.; and Hoi, S. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, 12888--12900. PMLR

    2022

  11. [19]

    Lin, H.; Chen, Z.; Luo, Z.; Cheng, M.; Ma, J.; and Chen, G. 2024. CofiPara: A Coarse-to-fine Paradigm for Multimodal Sarcasm Target Identification with Large Multimodal Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1...

    2024

  12. [20]

    Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2023. Visual instruction tuning. Advances in neural information processing systems, 36: 34892--34916

    2023

  13. [21]

    Lu, Q.; Sun, X.; Long, Y.; Zhao, X.; Zou, W.; Feng, J.; and Wang, X. 2025. Multimodal dual perception fusion framework for multimodal affective analysis. Information Fusion, 115: 102747

    2025

  14. [22]

    Lunando, E.; and Purwarianti, A. 2013. Indonesian social media sentiment analysis with sarcasm detection. In 2013 International Conference on Advanced Computer Science and Information Systems (ICACSIS), 195--198. IEEE

    2013

  15. [23]

    McDonald, S. 1999. Exploring the process of inference generation in sarcasm: A review of normal and clinical studies. Brain and language, 68(3): 486--506

    1999

  16. [24]

    Olkoniemi, H.; Ranta, H.; and Kaakinen, J. K. 2016. Individual differences in the processing of written sarcasm and metaphor: Evidence from eye movements. Journal of Experimental Psychology: Learning, Memory, and Cognition, 42(3): 433

    2016

  17. [25]

    Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 27730--27744

    2022

  18. [26]

    Qin, L.; Huang, S.; Chen, Q.; Cai, C.; Zhang, Y.; Liang, B.; Che, W.; and Xu, R. 2023. MMSD 2.0: Towards a Reliable Multi-modal Sarcasm Detection System. In Rogers, A.; Boyd-Graber, J.; and Okazaki, N., eds., Findings of the Association for Computational Linguistics: ACL 2023,...

    2023

  19. [27]

    Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; Sutskever, I.; et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8): 9

    2019

  20. [28]

    Raffel, C.; Shazeer, N.; Roberts, A.; Lee, K.; Narang, S.; Matena, M.; Zhou, Y.; Li, W.; and Liu, P. J. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research, 21(140): 1--67

    2020

  21. [29]

    Sun, X.; Chen, Y.; Huang, Y.; Xie, R.; Zhu, J.; Zhang, K.; Li, S.; Yang, Z.; Han, J.; Shu, X.; et al. 2024. Hunyuan-large: An open-source moe model with 52 billion activated parameters by tencent. arXiv preprint arXiv:2411.02265

    2024 arXiv

  22. [30]

    Team, Q. 2025. Qwen3 Technical Report. arXiv:2505.09388

    2025 arXiv

  23. [31]

    Tian, Y.; Xu, N.; Zhang, R.; and Mao, W. 2023. Dynamic routing transformer network for multimodal sarcasm detection. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2468--2480

    2023

  24. [32]

    Wang, J.; Sun, L.; Liu, Y.; Shao, M.; and Zheng, Z. 2022 a . Multimodal sarcasm target identification in tweets. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8164--8175

    2022

  25. [33]

    Wang, J.; Yang, Z.; Hu, X.; Li, L.; Lin, K.; Gan, Z.; Liu, Z.; Liu, C.; and Wang, L. 2022 b . GIT: A Generative Image-to-text Transformer for Vision and Language. arXiv preprint arXiv:2205.14100

    2022 arXiv

  26. [34]

    Wei, Y.; Yuan, S.; Zhou, H.; Wang, L.; Yan, Z.; Yang, R.; and Chen, M. 2024. G\^ 2SAM: Graph-Based Global Semantic Awareness Method for Multimodal Sarcasm Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 9151--9159

    2024

  27. [35]

    Wellman, H. M. 2014. Making minds: How theory of mind develops. Oxford University Press

    2014

  28. [36]

    Wen, C.; Jia, G.; and Yang, J. 2023. Dip: Dual incongruity perceiving network for sarcasm detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2540--2550

    2023

  29. [37]

    Yue, T.; Mao, R.; Wang, H.; Hu, Z.; and Cambria, E. 2023. KnowleNet: Knowledge fusion network for multimodal sarcasm detection. Information Fusion, 100: 101921

    2023

  30. [38]

    V.; Mihaylov, T.; Ott, M.; Shleifer, S.; Shuster, K.; Simig, D.; Koura, P

    Zhang, S.; Roller, S.; Goyal, N.; Artetxe, M.; Chen, M.; Chen, S.; Dewan, C.; Diab, M.; Li, X.; Lin, X. V.; Mihaylov, T.; Ott, M.; Shleifer, S.; Shuster, K.; Simig, D.; Koura, P. S.; Sridhar, A.; Wang, T.; and Zettlemoyer, L. 2022. OPT: Open Pre-trained Transformer Language Mo...

    2022 arXiv

  31. [39]

    Zhao, W.; Huang, Q.; Xu, D.; and Zhao, P. 2023. Multi-modal sarcasm generation: Dataset and solution. In Findings of the Association for Computational Linguistics: ACL 2023, 5601--5613

    2023

  32. [40]

    Zhuang, X.; Li, Z.; Zhang, C.; and Ma, H. 2025. A cross-modal collaborative guiding network for sarcasm explanation in multi-modal multi-party dialogues. Engineering Applications of Artificial Intelligence, 142: 109884

    2025

Pith tools

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