Pith. sign in

REVIEW 4 major objections 5 minor 34 references

RAMer: Reconstruction-based Adversarial Model for Multi-party Multi-modal Multi-label Emotion Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read RAMer claims state-of-the-art performance on multi-party and dyadic multi-modal multi-label emotion recognition.

desk verdict A plausible and well-engineered MMER framework with strong reported numbers, but the stack shuffle augmentation is under-specified and never ablated, and one results claim contradicts its own table. read the letter →

arxiv 2502.10435 v2 pith:VQP4PF36 submitted 2025-02-09 cs.CV cs.AI

Xudong Yang , Yizhang Zhu , Hanfeng Liu , Zeyi Wen , Nan Tang , Yuyu Luo This is my paper · ORCID
classification cs.CVcs.AI
keywords multi-partyemotionrecognitionmulti-modalmulti-labelincompletemodalitiesadversarialtrainingfeaturereconstructioncontrastivelearningpersonality-awareattentionstackshuffleaugmentation
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 claims that multi-party multi-modal multi-label emotion recognition can be pushed past current results by explicitly separating each modality's shared and unique information, reconstructing missing features, and enriching the feature space with a shuffle augmentation. It introduces RAMer, which combines adversarial training with a reconstruction network and contrastive learning, plus an auxiliary personality task and a stack shuffle strategy. On MEmoR, CMU-MOSEI, and M3ED, the paper reports state-of-the-art accuracy and Micro-F1, including gains in multi-party settings where non-speakers lack acoustic and textual inputs. A sympathetic reader would care because real conversations rarely give every participant complete signals, and the paper's approach is designed for exactly that incompleteness.

What carries the argument

The load-bearing mechanism is the reconstruction-based adversarial learning module: a generator $G$ projects reconstructed features into a shared subspace while a discriminator $D$ tries to identify the source modality, forcing the commonality representations to be modality-agnostic, while specificity encoders keep modality-unique information and an orthogonality constraint separates the two. A second-level reconstruction pathway rebuilds each modality from the others using intrinsic vectors from supervised contrastive learning, and the stack shuffle operation augments the commonality and specificity features before the final binary cross-entropy classifier.

What would settle it

Remove the stack shuffle loss $L^{suf}_{cls}$ from the final objective and retrain on the same aligned CMU-MOSEI protocol; if Micro-F1 stays at or above 0.604, the shuffle component is unnecessary for the claimed state-of-the-art result.

Watch

Extended reading notes

Core claim

RAMer's central claim is that a reconstruction-based adversarial framework can simultaneously preserve modality-specific information, extract cross-modal commonality, and recover missing modalities, leading to better multi-label emotion predictions than prior fusion methods. The model encodes each modality separately, uses a generator-discriminator pair to align common subspaces while an orthogonal loss keeps specificity distinct, and reconstructs each modality from the others with contrastively learned intrinsic vectors. Personality embeddings are attached through modality-level attention to compensate for missing inputs, and a stack shuffle operation rearranges samples and modalities to strengthen label-feature correlations. The paper reports that this full system outperforms prior methods on MEmoR, CMU-MOSEI, and M3ED.

Load-bearing premise

The stack shuffle augmentation is assumed to enrich label-modality correlations without destroying the true feature-to-label correspondence, and the paper never ablates it, so the reported gains could partly ride on that untested assumption.

Editorial extensions

If this is right

  • On the MEmoR dataset, RAMer reports Micro-F1 of 0.499, Macro-F1 of 0.402, and Weighted-F1 of 0.503 in the primary setting, above all listed baselines including AMER with personality.
  • On CMU-MOSEI, RAMer reports the best accuracy and Micro-F1 in both aligned and unaligned settings, with Micro-F1 0.604 (aligned) and 0.560 (unaligned).
  • On M3ED, RAMer matches or exceeds the best prior result across accuracy, precision, recall, and Micro-F1.
  • The personality auxiliary task and modality-level attention are claimed to be especially useful when non-speakers lack acoustic and textual inputs, allowing inter-person context to fill in missing signals.
  • The paper's ablation study shows that removing adversarial training, the reconstruction network, or contrastive learning each lowers Micro-F1, supporting the claim that these parts matter.

