Pith. sign in

REVIEW 5 major objections 6 minor 48 references

CLAMP: Contrastive Learning with Adaptive Multi-loss and Progressive Fusion for Multimodal Aspect-Based Sentiment Analysis

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CLAMP claims top F1 scores on two multimodal sentiment benchmarks by aligning text and images at both global and local levels.

desk verdict Coherent architecture with a sensible new combination, but the SOTA claim leans on single-run F1 gaps of 0.1–0.6 points; the paper deserves review, but the empirical claim needs more support. read the letter →

arxiv 2507.16854 v1 pith:KDU4OKKZ submitted 2025-07-21 cs.CV cs.AI

Xiaoqiang He This is my paper · ORCID
classification cs.CVcs.AI
keywords multimodalaspect-basedsentimentanalysiscontrastivelearningprogressiveattentionfusionuncertainty-weightedmulti-lossoptimaltransportalignmentTwitter-2015Twitter-2017RoBERTa-ViT
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 proposes CLAMP, an end-to-end framework for multimodal aspect-based sentiment analysis that combines progressive attention fusion with multi-task contrastive learning and adaptive loss weighting. The central claim is that aligning text and images at two granularities, globally through contrastive learning and locally through word-region optimal transport, then fusing them across staged attention layers and balancing the resulting losses by uncertainty, yields better aspect extraction and sentiment classification than prior methods. The reported evidence is an F1 of 67.7% on Twitter-2015 and 68.9% on Twitter-2017, edging past the strongest existing baselines by small margins. If the claim holds, the paper shows that fine-grained cross-modal alignment and adaptive gradient balancing, rather than any single fusion trick, are what move the needle on this task.

What carries the argument

The load-bearing mechanism is the combination of a Progressive Attention Fusion network and a Multi-task Contrastive Learning module coordinated by Adaptive Multi-loss Aggregation. PAF performs staged self-attention, cross-attention with text as query and image patches as key-value, relative position biases, and a gated residual so that cross-modal alignment deepens layer by layer. MCL adds global InfoNCE contrast, local word-region alignment via the IPOT optimal transport algorithm, and two CRF sequence classifiers, one on fused features and one on text-only features. AMA then combines fixed weights, softmax-normalized task priorities, and uncertainty-based scaling to form the total loss, which is intended to stop any single task from dominating the gradient.

What would settle it

Re-run CLAMP, M2DF, and MCPL on the Ling et al. revised Twitter-2015 and Twitter-2017 splits with identical train/dev/test partitions, the same pretrained RoBERTa and ViT backbones, and multiple random seeds; if M2DF or MCPL matches or exceeds CLAMP's F1 under this controlled comparison, the claimed state-of-the-art result does not hold.

Watch

Extended reading notes

Core claim

CLAMP is a RoBERTa-plus-ViT architecture in which a three-stage Progressive Attention Fusion network lets text features iteratively attend to image regions, a Multi-task Contrastive Learning module applies InfoNCE at the global level and optimal-transport word-region alignment at the local level while two CRF classifiers annotate sentiment, and an Adaptive Multi-loss Aggregation module weights the four task losses by learned uncertainty and task priority. The paper reports that this combination reaches 67.7% F1 on Twitter-2015 and 68.9% on Twitter-2017, which it states outperforms the vast majority of existing state-of-the-art methods, including 0.1-point gains over M2DF and MCPL on Twitter-2015 and larger gains on Twitter-2017. Ablations attribute the largest drop to removing progressive attention fusion, followed by removing adaptive loss aggregation and then multi-task contrastive learning, indicating that all three modules contribute to the result.

Load-bearing premise

The benchmark comparison is apples-to-apples: every baseline was scored on the same revised dataset version with the same training and evaluation protocol, which matters because CLAMP's lead over the closest methods on Twitter-2015 is only 0.1 F1.

Editorial extensions

If this is right

  • If the reported numbers hold, CLAMP is the current best performer on the Twitter-2015 and Twitter-2017 MABSA benchmarks for joint aspect extraction and sentiment classification.
  • The ablation results imply that progressive, stage-by-stage cross-modal fusion contributes more than either the contrastive learning module or the adaptive loss weighting, so future designs should prioritize multi-stage interaction.
  • The combination of global contrastive learning with local optimal-transport alignment suggests that both coarse and fine alignment signals are needed, not either one alone.
  • The uncertainty-based loss aggregation offers a general recipe for balancing several auxiliary tasks in a multimodal sequence-labeling model without hand-tuned weights.

