Pith. sign in

REVIEW 3 major objections 6 minor 33 references

Multimodal Sentiment Analysis Based on Causal Reasoning

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read A counterfactual score that subtracts text-only and image-only effects from the fused prediction removes modality bias and yields new best results on MVSA-Single and MVSA-Multiple.

desk verdict A sensible adaptation of Counterfactual VQA to multimodal sentiment, but the central debiasing formula is arithmetically wrong and the empirical gains are too thin to rescue it. read the letter →

arxiv 2412.07292 v1 pith:UAECGA2E submitted 2024-12-10 cs.MM cs.CL

classification cs.MMcs.CL
keywords multimodalsentimentanalysiscounterfactualinferencecausaldebiasingmodalitybiastotalindirecteffecttext-imagefusionMVSA-SingleMVSA-Multiple
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 tries to establish that modality bias in multimodal sentiment analysis can be modeled as a direct causal effect and then removed by counterfactual reasoning. The proposed CF-MSA framework trains three branches--text-only, image-only, and text-image fusion--and at inference scores each label by subtracting the estimated text-only and image-only contributions from the fused score. On the MVSA-Single and MVSA-Multiple datasets the authors report that this subtraction improves accuracy and F1 over biased baselines and over earlier fusion methods, with the largest gains when text bias is removed. If the causal interpretation holds, the same recipe provides a general debiasing layer that can be attached to existing multimodal classifiers.

What carries the argument

The load-bearing object is the counterfactual decomposition of a model score into total effect, natural direct effect, and total indirect effect on the graph where text $T$ and image $I$ influence the label $Y$ both through direct paths and through the joint multimodal knowledge $K$. The implementation uses three branches $Z_t$, $Z_i$, $Z_k$, a SUM fusion $h(Z_t,Z_i,Z_k)=\log\sigma(Z_t+Z_i+Z_k)$, and two auxiliary losses: $L_{ti}$ symmetrizes the text-only and image-only distributions, while $L_{kl}$ fits the missing-input constants so the estimated natural direct effect has the same sharpness as the total effect. This machinery converts the debiasing intuition into a concrete label-scoring rule.

What would settle it

Construct a held-out set of image-text pairs in which the text contains explicit sentiment words opposite to the sentiment of the image, and compare CF-MSA against the biased baseline on those pairs. If the method truly removes text bias, its predictions should track the image content substantially more often than the baseline does; a failure to flip those predictions would show that the subtracted term is not capturing the text direct effect.

Watch

Extended reading notes

Core claim

The central claim is that the score $TIE_{joint}=2Z_{t,i,k}-Z_{t,i^*,k^*}-Z_{t^*,i,k^*}$ is a causal total indirect effect: it keeps the part of the fused multimodal prediction that flows through the joint multimodal mediator and subtracts the parts contributed directly by text alone and by image alone. The authors assert that choosing the sentiment label that maximizes this counterfactual score mitigates the direct effect of unimodal bias and, on MVSA-Single and MVSA-Multiple, gives the best accuracies and F1 scores among the methods compared. The argument rests on estimating the natural direct effects of the missing modalities with learnable constants $c_1,\dots,c_4$, fitted by KL divergence so that the sharpness of the direct-effect distributions matches the sharpness of the total-effect distribution.

Load-bearing premise

The load-bearing premise is that the learnable constants $c_1,\dots,c_4$, fitted by matching the sharpness of the direct-effect distributions to the total-effect distribution, correctly estimate the real direct effects of the missing modalities; if that sharpness match fails, the subtraction removes genuine signal rather than bias.

Editorial extensions

If this is right

  • If the total indirect effect score is a faithful causal quantity, replacing softmax inference with the argued TIE scoring should reduce misclassifications on samples where explicit sentiment words in text contradict the image.
  • The counterfactual layer should transfer across fusion architectures, since the paper reports gains when the same inference strategy is applied to OTE, CMAC, HSTEC, and NaiveCat.
  • The intermodal bias mutual elimination loss $L_{ti}$ should make text-only and image-only prediction distributions more alike, reducing the fused model's sensitivity to whichever modality dominates.
  • Non-uniform learnable constants should beat uniform, random, or prior-statistic settings because they allow per-modality calibration of the direct-effect estimate.

