Pith. sign in

REVIEW 5 major objections 4 minor 41 references

Missing modalities should not always be repaired: a per-sample oracle shows full-modality input is best for only ~15–33% of samples, and the SIEVE gate learns when to skip repair, improving three backbones and nearing the per-sample optimum

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

SIEVE learns a per-sample 'should we repair?' decision from the loss gap between direct and repair branches, improving three missing-modality MSA backbones on CMU-MOSI and IEMOCAP.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection SIEVE's 'whether to repair' question is the right one, but the paper's own Table 2 contradicts 'consistently improves,' and the evidential gate's training signal is a non-stationary loss gap that lacks stability analysis. the 5 major comments →

arxiv 2607.17262 v1 pith:2RDWXVYO submitted 2026-07-19 cs.CL

Should Missing Modalities Always Be Necessary to Repair for Multi-modal Sentiment Analysis?

Yubo Gao , Haotian Wu , Xiaoyu Xu , Yibo Yan , Hong Chen , Ruoshui Peng , Fei Pan , Puay Siew Tan
show 4 more authors
Zhuoran Gao Yonghua Hei Jie Zhang Xuming Hu
This is my paper
classification cs.CL
keywords multimodal sentiment analysismissing modalitiesmodality repairsample-dependent modality utilityevidential deep learninguncertainty-aware routingselective repairoracle analysis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 challenges the repair-first assumption in multimodal sentiment analysis: when a modality is missing, most systems try to reconstruct or compensate for it before predicting. A per-sample oracle analysis on two standard benchmarks shows this default is wrong for a majority of samples — full-modality input is the best choice for only 15.5% (MISA) and 33.2% (GCNet) of test samples, and every modality subset is preferred by some samples. The authors propose SIEVE, a plug-and-play gate that compares a direct prediction branch with a repair branch, uses their per-sample loss gap as a self-supervised sufficiency signal, and routes each input through an evidential Beta gate that accounts for uncertainty and the missing pattern. On CMU-MOSI and IEMOCAP, SIEVE consistently improves three repair backbones (explicit, distillation-based, and evidence-compensation) across missing rates, and closes most of the gap to the per-sample dual-branch optimum. If correct, the work reframes missing-modality MSA from 'how to repair' to 'whether to repair' at the sample level.

Core claim

On its own terms, the paper establishes that modality utility is sample-dependent: for a substantial fraction of inputs, the observed modalities already carry enough evidence for accurate prediction, and forcing repair adds noise. The discovery is operationalized as an explicit per-sample routing decision: SIEVE trains a direct branch and a repair branch, defines empirical sufficiency as 'direct loss ≤ repaired loss' for that sample, and supervises an evidential valve with the stop-gradient loss gap. The valve predicts a Beta distribution over sufficiency, and routes hard at inference using a dynamic threshold that combines epistemic uncertainty with a missing-pattern base rate maintained as

What carries the argument

The load-bearing mechanism is the dual-branch loss gap Ai = sg(Lr,i − Ld,i) between the repair and direct branches, converted into a soft sufficiency target s⋆i = σ(Ai/η). This self-supervised signal, with stop-gradient, trains an evidential sufficiency valve — a Beta parameterization over the probability that the observed modalities suffice — whose posterior mean ¯si = αi/(αi + βi) is compared against a dynamic threshold θ(ui) = (1 + (1 − 2πm)ui)/2, where ui is epistemic uncertainty and πm is the online-estimated base rate for the sample's missing pattern. The threshold is the exact decision boundary of a Subjective Logic projection; training uses Gumbel-Sigmoid relaxation and a two-stage p

Load-bearing premise

The whole signal is the per-sample loss gap between two branches being trained simultaneously; if that gap keeps moving as the branches co-evolve, the sufficiency target that supervises the valve is not stable enough to learn a reliable routing decision.