Reading between the lines

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

  • The 0.1-point F1 margins over M2DF and MCPL on Twitter-2015 are within typical run-to-run variation, so the strongest defensible reading is that CLAMP is competitive with, rather than clearly better than, the top prior methods until multi-seed significance testing is reported.
  • The same progressive-fusion-plus-multi-task-contrastive recipe could transfer to other grounded sequence-labeling tasks, such as multimodal named entity recognition or visual question answering, where local image regions need to be tied to specific tokens.
  • A direct test of whether the word-region alignment loss is doing interpretable work would be to visualize the learned transport matrix on failure cases, since the paper does not show what the aligned regions look like.
  • The adaptive loss mechanism, which couples uncertainty weights with softmax task priorities, could be compared against simpler fixed-weight schedules to see whether the added complexity is responsible for the gains or whether the fusion and contrastive modules carry the result.
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 / 6 minor

Summary. The paper proposes CLAMP, an end-to-end joint MABSA model that combines a three-stage progressive attention fusion network (PAF), a multi-task contrastive learning module (MCL) with global and word-region alignment, and an adaptive uncertainty-weighted multi-loss aggregation (AMA). The model is evaluated on Twitter-2015 and Twitter-2017 for joint aspect extraction and sentiment classification, reporting F1 scores of 67.7 and 68.9 respectively, and is claimed to consistently outperform existing state-of-the-art methods. The paper also includes an ablation study, a case study, and dataset statistics, but provides no released code or significance testing.

Significance. If the reported numbers are reproducible and the comparison is controlled, the contribution is a useful incremental improvement: a small consistent gain over strong recent baselines (M2DF, MCPL) on two standard benchmarks, with a plausible architectural story and a three-way ablation. The method is empirical rather than theoretical, and its value depends almost entirely on the soundness of the benchmark claims. The main strengths are the clearly formulated joint task, the ablated contribution of each module, and the qualitative case study. The main weakness is that the claimed state-of-the-art status rests on single-run F1 values with no variance estimates, no significance tests, and no code or data release, so the central claim is not yet established at the claimed precision.

major comments (5)
  1. [4.3, Table 4] The central SOTA claim is supported only by single-run F1 numbers with no error bars, multiple seeds, or significance tests. The Twitter-2015 margin over M2DF and MCPL is 0.1 F1 points (67.7 vs 67.6), which is within the plausible run-to-run variation of deep multimodal models. Please report mean and variance over at least five seeds, perform a significance test (e.g., bootstrap or paired test over the test set), and release the code so that the comparison can be audited.
  2. [4.1, Table 4] The comparison is not controlled for dataset version or evaluation protocol. Section 4.1 states that Ling et al. [6] revised the Twitter datasets, yet Table 4 appears to aggregate numbers from different source papers without stating which dataset version each baseline used. If any baseline was evaluated on the earlier version or with a different tokenizer or decoding routine, the reported 0.1 and 0.6 F1 advantages could vanish. Please verify that every baseline was evaluated on exactly the same revised splits and labels, or rerun all baselines under a common protocol.
  3. [2.3, Table 4] The related-work section discusses AoM [7], Atlantis [9], and TCMT [10] as recent MABSA methods, but none appears in Table 4. Because the abstract and Section 4.3 claim superiority over 'state-of-the-art' methods and 'the vast majority' of existing methods, the benchmark table should include these discussed baselines, or the claims should be explicitly restricted to the methods listed in the table.
  4. [3.6, Eq. (32)] The total loss in Eq. (32) is defined over M=4 tasks including L_CLS, but L_CLS is nowhere defined. Eq. (29) defines L_CRF as the sum of two CRF losses, and Section 4.4 lists L = L_CRF + L_CLS + L_GCL + L_WRA with no specification of L_CLS. Without a definition of L_CLS, the objective and the AMA ablation are not reproducible. Please define L_CLS, including which classifier and label set it uses, or remove it from the objective.
  5. [3.4, Eq. (13)] The recurrence for the progressive stages is under-specified. Eq. (13) writes H̄_t,s = LayerNorm(H̄_t + H̄_ffn^s), but H̄_t is not clearly the stage input, H̄_ffn^s is not defined, and it is not stated whether the self-attention and cross-attention sublayers inside a stage have their own residual connections or only the final layer norm. Similar ambiguity affects Eq. (20), where H_t,s and H_v,s are used as queries and keys without a definition after Eq. (13). Please rewrite the per-stage update with explicit sublayer indices and specify how the output of stage s becomes the input of stage s+1.
minor comments (6)
  1. [3.4, Eq. (15)] The relative position bias matrix P is L_max by L_max and is added to every attention head, but Eq. (20) uses P_tv for text-image attention; please clarify the relationship between P and P_tv and state their shapes for the cross-modal case.
  2. [3.5, Eq. (28)] The notation switches between h_PAF and h_PAM in Eq. (28); please define both symbols consistently and state the dimensions involved in the product W_r h_t.
  3. [4.4, Table 5] The text says the order of importance is PAF, then AMA, then MCL, but on Twitter-2017 the F1 drops are PAF 2.0, MCL 1.6, and AMA 1.2, while on Twitter-2015 AMA (3.8) slightly exceeds PAF (3.7); please report the per-dataset ordering or temper the summary to match the table.
  4. [4.5, Figure 4] The VLP-MABSA row for example (a) shows empty parentheses with '×' markers; please explain what these symbols denote so that the case study is interpretable.
  5. [References] References [4] and [37] are the same BART paper, and references [5], [27], and [43] are the same JML paper; please consolidate duplicate references.
  6. [Data availability] The paper says data are available on request but gives no URL or repository identifier; releasing the code with a public link would materially improve reproducibility and support the empirical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CLAMP is an empirical architecture whose benchmark claims are standard evaluations, not predictions derived from fitted constants or self-citations.