Reading between the lines

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

  • Beyond the paper, the identity $TIE_{joint}=Z_{t,i,k}+(Z_{t,i,k}-Z_{t,i^*,k^*})+(Z_{t,i,k}-Z_{t^*,i,k^*})$ suggests a general ensemble-debiasing rule for any multimodal classifier with unimodal branches, not only sentiment models.
  • A natural testable extension is to apply the same three-branch counterfactual subtraction to other tasks with known modality bias, such as visual question answering or hate-speech detection, where a dominant modality is known to drive spurious predictions.
  • Because the constants $c_1,\dots,c_4$ are fitted on the training distribution, the debiasing may be effective in-distribution but could weaken under distribution shift; evaluating on deliberately perturbed modality distributions would reveal how much of the gain is causal rather than distributional.
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

3 major / 6 minor

Summary. The paper proposes CF-MSA, a counterfactual causal inference framework for multimodal sentiment analysis. It models text and image as treatments, the multimodal fusion representation K as a mediator, and the sentiment label Y as the outcome. The method defines total effect (TE), natural direct effect (NDE), and total indirect effect (TIE), and performs debiased inference by selecting the label that maximizes a joint indirect effect TIE_joint, which is computed by subtracting learned counterfactual baselines from the fused prediction. The framework is instantiated with three branches (text, image, text-image fusion) and a fusion function, plus a KL-divergence-based loss L_ti that aligns the text and image counterfactual distributions. Experiments on MVSA-Single and MVSA-Multiple report accuracy and F1 improvements over several baselines, with ablations on the loss, the counterfactual conditions, and the learnable constants c.

Significance. If the causal derivation were correct, the paper would offer a principled way to remove unimodal bias in multimodal sentiment analysis by leveraging mediation analysis, and the proposed framework could potentially transfer to other multimodal tasks. The paper contains extensive experiments, ablations, and a promise of code release, which are strengths. However, the central theoretical claim is internally inconsistent: Eq. (24) does not follow from the TE/NDE decomposition defined in Eqs. (5)-(10). Moreover, the learnable constants c that define the counterfactual baselines are fit to match the observed joint distribution, so the 'direct effects' subtracted are not identified as natural direct effects from a causal model. These issues undermine the claim that the experimental gains are due to the proposed causal debiasing mechanism. The paper may still contain a useful heuristic debiasing method, but the causal interpretation as presented is not established.

major comments (3)
  1. [Section 3.1, Eqs. (8)-(10) and Eq. (24)] The joint indirect effect is defined as TIE_joint = TIE_text + TIE_image = 2Z_{t,i,k} - Z_{t,i*,k*} - Z_{t*,i,k*}. However, subtracting the two NDEs from the TE defined in Eq. (5) yields TE - NDE_text - NDE_image = Z_{t,i,k} - Z_{t,i*,k*} - Z_{t*,i,k*} + Z_{t*,i*,k*}. The reference term Z_{t*,i*,k*} is nonzero in general and is never set to zero or otherwise constrained in the paper, so Eq. (24) is not the TE - NDE decomposition claimed in the text. Consequently, the inference rule 'select the label maximizing TIE_joint' has no demonstrated connection to removing direct modality effects; the causal claim that motivates the method is formally unsupported.
  2. [Section 3.2, Eqs. (16)-(19)] The learnable parameters c1..c4 define the counterfactual baselines Z_{t,i*,k*} and Z_{t*,i,k*}. These parameters are fit by minimizing the KL divergence between the counterfactual distributions and the total-effect distribution (Lkl) and between the text and image counterfactual distributions (Lti). This means the 'direct effects' subtracted in Eq. (24) are not estimated as natural direct effects from a structural causal model; they are calibrated to match the observed joint distribution. The assumption that the sharpness of the NDE should be similar to that of the TE, stated before Eq. (17), is asserted without any test or justification. As a result, the ablation in Table 5 cannot distinguish a genuine causal debiasing effect from a learned reweighting that happens to improve accuracy on these particular datasets.
  3. [Section 4.3, Tables 2 and 3] The abstract and Section 4.3 claim new state-of-the-art performance, but the only condition that clearly exceeds all baselines on MVSA-Single is 'Removing Text Bias' (76.17 ACC); the full 'Removing Text-Image Bias' condition achieves 74.80, which is lower, and 'Removing Image Bias' achieves 73.24. In Table 3, several counterfactual variants degrade performance relative to the biased baselines (e.g., CF-CMAC removing image bias drops by 2.39 ACC on MVSA-Multiple; CF-CMAC removing text-image bias drops by 0.36 ACC on MVSA-Multiple). The paper does not specify which condition corresponds to the model referred to as 'CF-MSA' in the headline claim, and the inconsistent improvements make it difficult to attribute the observed gains to the proposed debiasing mechanism.
minor comments (6)
  1. [Section 1] There are several typos and grammatical errors, e.g., 'Multimoal data' and 'attracts increasing research and industrial attention, as Therefore' in the first two paragraphs; these should be corrected.
  2. [Eq. (9)] The subscript in TIE_image is written as Z_i,t*,k*, which is inconsistent with the notation used elsewhere; it should be Z_{t*,i,k*} to match the counterfactual where text is masked and image is present.
  3. [Section 4.1 and Table 1] The dataset statistics in Table 1 (4,511 and 16,779) do not match the original dataset sizes mentioned in the text (5,129 and 19,600); the preprocessing steps that lead to the reduced counts should be described in more detail.
  4. [Section 3.1] The mediator K is described as 'multimodal fusion knowledge', but its causal role is not formalized with a structural equation or explicit graph; the current presentation leaves the relation between K and the branching models ambiguous.
  5. [Section 1] The claim of being 'the first to introduce the causal effects into multimodal sentiment analysis' is too strong given the existing causal and counterfactual methods cited later (e.g., Niu et al. 2021); the novelty should be framed more precisely.
  6. [Table 5] The four c-parameter hypotheses (Random, CF-MSA Prior, Uniform, Non-uniform) are not precisely defined; in particular, 'prior knowledge from dataset statistic' and 'non-uniformly distributed Niu et al. (2021)' need explicit mathematical descriptions for the experiments to be reproducible.