Reading between the lines

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

  • Editorial inference: The stack shuffle loss in Eq. (12) is not isolated in the ablation table, so a natural follow-up is to retrain with that loss removed to see how much of the reported gain comes from the shuffle rather than from the reconstruction and adversarial components.
  • Editorial inference: Because personality labels are only available on MEmoR, applying RAMer to datasets without personality would require a surrogate trait estimator; the paper does not report such a variant.
  • Editorial inference: The same reconstruction-plus-commonality design could transfer to other incomplete-modality tasks such as sentiment or depression detection in group conversations, where non-speakers also lack audio and text.
  • Editorial inference: The t-SNE analysis suggests the model changes representation geometry, which is a qualitative claim; a quantitative test would be to measure label separability (for example, nearest-neighbor purity) before and after reconstruction training.
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

4 major / 5 minor

Summary. The paper proposes RAMer, a framework for multi-party multi-modal multi-label emotion recognition (MMER) that combines reconstruction-based adversarial learning, contrastive learning, a personality auxiliary task, and a stack shuffle augmentation strategy. The model is evaluated on MEmoR, CMU-MOSEI, and M3ED, reporting state-of-the-art performance in both dyadic and multi-party scenarios.

Significance. If the reported results hold, the paper addresses a practically important gap: MMER under incomplete modality conditions in multi-party conversations. The proposed reconstruction-based adversarial architecture is a reasonable and potentially useful contribution, and the inclusion of the auxiliary personality task is interesting. The paper also provides extensive comparisons and an ablation study, which strengthens the empirical narrative. However, the central state-of-the-art claim is currently undermined by missing ablations, a contradictory textual claim about recall, and lack of statistical significance testing. The stack shuffle component, one of the three named contributions, is never isolated in the ablation study, making its actual contribution unclear.

major comments (4)
  1. [§4.3 and Eq. (12)] The stack shuffle description specifies only how feature vectors are reordered, not whether the label vector Y is permuted consistently. If labels are not shuffled, Eq. (12) trains the classifier on mismatched feature-label pairs, so the gradient is not a valid estimate of emotion-recognition risk. If labels are shuffled with the features, the classifier is trained on arbitrary cross-sample pairs that do not occur at inference, creating a train-test distribution mismatch. Moreover, Table 4 contains no ablation that removes or modifies the stack shuffle, so the contribution of Lsuf_cls to the final objective in Eq. (13) is empirically untested.
  2. [§5.2 and Table 2] The text states that RAMer surpasses state-of-the-art methods on all metrics except recall, but Table 2 shows that RAMer achieves the highest recall in both the aligned (0.551 vs. 0.518 for CARAT) and unaligned (0.486 vs. 0.466) settings. This direct contradiction between the reported results and the textual claim needs to be corrected.
  3. [Tables 1–3] All reported numbers are point estimates without error bars or significance tests. On M3ED, the improvement over CARAT is 0.001 in accuracy, 0.002 in precision, 0.004 in recall, and 0.003 in Micro-F1. These differences are within typical run-to-run variance, so the state-of-the-art claim is not yet statistically supported. The authors should report multiple runs with standard deviations or perform significance tests.
  4. [§4 and §5] The paper motivates the problem by incomplete modalities, but it does not specify how missing inputs are represented during training and inference, nor how the reconstruction module is used at test time. The case study in Figure 5 illustrates missing modalities, but the experimental protocol for handling them is not described. This omission hinders reproducibility of the claimed robustness in multi-party scenarios.
minor comments (5)
  1. [§4.2, Eq. (1)–(2)] The notation U ∈ {U^v, U^t, U^a} is ambiguous: it is described as the ground truth label for the discriminator, but it is unclear whether this is a modality indicator or an emotion label. Please clarify the definition and shape of U.
  2. [§4.2, Eq. (9)] The prototype vector is introduced as μ^m_{j,k} but used as u^m_{j,k} in the exponent; please make the notation consistent.
  3. [§5.4 and Figure 3] The t-SNE visualizations are qualitative and would benefit from a stated criterion for separation; also, the figure caption uses lowercase 'figure' in the text when referring to it.
  4. [References] References [Liu et al., 2018a] and [Liu et al., 2018b] are the same paper (LMF); please merge or differentiate.
  5. [§5.1] The datasets section says MEmoR includes personality, while CMU-MOSEI and M3ED do not, but it is not explained how the personality auxiliary task is used for datasets without personality annotations; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RAMer's state-of-the-art claim rests on external benchmark comparisons, and its losses are standard training objectives rather than fitted inputs renamed as predictions.