What would settle it

Train SIEVE with the valve supervised by the oracle best-branch label (or by a fixed, pretrained pair of branches) and compare routing agreement with the loss-gap-supervised version on the test set. If the loss-gap valve shows near-chance agreement with the oracle or fails to beat always-repair on held-out missing patterns, the co-evolved loss gap is not a reliable sufficiency signal.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the oracle analysis generalizes, missing-modality MSA systems should include a routing decision instead of assuming repair-first; a default repair policy misallocates computation and can hurt accuracy.
  • SIEVE's gains hold at all evaluated missing rates, including MR = 0.0, meaning selective routing can suppress unnecessary repair even when all modalities are present.
  • The plug-and-play nature means any existing explicit or implicit repair backbone can be wrapped by the gate without internal modification, so the result is a drop-in upgrade for current repair-based MSA models.
  • Because the sufficiency signal is self-supervised, it can be applied in settings where complete-modality supervision is unavailable, unlike distillation-based methods that need full-modality teachers.
  • The approach approaches (but does not match) the per-sample dual-branch achievable optimum, indicating the remaining gap is a router-quality problem rather than a representation problem.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond sentiment analysis, the 'whether to repair' principle should transfer to any multimodal task with missing inputs (VQA, action recognition, medical fusion); the same dual-branch loss-gap gate could be attached to unimodal or multimodal predictors, though the oracle proportions will differ by task.
  • A testable extension: use the valve's predicted sufficiency to skip forwarding through the repair branch at inference entirely, using only the direct branch when ¯si is high — SIEVE currently evaluates both branches during training and the repair branch at inference only when routed; a learned predictor of the loss gap could amortize the cost.
  • The oracle analysis itself suggests a diagnostic tool: for any dataset, plotting the best-modality-subset distribution per sample reveals how text-dominant or redundancy-heavy the data is; this could guide modality engineering before model training.
  • If the loss gap is too noisy when branches co-evolve, a robust variant could anneal the stop-gradient or use a lagged copy of the repair branch to compute the gap; this would test the stability assumption underlying the valve.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper challenges the repair-first paradigm in multimodal sentiment analysis under missing modalities. It presents a per-sample oracle analysis on CMU-MOSI showing that the best modality subset varies by sample, then proposes SIEVE, a plug-and-play module that learns whether to invoke an existing repair module for each input. SIEVE computes an empirical sufficiency target from the per-sample loss gap between direct and repair branches, models sufficiency as a Beta distribution with evidential uncertainty, and routes samples through a dynamic threshold. Experiments on CMU-MOSI and IEMOCAP compare SIEVE against and on top of three repair backbones (GCNet, CorrKD, MoMKE) across missing rates. The central claims are that SIEVE consistently improves these backbones and approaches the per-sample achievable optimum.

Significance. If the claims hold, the paper makes a useful conceptual contribution: it reframes missing-modality robustness as a sample-level decision problem rather than a universal repair obligation, and the SIEVE mechanism is repair-agnostic. The oracle analysis (Figure 1) is a simple but effective diagnostic that supports the sample-dependent utility of modalities. The evidential formulation for uncertainty-aware routing is elegant and the Appendix A derivation of the threshold is clean. However, the paper's own data contradict the 'consistently improves' claim, and the lack of significance testing plus the non-stationary target issue leave the empirical case incomplete. The idea is promising and in-scope for cs.CL, but the current presentation overstates the evidence.