full rationale

The paper makes no theoretical derivation or first-principles prediction; it presents an end-to-end neural architecture and reports F1 scores on Twitter-2015 and Twitter-2017. The adaptive multi-loss weights (Eq. 30-32) are learned during normal training and are not fitted to the benchmark labels to manufacture the reported F1 values. The claimed SOTA results depend on Table 4 comparisons against baselines from multiple papers, and the dataset revision by Ling et al. [6] raises an apples-to-apples concern, but that is a correctness or verifiability issue, not circularity. There are no self-citations by the author and no imported uniqueness theorem. Ablations and case studies are empirical observations. Therefore no load-bearing step reduces to its own inputs by construction.

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

No new theoretical entities are introduced; all components are standard neural modules. The main free parameters are the multi-loss aggregation hyperparameters, which are not reported, making the method hard to reproduce.

free parameters (4)
  • alpha (smoothing parameter) = unspecified
    In Eq (31), alpha blends uniform weights with priority softmax weights; no value or tuning range is reported.
  • temperature tau = unspecified
    Used in both InfoNCE (Eq 25) and priority softmax (Eq 30); value is not reported, and it is unclear if the same tau is shared.
  • initial uncertainty sigma_i and task priorities pi_i = unspecified
    Learnable parameters in Eq (32) with no initialization scheme or constraint details given.
  • PAF stage count and IPOT iterations = 3 stages; IPOT iteration count not reported
    The choice of three stages is fixed with no ablation, and the number of IPOT iterations used in Eq (27) is not specified.
assumptions (5)
  • standard math InfoNCE contrastive loss aligns global text-image semantics
    Eq (25) relies on the standard property that InfoNCE maximizes mutual information between paired modalities.
  • standard math Optimal transport (IPOT) produces a meaningful word-region alignment
    Eq (27) assumes the transport matrix T from IPOT provides one-to-one correspondences with negligible error.
  • standard math CRF decoding improves sequence tagging over independent classification
    The two CRF losses in Eq (29) assume label dependencies captured by CRF improve token-level F1.
  • domain assumption Pretrained RoBERTa and ViT encoders transfer to this domain
    The model relies on frozen or fine-tuned pretrained encoders without verification that the visual features are relevant to aspect sentiment.
  • domain assumption The revised Twitter datasets used by Ling et al. are the same evaluation set as baseline papers
    Baseline scores in Table 4 are taken from papers that may use different dataset splits or annotations; comparability is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLAMP: Contrastive Learning with Adaptive Multi-loss and Progressive Fusion for Multimodal Aspect-Based Sentiment Analysis." pith.science (2026). https://pith.science/paper/KDU4OKKZ