full rationale

The paper makes an empirical performance claim: 'RAMer achieves state-of-the-art performance in dyadic and multi-party MMER scenarios' on MEmoR, CMU-MOSEI, and M3ED. This claim is supported by Tables 1-3, which compare RAMer against externally published baselines on standard dataset protocols, so the central result is not derived from the paper's own assumptions. The method's components (adversarial commonality/specificity losses, contrastive learning, reconstruction, and stack shuffle) are training objectives; none of them is a fitted constant that is later relabeled as a prediction. The prototype vectors in Eq. (9) are online moving averages used inside the contrastive/reconstruction objective, not calibrated inputs whose output is then reported as a forecast of the same data. No load-bearing premise is justified only by a self-citation: the references to prior work (e.g., MEmoR, M3ED, CMU-MOSEI, transformer, contrastive learning) are external and independent. No 'uniqueness theorem' is imported from the authors' own prior work, and no known empirical pattern is renamed as a new theoretical result. The reader's concern about stack shuffle (Eq. 12) is a reproducibility and ablation-coverage issue: the shuffle step does not state explicitly whether labels are permuted consistently, and Table 4 never abates this component. That is a missing-evidence or specification problem, not a circularity problem, because it does not make the reported improvement reduce by construction to the paper's own inputs. Overall, the derivation chain is self-contained and externally falsifiable, so the appropriate circularity score is 0.

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

The model introduces no new physical or conceptual entities. It uses learned prototype vectors and internal representations, but these are ordinary model parameters, not externally postulated entities. The free parameters are mostly unstated hyperparameters, which limits the reader's ability to reproduce the exact results.

free parameters (4)
  • loss_trade_off_weights = not reported
    Equations 5, 8, and 13 contain weights lambda_a, lambda_o, lambda_c, lambda_alpha, lambda_beta, lambda_gamma, lambda_r, lambda_s that balance adversarial, reconstruction, contrastive, and classification losses; their values are not given, and the reported results likely depend on tuning them on validation.
  • contrastive_temperature_eta = not reported
    The supervised contrastive loss (Eq 11) uses a temperature eta that is not specified.
  • stack_count_k = not reported
    The stack shuffle splits features into k stacks (Section 4.3), but k is not specified.
  • model_dimensions_and_layers = not reported
    Transformer encoder layers, MLP depths and widths, and latent dimensions are not reported for any dataset.
assumptions (4)
  • domain assumption Benchmark datasets and splits provide reliable emotion labels and are comparable across methods.
    The evaluation claims state-of-the-art performance against published numbers on MEmoR, CMU-MOSEI, and M3ED; this assumes the standard protocols are used correctly.
  • domain assumption Personality annotations in MEmoR are accurate and act as complementary information rather than a leaking shortcut.
    The personality auxiliary task concatenates personality embeddings in Section 4.1; if these annotations correlate strongly with emotion labels, the gain on MEmoR could be inflated.
  • ad hoc to paper The adversarial objective (Eq 5) leads to a useful separation of commonality and specificity without discarding label-relevant information.
    This is a heuristic design choice; the paper provides only t-SNE visualizations as evidence, no quantitative measure of separation quality.
  • domain assumption The reconstruction loss (Eq 7) trained on available modalities transfers to the recovery of completely absent modalities.
    The paper never explicitly masks modalities during training to simulate absences on CMU-MOSEI and M3ED, so the recovery mechanism's behavior on missing inputs is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAMer: Reconstruction-based Adversarial Model for Multi-party Multi-modal Multi-label Emotion Recognition." pith.science (2026). https://pith.science/paper/VQP4PF36