Circularity Check

2 steps flagged · score 7.0 of 10

CF-MSA's causal debiasing reduces to a fit: the subtracted NDEs are learned constants optimized to match the TE distribution, and Eq. 24's 'TE − NDE' label is definitional, not a derivation.

  1. fitted input called prediction [Section 3.2, 'The learnable parameter c' (Eqs. 12-14, 17-19) and Inference (Eqs. 22-24)]
    "The learnable parameter c, which controls the sharpness of theZt,i∗,k∗ and Zt∗,i,k∗ distributions, assumes that the sharpness of the NDE should be similar to that of the TE. Otherwise, the study concluded that inappropriate c could lead to TIE being dominated by either TE or NDE. Therefore, the experiment uses the Kullback-Leibler divergence to estimate c, which is expressed by the simplified optimization objective function as follows::"

    The NDE terms subtracted in Eqs. 22-24 are Zt,i*,k* = h(zt, c2, c3) and Zt*,i,k* = h(c1, zi, c4), with c1..c4 learned. The KL losses (17)-(19) minimize the divergence between p(y|t,i,k) (the TE distribution) and p(y|t,i*,k*), p(y|t*,i,k*), so the 'direct effects' are optimized to reproduce the very total effect from which they are later subtracted. The resulting TIE is a residual of a distribution-matching fit, not an independently identified causal quantity; the sharpness-similarity assumption is imposed to make this fit well-posed. Table 5 then selects the c-distribution by accuracy, so the causal mechanism is post-hoc tuned to the outcomes it is said to predict.

  2. self definitional [Section 3.1 Eq. 10; Section 3.2 Inference Eq. 24]
    "The joint total indirect effect (TIE) is a composite effect that adds the above two indirect effects so as to capture the combined effect of text and image jointly influencing sentiment polarity through mediating variables ... T IEjoint = T IEtext + T IEimage (10) ... Combining the two yields a text-image bias implementation T IE= T E− N DE= 2× Zt,i,k − Zt,i∗,k∗ − Zt∗,i,k∗ = 2× h (zt, zi, zk) − h (zt, z∗ i , z∗ k) − h (z∗ t , zi, z∗ k) (24)"

    Eq. 24's RHS is exactly Eq. 10's TIE_text + TIE_image, so calling it 'TE − NDE' is a definitional label, not a consequence of Eqs. 5-7. From the paper's own definitions, TE − NDE_text − NDE_image = Zt,i,k − Zt,i*,k* − Zt*,i,k* + Zt*,i*,k*, which differs from Eq. 24 by Zt*,i*,k*. Nothing in the paper sets this baseline to zero. The inference rule 'select the label maximizing TIE_joint' is therefore not derived from the stated TE/NDE decomposition; the causal-debiasing score is the sum of two indirect-effect differences by construction.

full rationale

The central claim—that selecting the label maximizing Eq. 24's TIE_joint removes unimodal direct effects—is not supported by an independent causal derivation. The counterfactual baselines Zt,i*,k* and Zt*,i,k* are constructed from learnable constants c1..c4, and those constants are fit by KL divergence to make the NDE distributions approximate the TE distribution (Eqs. 17-19). The 'direct effects' are therefore fitted to the total effect they are meant to be subtracted from, so the residual TIE is a fitted quantity rather than a first-principles causal effect. Separately, Eq. 24 labels as 'TE − NDE' an expression that is just the definitional sum of Eqs. 8 and 9; under the paper's own TE/NDE definitions the stated identity omits the Zt*,i*,k* baseline term and is false unless an unstated assumption is added. No load-bearing self-citation chain is present (the closest external source, Niu et al. 2021, is not by the present authors), so the circularity is internal: the causal prediction reduces, by the paper's own equations and fitting procedure, to its own inputs.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method rests on a small set of learned constants and several modeling assumptions carried over from counterfactual VQA. No new physical or mathematical entities are introduced.