major comments (5)
  1. [§4.2, Table 2 / Table 5] The abstract and §4.2 claim that SIEVE 'consistently improves' all three backbones at every missing rate. This is contradicted by Table 2: CorrKD+SIEVE on CMU-MOSI at MR=0.2 has Acc-2 0.7500 vs CorrKD 0.7546 (F1 0.7515 vs 0.7551, Corr 0.7090 vs 0.7193 in Table 5). This single counterexample invalidates the literal 'every missing rate' claim. The authors must either restrict the claim to average or 'nearly all' settings, and provide significance tests to support the improvements where they do occur. As written, a headline claim is false.
  2. [§4.1, Table 2] All results are five-seed averages with no standard deviations or significance tests. Many differences are small (e.g., GCNet+SIEVE at MR=0.3 on CMU-MOSI: 0.8201 vs 0.8171) and could easily be noise. Without variance estimates or pairwise significance tests, 'consistently improves' is not supported even for the cells that do improve. Please report std/confidence intervals and run paired tests (or at least report effect sizes) for the key comparisons.
  3. [§3.3, §3.4, Eqs. (8)–(9) and (20)] The sufficiency target s*_i is computed from the per-sample loss gap between two branches that are jointly trained with the valve. The stop-gradient in Eq. (8) prevents the valve loss from directly using the target, but the routing weight g_i in Eq. (20) updates both branches, which in turn changes L_d,i and L_r,i and hence moves s*_i. The valve is thus chasing a moving and self-referential target. The two-stage warmup mitigates initial instability but does not prevent drift during joint training. The paper provides no analysis of target stability or sensitivity to η, κ, or the EMA rate. Please add a diagnostic (e.g., correlation between early s* and final routing decisions, or comparison with a frozen-branch target) to show the valve is learning a stable sufficiency signal rather than fitting noise.
  4. [§4.2, Figure 3] The 'Achievable Optimum' is computed from test-time per-sample losses of the two branches as trained under the SIEVE routing policy. Because the routing policy affects branch training, this oracle is not an independent upper bound; it is an in-sample descriptive reference. The claim that SIEVE 'approaches the per-sample achievable optimum' is therefore weaker than it appears. Please either compute the oracle from independently trained direct/repair branches (a true upper bound) or explicitly relabel it as a 'descriptive best-branch reference' and adjust the text accordingly.
  5. [§4.1, missing-rate protocol] The text states 'M R = 1 − (Σ c_i)/(3L)' with c_i ≥ 1, and then says 'hence M R ≤ 2/3 ≈ 0.67; M R = 0.7 approximates this limit'. This is internally inconsistent: if every sample has at least one observed modality, the maximum possible MR under this definition is 2/3, so MR=0.7 cannot occur. Either the protocol is mis-specified or MR is used as a nominal drop probability before the conditioning, in which case the text is wrong about the upper bound. Since all extreme-setting experiments (Table 1, Table 2, Table 3) use MR=0.7, this needs clarification. Please define the protocol precisely and reconcile the numbers.
minor comments (4)
  1. [Table 4] The table lists 'κ' as the 'soft sufficiency-label temperature', but Eq. (9) names this temperature 'η', while κ is used in Eq. (15) as the EMA update rate. Please align the notation.
  2. [§4.2, Figure 3] The subplot labels (e.g., '4.17%', '3.87%') are not explained in the text. State explicitly whether these are the gap between SIEVE and the Achievable Optimum, and clarify what 'Repair' denotes in the lower row.
  3. [Eq. (22)] The description 'mask out the target-aligned evidence and retain only the wrong-direction component' is accurate, but the equations for ilde α and ilde β could be annotated more directly (e.g., 'when s*=1, ilde β=β_i') to help readers see that only the contradicting evidence survives.
  4. [§1, Figure 1] The figure in the text is hard to parse because the two rows of numbers are not clearly associated with 'GCNet' and 'MISA'. Please label the bars directly or use a legend.

Circularity Check

2 steps flagged

SIEVE's 'achievable optimum' and its sufficiency label both derive from the same dual-branch loss gap; reported closeness to the optimum is partly by construction, while the main benchmark gains remain external.

specific steps
  1. self definitional [Section 4.2, 'Anatomy of SIEVE's Routing Behavior' (Figure 3)]
    "As an upper reference, we compute the Achievable Optimum, which selects, for each sample, the branch with the smaller per-sample loss; this denotes the highest accuracy reachable by any binary router restricted to the two branches."

    The 'Achievable Optimum' is exactly the per-sample argmin of {L_d,i, L_r,i}, i.e. the sign of the loss gap A_i = sg(L_r,i - L_d,i) used in Eq. (8). That same gap is converted by Eq. (9) into the valve's soft training target s*_i = sigma(A_i/eta). Thus the red-dashed upper bound in Figure 3 is constructed from the same quantity used to supervise the valve (Eq. 21). The claim that SIEVE 'approaches the Achievable Optimum' therefore partly reduces to checking whether the gate reproduces its own training signal, rather than comparing against an externally fixed ceiling. The bound is also not independent of the routing policy, because the branch losses it is computed from were produced by branches trained under Eq. (20).

  2. other [Sections 3.3-3.4, Eqs. (7)-(9) and (18)-(20)]
    "The empirical advantage of direct prediction is the loss gap A_i = sg(L_r,i - L_d,i) ... L_pred = 1/N sum_i [ g_i loss(y_hat^d_i, y_i) + (1 - g_i) loss(y_hat^r_i, y_i) ]."