@misc{pith2026250716854,
  author       = {Pith},
  title        = {Pith review of: CLAMP: Contrastive Learning with Adaptive Multi-loss and Progressive Fusion for Multimodal Aspect-Based Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KDU4OKKZ}},
  note         = {Machine review of arXiv:2507.16854}
}
read the original abstract

Multimodal aspect-based sentiment analysis(MABSA) seeks to identify aspect terms within paired image-text data and determine their fine grained sentiment polarities, representing a fundamental task for improving the effectiveness of applications such as product review systems and public opinion monitoring. Existing methods face challenges such as cross modal alignment noise and insufficient consistency in fine-grained representations. While global modality alignment methods often overlook the connection between aspect terms and their corresponding local visual regions, bridging the representation gap between text and images remains a challenge. To address these limitations, this paper introduces an end to end Contrastive Learning framework with Adaptive Multi-loss and Progressive Attention Fusion(CLAMP). The framework is composed of three novel modules: Progressive Attention Fusion network, Multi-task Contrastive Learning, and Adaptive Multi-loss Aggregation. The Progressive Attention Fusion network enhances fine-grained alignment between textual features and image regions via hierarchical, multi-stage cross modal interactions, effectively suppressing irrelevant visual noise. Secondly, multi-task contrastive learning combines global modal contrast and local granularity alignment to enhance cross modal representation consistency. Adaptive Multi-loss Aggregation employs a dynamic uncertainty based weighting mechanism to calibrate loss contributions according to each task's uncertainty, thereby mitigating gradient interference. Evaluation on standard public benchmarks demonstrates that CLAMP consistently outperforms the vast majority of existing state of the art methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 26 canonical work pages

  1. [6]

    Ling, Y ., Yu, J., & Xia, R. (2022). Vision -language pre -training for multimodal aspect -based sentiment analysis. arXiv preprint arXiv:2204.07955. https://doi.org/10.48550/arXiv.2204.07955

    work page Pith review arXiv doi:10.48550/arxiv.2204.07955 2022

  2. [7]

    Zhou, R., Guo, W., Liu, X., Yu, S., Zhang, Y ., & Yuan, X. (2023). Aom: Detecting aspect-oriented information for multimodal aspect -based sentiment analysis. arXiv preprint arXiv:2306.01004. https://doi.org/10.48550/arXiv.2306.01004

    work page Pith review arXiv doi:10.48550/arxiv.2306.01004 2023

  3. [10]

    Zou, W., Sun, X., Wu, W., Lu, Q., Zhao, X., Bo, Q., & Yan, J. (2025). TCMT: Target -oriented Cross Modal Transformer for Multimodal Aspect -Based Sentiment Analysis. Expert Systems with Applications, 264, 125818. https://doi.org/10.1016/j.eswa.2024.125818

    arXiv 2025

  4. [1]

    X., Hu, Q

    Zhou, J., Zhao, J., Huang, J. X., Hu, Q. V ., & He, L. (2021). MASAD: A large-scale dataset for multimodal aspect-based sentiment analysis. Neurocomputing, 455, 47-58. https://doi.org/10.1016/j.neucom.2021.05.040

    work page doi:10.1016/j.neucom.2021.05.040 2021

  5. [2]

    H., Prasad, P

    Do, H. H., Prasad, P. W., Maag, A., & Alsadoon, A. (2019). Deep learning for aspect-based sentiment analysis: a comparative review. Expert systems with applications, 118, 272-299. https://doi.org/10.1016/j.eswa.2018.10.003

    doi:10.1016/j.eswa.2018.10.003 2019

  6. [3]

    Gandhi, A., Adhvaryu, K., Poria, S., Cambria, E., & Hussain, A. (2023). Multimodal sentiment analysis: A systematic review of history, datasets, multimodal fusion methods, applications, challenges and future directions. Information Fusion, 91, 424-444. https://doi.org/10.1016/j.inffus.2022.09.025

    doi:10.1016/j.inffus.2022.09.025 2023

  7. [11]