@misc{pith2026250210435,
  author       = {Pith},
  title        = {Pith review of: RAMer: Reconstruction-based Adversarial Model for Multi-party Multi-modal Multi-label Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VQP4PF36}},
  note         = {Machine review of arXiv:2502.10435}
}
abstract

Conventional Multi-modal multi-label emotion recognition (MMER) assumes complete access to visual, textual, and acoustic modalities. However, real-world multi-party settings often violate this assumption, as non-speakers frequently lack acoustic and textual inputs, leading to a significant degradation in model performance. Existing approaches also tend to unify heterogeneous modalities into a single representation, overlooking each modality's unique characteristics. To address these challenges, we propose RAMer (Reconstruction-based Adversarial Model for Emotion Recognition), which refines multi-modal representations by not only exploring modality commonality and specificity but crucially by leveraging reconstructed features, enhanced by contrastive learning, to overcome data incompleteness and enrich feature quality. RAMer also introduces a personality auxiliary task to complement missing modalities using modality-level attention, improving emotion reasoning. To further strengthen the model's ability to capture label and modality interdependency, we propose a stack shuffle strategy to enrich correlations between labels and modality-specific features. Experiments on three benchmarks, i.e., MEmoR, CMU-MOSEI, and $M^3ED$, demonstrate that RAMer achieves state-of-the-art performance in dyadic and multi-party MMER scenarios.

Figures

Figures reproduced from arXiv: 2502.10435 by the authors.

Figure 1
Figure 1. (a) shows the conventional approach for MMER in mono [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of RAMer. Given incomplete multi-modal inputs, RAMer first encodes each individual modality through an [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. t-SNE visualizations of modality embeddings. (a)(b): Specificity and commonality features without/with adversarial training. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: shows an example of MMER on the MEmoR dataset where specific target persons have incomplete modality sig￾nals. The top three rows display different modalities from a video clip, segmented semantically with aligned multi￾modal signals. Key observations include: 1) The t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 22 canonical work pages

  1. [1]

    Learning multi-label scene classification

    [Boutell et al., 2004] Matthew R Boutell, Jiebo Luo, Xipeng Shen, and Christopher M Brown. Learning multi-label scene classification. Pattern recognition , 37(9):1757– 1771,

    work page 2004

  2. [7]

    Gen- erative adversarial nets

    [Goodfellow et al., 2014] Ian Goodfellow, Jean Pouget- Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen- erative adversarial nets. Advances in neural information processing systems, 27,

    2014

  3. [10]

    Fusenet: Incorporating depth into semantic segmentation via fusion-based cnn ar- chitecture

    [Hazirbas et al., 2017] Caner Hazirbas, Lingni Ma, Csaba Domokos, and Daniel Cremers. Fusenet: Incorporating depth into semantic segmentation via fusion-based cnn ar- chitecture. In Computer Vision–ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, Novem- ber 20-24, 2016, Revised Selected Papers, Part I 13, pages 213–228. Springer,

    work page 2017

  4. [13]

    Supervised Con- trastive Learning

    [Khosla et al., 2020] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Con- trastive Learning. arXiv e-prints, page arXiv:2004.11362, April

    arXiv 2020

  5. [15]

    Efficient low-rank multimodal fusion with modality-specific factors

    [Liu et al., 2018b] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv preprint arXiv:1806.00064,

    arXiv

  6. [16]

    Progressive modality re- inforcement for human multimodal emotion recognition from unaligned multimodal sequences

    [Lv et al., 2021] Fengmao Lv, Xiang Chen, Yanyong Huang, Lixin Duan, and Guosheng Lin. Progressive modality re- inforcement for human multimodal emotion recognition from unaligned multimodal sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 2554–2562,

    work page 2021

  7. [17]

    Label-specific dual graph neural network for multi-label text classification

    [Ma et al., 2021] Qianwen Ma, Chunyuan Yuan, Wei Zhou, and Songlin Hu. Label-specific dual graph neural network for multi-label text classification. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 3855–3864,

    work page 2021

  8. [18]

    Multimodality representation learning: A survey on evolution, pretraining and its applications