    The soft routing weight g_i directly weights the two branch losses in Eq. (20), so updates to both branches depend on the valve's routing decision. But the valve's own target s*_i = sigma(A_i/eta) is recomputed from those same branch losses after every update (Eqs. 7-9). The gate is therefore chasing a target that its own routing decisions move: if the gate routes a sample to the direct branch, the direct branch is trained more and the gap widens, reinforcing the gate's choice. The warmup stage (Eq. 19) stabilizes the initial gap but does not remove the coupling during joint training. The paper provides no frozen-branch, delayed-target, or convergence analysis for this feedback loop, so the reported agreement between the valve and the loss-gap signal can partly reflect a self-fulfilling dy

full rationale

Most of SIEVE is a normal weakly-supervised routing method: sufficiency is explicitly defined operationally through the dual-branch loss gap, and the headline benchmark gains over GCNet, CorrKD, and MoMKE are measured against external labels on held-out test sets, so the main empirical claim is not forced by construction. No load-bearing self-citation chain is present. However, two parts of the framework and evaluation are partially self-referential. First, the 'Achievable Optimum' used as the upper reference is exactly the per-sample argmin of the two branch losses, which is the same sign information encoded in the valve's training target s*_i = sigma(sg(L_r,i - L_d,i)/eta); closeness to this optimum partly measures how well the gate replicates its own supervisory signal, not an externally fixed ceiling. Second, during joint training the routing weight g_i modulates both branch losses, which are then used to recompute the target, so the sufficiency signal is non-stationary and endogenous to the gate; this is a feedback loop rather than an independent label. These issues are partial: SIEVE could still be a useful and empirically effective router, and the headline accuracy improvements are externally measured. Score 4 reflects one evaluation bound constructed from the training target plus an endogenous supervision loop, while the central benchmark claim retains independent content.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced; the evidential valve is a learned module. The main free choices are gate hyperparameters tuned per dataset/backbone, the EMA base rate, and the unspecified warmup schedule. The threshold derivation is internally consistent given the chosen Subjective Logic axioms.

free parameters (6)
  • Loss-gap temperature η/κ = 0.08–0.50 (Table 4, labeled κ)
    Controls sharpness of the soft sufficiency target s* in Eq. (9); tuned per backbone and dataset, and symbol-confused with the EMA rate κ in Eq. (15).
  • EMA update rate κ in Eq. (15) = not reported separately
    Updates the missing-pattern base rate π_m; because κ is also used for the temperature, the actual value is ambiguous.
  • Valve calibration weight λ_v = 0.05–0.30 (Table 4)
    Weights the Beta-likelihood valve loss in Eq. (21); tuned per method and dataset.
  • Evidence regularizer weight λ_e = 0.01–0.12 (Table 4)
    Controls the KL divergence of wrong-direction evidence in Eq. (23); tuned per method and dataset.
  • Final Gumbel-Sigmoid temperature τ_min = 0.10–0.45 (Table 4)
    Annealed from 1.0; controls how close routing is to a hard binary decision; tuned per method and dataset.
  • Valve MLP hidden dimension = 96 or 128 (Table 4)
    Architecture hyperparameter for the valve network; chosen per method and dataset.
axioms (4)
  • domain assumption The per-sample loss gap A_i = L_r − L_d is a reliable and stable training signal for sufficiency.
    Core to the method (Section 3.3, Eqs. 7–9): if the co-evolving branches make the gap non-stationary or noisy, the valve learns a moving target; the paper only offers warmup without analysis.
  • domain assumption The oracle per-sample best subset in Figure 1 reflects sample-dependent modality utility rather than model noise or overfitting.
    Load-bearing for the 'repair is not always needed' motivation; computed from two fixed models and test labels, with no control for label noise or model capacity.
  • standard math The Beta/Subjective Logic parameterization with α,β > 1 and the decision threshold θ(u) in Eq. (14) is the appropriate uncertainty-aware decision rule.
    Assumes the adopted evidential framework; the derivation in Appendix A is internally consistent within that framework.
  • domain assumption The random missing protocol with identical masks across train/validation/test is representative of deployment-time missingness.
    Evaluation depends on this protocol; it may not transfer to structured missingness such as privacy-based or sensor-failure patterns.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Should Missing Modalities Always Be Necessary to Repair for Multi-modal Sentiment Analysis?." pith.science (2026). https://pith.science/paper/2RDWXVYO