    S., Nahta, R., Meena, Y

    Chauhan, G. S., Nahta, R., Meena, Y . K., & Gopalani, D. (2023). Aspect based sentiment analysis using deep learning approaches: A survey. Computer Science Review, 49, 100576. https://doi.org/10.1016/j.cosrev.2023.100576

    work page arXiv 2023

  8. [12]

    Zou, W., Zhang, W., Tian, Z., & Wu, W. (2024). A syntactic features and interactive learning model for aspect- based sentiment analysis. Complex & Intelligent Systems, 10(4), 5359 -5377. https://doi.org/10.1007/s40747-024- 01449-5

    doi:10.1007/s40747-024- 2024

Show all 48 references
  1. [13]

    K., Jena, M., Rath, S

    Behera, R. K., Jena, M., Rath, S. K., & Misra, S. (2021). Co-LSTM: Convolutional LSTM model for sentiment analysis in social big data. Information Processing & Management, 58(1), 102435. https://doi.org/10.1016/j.ipm.2020.102435

    2021

  2. [14]

    Chen, Y . (2015). Convolutional neural network for sentence classification (Master's thesis, University of Waterloo). http://hdl.handle.net/10012/9592

    2015

  3. [15]

    Shang, W., Chai, J., Cao, J., Lei, X., Zhu, H., Fan, Y ., & Ding, W. (2024). Aspect-level sentiment analysis based on aspect -sentence graph convolution network. Information Fusion, 104, 102143. https://doi.org/10.1016/j.inffus.2023.102143

    2024

  4. [16]

    J., Dahlmeier, D., & Xiao, X

    Wang, W., Pan, S. J., Dahlmeier, D., & Xiao, X. (2017, February). Coupled multi -layer attentions for co - extraction of aspect and opinion terms. In Proceedings of the AAAI conference on artificial intelligence (V ol. 31, No. 1). https://dl.acm.org/doi/abs/10.5555/3298023.3298050

    2017

  5. [17]

    (2020, April)

    Peng, H., Xu, L., Bing, L., Huang, F., Lu, W., & Si, L. (2020, April). Knowing what, how and why: A near complete solution for aspect -based sentiment analysis. In Proceedings of the AAAI conference on artificial intelligence (V ol. 34, No. 05, pp. 8600-8607). https://doi.org/...

    2020 doi

  6. [18]

    (2023, July)

    Zhang, M., Zhu, Y ., Liu, Z., Bao, Z., Wu, Y ., Sun, X., & Xu, L. (2023, July). Span-level aspect-based sentiment analysis via table filling. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers) (pp. 9273-9284). ht...

    2023 doi

  7. [19]

    Zheng, Y ., Gong, J., Wen, Y ., & Zhang, P. (2024). DJMF: A discriminative joint multi-task framework for multimodal sentiment analysis based on intra -and inter -task dynamics. Expert Systems with Applications, 242, 122728. https://doi.org/10.1016/j.eswa.2023.122728

    2024

  8. [20]

    Zou, W., Zhang, W., Wu, W., & Tian, Z. (2024). A multi -task shared cascade learning for aspect sentiment triplet extraction using bert -mrc. Cognitive Computation, 16(4), 1554 -1571. https://doi.org/10.1007/s12559-024- 10247-7

    2024 doi

  9. [21]

    Liu, Z., Zhang, T., Yang, K., Thompson, P., Y u, Z., & Ananiadou, S. (2024). Emotion detection for misinformation: A review. Information Fusion, 107, 102300. https://doi.org/10.1016/j.inffus.2024.102300

    2024

  10. [22]

    Yue, T., Mao, R., Wang, H., Hu, Z., & Cambria, E. (2023). KnowleNet: Knowledge fusion network for multimodal sarcasm detection. Information Fusion, 100, 101921. https://doi.org/10.1016/j.inffus.2023.101921

    2023

  11. [23]

    (2015, October)

    You, Q., Luo, J., Jin, H., & Yang, J. (2015, October). Joint visual-textual sentiment analysis with deep neural networks. In Proceedings of the 23rd ACM international conference on Multimedia (pp. 1071 -1074). https://doi.org/10.1145/2733373.2806284

    2015

  12. [24]

    (2014, July)