    [Manzoor et al., 2023] Muhammad Arslan Manzoor, Sarah Albarri, Ziting Xian, Zaiqiao Meng, Preslav Nakov, and Shangsong Liang. Multimodality representation learning: A survey on evolution, pretraining and its applications. ACM Transactions on Multimedia Computing, Communi- cations and Applications, 20(3):1–34,

    work page 2023

Show all 34 references
  1. [19]

    Adversarial training methods for semi-supervised text classification

    [Miyato et al., 2016] Takeru Miyato, Andrew M Dai, and Ian Goodfellow. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725,

    2016 arXiv

  2. [21]

    Classifier chains for multi-label classification

    [Read et al., 2011] Jesse Read, Bernhard Pfahringer, Geoff Holmes, and Eibe Frank. Classifier chains for multi-label classification. Machine learning, 85:333–359,

    2011

  3. [22]

    Towards emotion-aided multi-modal dialogue act classification

    [Saha et al., 2020] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In Proceedings of the 58th Annual Meeting of the Association for Computa- tional Linguistics, pages 4361–4372,

    2020

  4. [23]

    Memor: A dataset for multimodal emotion reasoning in videos

    [Shen et al., 2020] Guangyao Shen, Xin Wang, Xuguang Duan, Hongzhi Li, and Wenwu Zhu. Memor: A dataset for multimodal emotion reasoning in videos. In Proceedings of the 28th ACM international conference on multimedia , pages 493–502,

    2020

  5. [24]

    Order-free learning alleviating exposure bias in multi-label classification

    [Tsai and Lee, 2020] Che-Ping Tsai and Hung-Yi Lee. Order-free learning alleviating exposure bias in multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6038–6045,

    2020

  6. [25]

    Learning factorized multimodal represen- tations

    [Tsai et al., 2018] Yao-Hung Hubert Tsai, Paul Pu Liang, Amir Zadeh, Louis-Philippe Morency, and Ruslan Salakhutdinov. Learning factorized multimodal represen- tations. arXiv preprint arXiv:1806.06176,

    2018 arXiv

  7. [26]

    Multimodal transformer for un- aligned multimodal language sequences

    [Tsai et al., 2019] Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for un- aligned multimodal language sequences. In Proceedings of the conference. Association for computational linguis-...

    2019

  8. [27]

    Attention is all you need

    [Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30,

    2017

  9. [28]

    Gen- erative adversarial networks: introduction and outlook

    [Wang et al., 2017] Kunfeng Wang, Chao Gou, Yanjie Duan, Yilun Lin, Xinhu Zheng, and Fei-Yue Wang. Gen- erative adversarial networks: introduction and outlook. IEEE/CAA Journal of Automatica Sinica , 4(4):588–598,

    2017

  10. [29]

    Adversarial training for relation extraction

    [Wu et al., 2017] Yi Wu, David Bamman, and Stuart Russell. Adversarial training for relation extraction. InProceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1778–1783,

    2017

  11. [30]

    Cross-modality attention with semantic graph embedding for multi-label classification

    [You et al., 2020] Renchun You, Zhiyao Guo, Lei Cui, Xi- ang Long, Yingze Bao, and Shilei Wen. Cross-modality attention with semantic graph embedding for multi-label classification. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12709–12716,

    2020

  12. [31]

    Tensor fusion network for multimodal sentiment analysis

    [Zadeh et al., 2017] Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Tensor fusion network for multimodal sentiment analysis. arXiv preprint arXiv:1707.07250,

    2017 arXiv

  13. [32]

    Multimodal language analysis in the wild: Cmu- mosei dataset and interpretable dynamic fusion graph

    [Zadeh et al., 2018] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu- mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association...

    2018

  14. [33]

    Multi- modal multi-label emotion detection with modality and label dependence

    [Zhang et al., 2020] Dong Zhang, Xincheng Ju, Junhui Li, Shoushan Li, Qiaoming Zhu, and Guodong Zhou. Multi- modal multi-label emotion detection with modality and label dependence. In Proceedings of the 2020 confer- ence on empirical methods in natural language processing (EMN...