@misc{pith2026260717262,
  author       = {Pith},
  title        = {Pith review of: Should Missing Modalities Always Be Necessary to Repair for Multi-modal Sentiment Analysis?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2RDWXVYO}},
  note         = {Machine review of arXiv:2607.17262}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing methods for multimodal sentiment analysis (MSA) under missing modalities usually follow a repair-first paradigm. We revisit this assumption and ask: \emph{should every missing modality be repaired?} A per-sample oracle analysis shows the answer is not always: full-modality input is optimal for only a small fraction of samples, and every modality subset is preferred by some samples. These results suggest that adding or repairing modalities may not always improve prediction, and that the utility of each modality is sample-dependent. Building on this finding, we propose \textbf{S}ufficiency-\textbf{I}nformed \textbf{E}vidential \textbf{V}al\textbf{vE} (\textbf{SIEVE}) that turns ``whether to repair'' into an explicit, learnable decision at the sample level. SIEVE compares a direct prediction branch with a repair branch, derives an empirical sufficiency signal from their per-sample loss gap, and routes each input through an evidential gate that jointly models sufficiency and its epistemic uncertainty. SIEVE is repair-agnostic: it operates as a plug-and-play decision on top of any explicit or implicit repair module, without modifying its internal design. Experiments on CMU-MOSI and IEMOCAP show that SIEVE consistently improves representative repair backbones across evaluated missing rates, and approaches the per-sample dual-branch achievable optimum.

Figures

Figures reproduced from arXiv: 2607.17262 by Fei Pan, Haotian Wu, Hong Chen, Jie Zhang, Puay Siew Tan, Ruoshui Peng, Xiaoyu Xu, Xuming Hu, Yibo Yan, Yonghua Hei, Yubo Gao, Zhuoran Gao.

Figure 1
Figure 1. Figure 1: Best-performing modality combination per [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SIEVE. Direct and repair branch yield two predictions, whose per-sample loss gap supervises [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE projection of per-sample routed-branch [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 5 linked inside Pith

  1. [1]

    Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Multimodal prompt learning with missing modalities for sentiment analysis and emotion recognition , author=. Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  2. [2]

    arXiv preprint arXiv:2603.09874 , year=

    MissBench: Benchmarking Multimodal Affective Analysis under Imbalanced Missing Modalities , author=. arXiv preprint arXiv:2603.09874 , year=

    arXiv

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Incomplete multimodality-diffused emotion recognition , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    A unified self-distillation framework for multimodal sentiment analysis with uncertain missing modalities , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  5. [5]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Mmanet: Margin-aware distillation and modality-aware regularization for incomplete multimodal learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  6. [6]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Distribution-consistent modal recovering for incomplete multimodal learning , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  7. [7]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Leveraging knowledge of modality experts for incomplete multimodal learning , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  8. [8]

    Missing modality imagination network for emotion recognition with uncertain missing modalities , author=. 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=

  9. [9]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    CMAD: Correlation-Aware and Modalities-Aware Distillation for Multimodal Sentiment Analysis with Missing Modalities , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  10. [10]

    Advances in Neural Information Processing Systems , volume =

    Toward Robust Incomplete Multimodal Sentiment Analysis via Hierarchical Representation Learning , author =. Advances in Neural Information Processing Systems , volume =

  11. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Correlation-decoupled knowledge distillation for multimodal sentiment analysis with incomplete modalities , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Hyper-Modality Enhancement for Multimodal Sentiment Analysis with Missing Modalities , author=. Advances in Neural Information Processing Systems , volume=

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Towards robust multimodal sentiment analysis with incomplete data , author=. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Tmdc: A two-stage modality denoising and complementation framework for multimodal sentiment analysis with missing and noisy modalities , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  15. [15]

    IEEE Transactions on pattern analysis and machine intelligence , volume=

    Gcnet: Graph completion network for incomplete multimodal learning in conversation , author=. IEEE Transactions on pattern analysis and machine intelligence , volume=. 2023 , publisher=

    2023

  16. [16]

    arXiv preprint arXiv:1606.06259 , year=

    Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos , author=. arXiv preprint arXiv:1606.06259 , year=

    Pith/arXiv arXiv

  17. [17]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  18. [18]

    Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

    Multimodal transformer for unaligned multimodal language sequences , author=. Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

  19. [19]