    Wang, M., Cao, D., Li, L., Li, S., & Ji, R. (2014, July). Microblog sentiment analysis based on cross -media bag-of-words model. In Proceedings of international conference on internet multimedia computing and service (pp. 76-80). https://doi.org/10.1145/2632856.2632912

    2014

  13. [25]

    Huang, F., Zhang, X., Zhao, Z., Xu, J., & Li, Z. (2019). Image –text sentiment analysis via deep multimodal attentive fusion. Knowledge-Based Systems, 167, 26-37. https://doi.org/10.1016/j.knosys.2019.01.019

    2019 doi

  14. [28]

    Yu, J., Jiang, J., & Xia, R. (2019). Entity -sensitive attention and fusion network for entity -level multimodal sentiment classification. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28, 429 -439. https://doi.org/10.1109/TASLP.2019.2957872

    2019

  15. [29]

    C., & Y u, J

    Yang, L., Na, J. C., & Y u, J. (2022). Cross-modal multitask transformer for end -to-end multimodal aspect - based sentiment analysis. Information Processing & Management, 59(5), 103038. https://doi.org/10.1016/j.ipm.2022.103038

    2022

  16. [31]

    (2018, April)

    Zhang, Q., Fu, J., Liu, X., & Huang, X. (2018, April). Adaptive co -attention network for named entity recognition in tweets. In Proceedings of the AAAI conference on artificial intelligence (V ol. 32, No. 1). https://doi.org/10.1609/aaai.v32i1.11962

    2018 doi

  17. [32]

    (2018, July)

    Lu, D., Neves, L., Carvalho, V ., Zhang, N., & Ji, H. (2018, July). Visual attention model for name tagging in multimodal social media. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers) (pp. 1990-1999). https://...

    2018 doi

  18. [33]

    Hu, M., Peng, Y ., Huang, Z., Li, D., & Lv, Y . (2019). Open-domain targeted sentiment analysis via span-based extraction and classification. arXiv preprint arXiv:1906.03820. https://doi.org/10.48550/arXiv.1906.03820

    2019 doi arXiv

  19. [34]

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI blog, 1(8), 9. https://api.semanticscholar.org/CorpusID:160025533

    2019

  20. [35]

    (2020, December)

    Chen, G., Tian, Y ., & Song, Y . (2020, December). Joint aspect extraction and sentiment analysis with directional graph convolutional networks. In Proceedings of the 28th international conference on computational linguistics (pp. 272-279). https://doi.org/10.18653/v1/2020.col...

    2020 doi

  21. [36]

    & Stoyanov, V

    Liu, Y ., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., ... & Stoyanov, V . (2019). Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. https://doi.org/10.48550/arXiv.1907.11692

    2019 doi arXiv

  22. [37]

    & Zettlemoyer, L

    Lewis, M., Liu, Y ., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., ... & Zettlemoyer, L. (2019). Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461. https://doi.org/10.48550/...

    2019 doi arXiv

  23. [38]

    Yu, J., & Jiang, J. (2019). Adapting BERT for target -oriented multimodal sentiment classification. IJCAI. https://doi.org/10.24963/ijcai.2019/751

    2019 doi

  24. [39]

    Yu, J., Jiang, J., Yang, L., & Xia, R. (2020). Improving multimodal named entity recognition via entity span detection with unified multimodal transformer. Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.306

    2020 doi

  25. [40]

    F., & Li, Q

    Wu, Z., Zheng, C., Cai, Y ., Chen, J., Leung, H. F., & Li, Q. (2020, October). Multimodal representation with embedded visual guiding objects for named entity recognition in social media posts. In Proceedings of the 28th ACM International conference on multimedia (pp. 1038-104...

    2020

  26. [41]

    (2021, May)

    Sun, L., Wang, J., Zhang, K., Su, Y ., & Weng, F. (2021, May). RpBERT: a text-image relation propagation - based BERT model for multimodal NER. In Proceedings of the AAAI conference on artificial intelligence (V ol. 35, No. 15, pp. 13860-13868). https://doi.org/10.1609/aaai.v3...

    2021 doi

  27. [42]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S.,

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., ... & Sutskever, I. (2021, July). Learning transferable visual models from natural language supervision. In International conference on machine learning (pp. 8748-8763). PmLR. https://proceedings.mlr.press...