    2020

  15. [34]

    Tailor versatile multi-modal learning for multi-label emotion recognition

    [Zhang et al., 2022] Yi Zhang, Mingyuan Chen, Jundong Shen, and Chongjun Wang. Tailor versatile multi-modal learning for multi-label emotion recognition. arXiv e- prints, page arXiv:2201.05834, January

    2022 arXiv

  16. [2004]

    Multi-label image recognition with graph convolutional networks

    [Chen et al., 2019] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5177–5186,

    2019

  17. [2011]

    CARAT: Contrastive Feature Reconstruction and Aggregation for Multi-Modal Multi-Label Emotion Recognition

    [Peng et al., 2023] Cheng Peng, Ke Chen, Lidan Shou, and Gang Chen. CARAT: Contrastive Feature Reconstruction and Aggregation for Multi-Modal Multi-Label Emotion Recognition. arXiv e-prints, page arXiv:2312.10201, De- cember

    2023 arXiv

  18. [2013]

    Learning robust multi-modal representation for multi-label emotion recog- nition via adversarial masking and perturbation

    [Ge et al., 2023] Shiping Ge, Zhiwei Jiang, Zifeng Cheng, Cong Wang, Yafeng Yin, and Qing Gu. Learning robust multi-modal representation for multi-label emotion recog- nition via adversarial masking and perturbation. In Pro- ceedings of the ACM Web Conference 2023, pages 1510– 1518,

    2023

  19. [2014]

    Conversational memory net- work for emotion recognition in dyadic dialogue videos

    [Hazarika et al., 2018] Devamanyu Hazarika, Soujanya Po- ria, Amir Zadeh, Erik Cambria, Louis-Philippe Morency, and Roger Zimmermann. Conversational memory net- work for emotion recognition in dyadic dialogue videos. In Proceedings of the conference. Association for Com- putat...

    2018

  20. [2017]

    DialogueCRN: Contextual Reasoning Networks for Emo- tion Recognition in Conversations

    [Hu et al., 2021] Dou Hu, Lingwei Wei, and Xiaoyong Huai. DialogueCRN: Contextual Reasoning Networks for Emo- tion Recognition in Conversations. arXiv e-prints, page arXiv:2106.01978, June

    2021 arXiv

  21. [2018]

    Misa: Modality-invariant and-specific representations for multimodal sentiment analysis

    [Hazarika et al., 2020] Devamanyu Hazarika, Roger Zim- mermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM international conference on multimedia, pages 1122–1131,

    2020

  22. [2019]

    A simple framework for contrastive learning of visual representations

    [Chen et al., 2020] Ting Chen, Simon Kornblith, Moham- mad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In In- ternational conference on machine learning, pages 1597–

    2020

  23. [2020]

    Robust bloom fil- ters for large multilabel classification tasks

    [Cisse et al., 2013] Moustapha M Cisse, Nicolas Usunier, Thierry Artieres, and Patrick Gallinari. Robust bloom fil- ters for large multilabel classification tasks. Advances in neural information processing systems, 26,

    2013

  24. [2021]

    Audio- oriented multimodal machine comprehension via dynamic inter-and intra-modality attention

    [Huang et al., 2021] Zhiqi Huang, Fenglin Liu, Xian Wu, Shen Ge, Helin Wang, Wei Fan, and Yuexian Zou. Audio- oriented multimodal machine comprehension via dynamic inter-and intra-modality attention. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35,...

    2021

  25. [2022]

    M3ed: Multi-modal multi-scene multi-label emotional di- alogue database

    [Zhao et al., 2022] Jinming Zhao, Tenggan Zhang, Jingwen Hu, Yuchen Liu, Qin Jin, Xinchao Wang, and Haizhou Li. M3ed: Multi-modal multi-scene multi-label emotional di- alogue database. arXiv preprint arXiv:2205.10237, 2022

    2022 arXiv

  26. [2023]

    Dialoguegcn: A graph convolutional neural network for emotion recognition in conversation

    [Ghosal et al., 2019] Deepanway Ghosal, Navonil Ma- jumder, Soujanya Poria, Niyati Chhaya, and Alexander Gelbukh. Dialoguegcn: A graph convolutional neural network for emotion recognition in conversation. arXiv preprint arXiv:1908.11540,

    2019 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.