    Proceedings of the 28th ACM international conference on multimedia , pages=

    Misa: Modality-invariant and-specific representations for multimodal sentiment analysis , author=. Proceedings of the 28th ACM international conference on multimedia , pages=

  20. [20]

    arXiv preprint arXiv:2602.00811 , year=

    MissMAC-Bench: Building Solid Benchmark for Missing Modality Issue in Robust Multimodal Affective Computing , author=. arXiv preprint arXiv:2602.00811 , year=

    arXiv

  21. [21]

    Proceedings of the 2017 conference on empirical methods in natural language processing , pages=

    Tensor fusion network for multimodal sentiment analysis , author=. Proceedings of the 2017 conference on empirical methods in natural language processing , pages=

    2017

  22. [22]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Learning modality-specific representations with self-supervised multi-task learning for multimodal sentiment analysis , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  23. [23]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Found in translation: Learning robust joint representations by cyclic translations between modalities , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  24. [24]

    International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems , volume=

    A logic for uncertain probabilities , author=. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems , volume=. 2001 , publisher=

    2001

  25. [25]

    2016 , publisher=

    Subjective logic , author=. 2016 , publisher=

    2016

  26. [26]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Trusted multi-view classification with dynamic evidential fusion , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2022 , publisher=

    2022

  27. [27]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Asymmetric mutual alignment for unsupervised zero-shot sketch-based image retrieval , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  28. [28]

    Advances in neural information processing systems , volume=

    Evidential deep learning to quantify classification uncertainty , author=. Advances in neural information processing systems , volume=

  29. [29]

    Advances in neural information processing systems , volume=

    Deep evidential regression , author=. Advances in neural information processing systems , volume=

  30. [30]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Reliable conflictive multi-view learning , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  31. [31]

    Language resources and evaluation , volume=

    IEMOCAP: Interactive emotional dyadic motion capture database , author=. Language resources and evaluation , volume=. 2008 , publisher=

    2008

  32. [32]

    International joint conference on artificial intelligence , volume=

    The foundations of cost-sensitive learning , author=. International joint conference on artificial intelligence , volume=. 2001 , organization=

    2001

  33. [33]

    arXiv preprint arXiv:1904.05862 , year=

    wav2vec: Unsupervised pre-training for speech recognition , author=. arXiv preprint arXiv:1904.05862 , year=

    Pith/arXiv arXiv 1904

  34. [34]

    arXiv preprint arXiv:2006.03654 , year=

    Deberta: Decoding-enhanced bert with disentangled attention , author=. arXiv preprint arXiv:2006.03654 , year=

    Pith/arXiv arXiv 2006

  35. [35]

    IEEE Transactions on Image Processing , volume=

    Learning deep global multi-scale and local attention features for facial expression recognition in the wild , author=. IEEE Transactions on Image Processing , volume=. 2021 , publisher=

    2021

  36. [36]

    IEEE signal processing letters , volume=

    Joint face detection and alignment using multitask cascaded convolutional networks , author=. IEEE signal processing letters , volume=. 2016 , publisher=

    2016

  37. [37]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Merbench: A unified evaluation benchmark for multimodal emotion recognition , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  38. [38]

    9th International Conference on Learning Representations, ICLR 2021 , year=

    TRUSTED MULTI-VIEW CLASSIFICATION , author=. 9th International Conference on Learning Representations, ICLR 2021 , year=

    2021

  39. [39]

    arXiv preprint arXiv:1611.01144 , year=

    Categorical reparameterization with gumbel-softmax , author=. arXiv preprint arXiv:1611.01144 , year=

    Pith/arXiv arXiv

  40. [40]

    EMNLP 2025-2025 Conference on Empirical Methods in Natural Language Processing, Findings of EMNLP 2025 , pages=

    Two Challenges, One Solution: Robust Multimodal Learning through Dynamic Modality Recognition and Enhancement , author=. EMNLP 2025-2025 Conference on Empirical Methods in Natural Language Processing, Findings of EMNLP 2025 , pages=. 2025 , organization=

    2025

  41. [41]

    arXiv preprint arXiv:2604.05558 , year=

    Evaluation Before Generation: A Paradigm for Robust Multimodal Sentiment Analysis with Missing Modalities , author=. arXiv preprint arXiv:2604.05558 , year=

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.