    2021

  28. [43]

    (2021, November)

    Ju, X., Zhang, D., Xiao, R., Li, J., Li, S., Zhang, M., & Zhou, G. (2021, November). Joint multi-modal aspect- sentiment analysis with auxiliary cross-modal relation detection. In Proceedings of the 2021 conference on empirical methods in natural language processing (pp. 4395-...

    2021 doi

  29. [44]

    (2021, October)

    Khan, Z., & Fu, Y . (2021, October). Exploiting BERT for multimodal target sentiment classification through input space translation. In Proceedings of the 29th ACM international conference on multimedia (pp. 3034 -3042). https://doi.org/10.1145/3474085.3475692

    2021

  30. [45]

    & Poria, S

    Yang, X., Feng, S., Wang, D., Qi, S., Wu, W., Zhang, Y ., ... & Poria, S. (2023). Few-shot joint multimodal aspect-sentiment analysis based on generative multimodal prompt. arXiv preprint arXiv:2305.10169. https://doi.org/10.48550/arXiv.2305.10169

    2023 doi arXiv

  31. [46]

    Zhao, F., Li, C., Wu, Z., Ouyang, Y ., Zhang, J., & Dai, X. (2023). M2DF: multi -grained multi-curriculum denoising framework for multimodal aspect -based sentiment analysis. arXiv preprint arXiv:2310.14605. https://doi.org/10.48550/arXiv.2310.14605

    2023 doi arXiv

  32. [47]

    Mu, J., Nie, F., Wang, W., Xu, J., Zhang, J., & Liu, H. (2023). Mocolnet: A momentum contrastive learning network for multimodal aspect -level sentiment analysis. IEEE Transactions on Knowledge and Data Engineering, 36(12), 8787-8800. https://doi.org/10.1109/TKDE.2023.3345022

    2023

  33. [48]

    (2024, July)

    Liu, H., He, L., & Liang, J. (2024, July). Joint Modal Circular Complementary Attention for Multimodal Aspect-Based Sentiment Analysis. In 2024 IEEE International Conference on Multimedia and Expo Workshops (ICMEW) (pp. 1-6). IEEE. https://doi.org/10.1109/ICMEW63481.2024.10645483

    2024

  34. [49]

    T., & Nguyen, C

    Van Doan, N., Nguyen, D. T., & Nguyen, C. V . T. (2024). A Dual -Module Denoising Approach with Curriculum Learning for Enhancing Multimodal Aspect -Based Sentiment Analysis. arXiv preprint arXiv:2412.08489. https://doi.org/10.48550/arXiv.2412.08489

    2024 doi arXiv

  35. [50]

    Zhang, J., Qu, J., Liu, J., & Wang, Z. (2024). MCPL: Multi -model co -guided progressive learning for multimodal aspect -based sentiment analysis. Knowledge -Based Systems, 301, 112331. https://doi.org/10.1016/j.knosys.2024.112331

    2024

  36. [51]

    Li, Y ., Ding, H., Lin, Y ., Feng, X., & Chang, L. (2024). Multi-level textual-visual alignment and fusion network for multimodal aspect -based sentiment analysis. Artificial Intelligence Review, 57(4), 78. https://doi.org/10.1007/s10462-023-10685-z

    2024 doi

  37. [52]

    Das, R., & Singh, T. D. (2023). Multimodal sentiment analysis: a survey of methods, trends, and challenges. ACM Computing Surveys, 55(13s), 1-38. https://doi.org/10.1145/3586075

    2023 doi

  38. [53]

    A., & Song, D

    Zhao, T., Meng, L. A., & Song, D. (2024). Multimodal aspect -based sentiment analysis: a survey of tasks, methods, challenges and future directions. Information Fusion, 112, 102552. https://doi.org/10.1016/j.inffus.2024.102552

    2024

  39. [54]

    Zhang, W., Li, X., Deng, Y ., Bing, L., & Lam, W. (2022). A survey on aspect-based sentiment analysis: Tasks, methods, and challenges. IEEE Transactions on Knowledge and Data Engineering, 35(11), 11019 -11038. https://doi.org/10.1109/TKDE.2022.3230975

    2022

  40. [55]

    Brauwers, G., & Frasincar, F. (2022). A survey on aspect -based sentiment classification. ACM Computing Surveys, 55(4), 1-37. https://doi.org/10.1145/3503044

    2022 doi

Pith tools

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