free parameters (1)
  • c1, c2, c3, c4 (learnable constants for missing modality outputs) = not reported (learned via KL loss)
    These constants replace missing branch outputs in counterfactual inference and are fit on training data to align direct-effect distributions with the total-effect distribution. The debiasing score depends directly on their values.
assumptions (4)
  • domain assumption The causal graph Text, Image -> K -> Y with direct paths T->Y and I->Y is a valid representation of multimodal sentiment prediction.
    The method decomposes effects using this graph; if the branches do not isolate direct and indirect paths, the subtraction has no causal meaning. Invoked in Section 3.1 and Figure 2.
  • domain assumption The natural direct effect can be estimated by feeding a single modality and replacing the other branch with a constant.
    The counterfactual quantities Z_{t,i*,k*} and Z_{t*,i,k*} are computed with constants c, which assumes the missing modality's contribution is a constant independent of the input. This is the core approximation borrowed from Niu et al. (2021).
  • domain assumption The sharpness (entropy) of the direct-effect distribution should match the total-effect distribution, enabling KL-based estimation of c.
    Stated in Section 3.2: 'the sharpness of the NDE should be similar to that of the TE'. If false, the KL loss does not identify c.
  • ad hoc to paper The sum of the two indirect effects (TIE_text + TIE_image) is a meaningful quantity for selecting the label.
    Eq. 24 defines the final score as 2*Z - Z_{t,i*,k*} - Z_{t*,i,k*}; this sum double-counts the total effect and is not derived from the TE/NDE decomposition. It is adopted without a causal justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Sentiment Analysis Based on Causal Reasoning." pith.science (2026). https://pith.science/paper/UAECGA2E

@misc{pith2026241207292,
  author       = {Pith},
  title        = {Pith review of: Multimodal Sentiment Analysis Based on Causal Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UAECGA2E}},
  note         = {Machine review of arXiv:2412.07292}
}
read the original abstract

With the rapid development of multimedia, the shift from unimodal textual sentiment analysis to multimodal image-text sentiment analysis has obtained academic and industrial attention in recent years. However, multimodal sentiment analysis is affected by unimodal data bias, e.g., text sentiment is misleading due to explicit sentiment semantic, leading to low accuracy in the final sentiment classification. In this paper, we propose a novel CounterFactual Multimodal Sentiment Analysis framework (CF-MSA) using causal counterfactual inference to construct multimodal sentiment causal inference. CF-MSA mitigates the direct effect from unimodal bias and ensures heterogeneity across modalities by differentiating the treatment variables between modalities. In addition, considering the information complementarity and bias differences between modalities, we propose a new optimisation objective to effectively integrate different modalities and reduce the inherent bias from each modality. Experimental results on two public datasets, MVSA-Single and MVSA-Multiple, demonstrate that the proposed CF-MSA has superior debiasing capability and achieves new state-of-the-art performances. We will release the code and datasets to facilitate future research.

Figures

Figures reproduced from arXiv: 2412.07292 by the authors.

Figure 1
Figure 1. The upper part illustrates traditional likelihood-based multimodal sentiment prediction, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The analysis of the unimodal T IE effect for image and text is shown on the left, where T ∗ and I ∗ indicate that text and image information is masked respectively, and the red crosses indicate that the path is blocked. In this case, the model relies only on unimodal information for sentiment prediction. The joint text-image effect (T IEjoint) is shown on the right. In counterfactual notation, in the original tradit… view at source ↗
Figure 3
Figure 3. CF-MSA model training consists of three main branches: the text branch ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative analysis of test set examples. The first probability distribution chart is the [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Cause-and-effect diagram examples and counterfactual situations. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages

  1. [1]

    Gpt-4 technical report

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

  2. [2]

    HinglishNLP: Fine-tuned Language Models for Hinglish Sentiment Detection

    M. Bhange and N. Kasliwal. Hinglishnlp: Fine-tuned language models for hinglish sentiment detection. arXiv:2008.09820. Available: http://arxiv.org/abs/2008.09820, August 2020. Accessed: Aug. 09, 2024

  3. [3]

    G. L. de Toledo and R. M. Marcacini. Transfer learning with joint fine-tuning for multimodal sentiment analysis. arXiv:2210.05790. Available: http://arxiv.org/abs/2210.05790, October 2022. Accessed: Aug. 09, 2024

  4. [4]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

  5. [5]

    Gandhi, K

    A. Gandhi, K. Adhvaryu, S. Poria, E. Cambria, and A. Hussain. Multimodal sentiment analysis: A systematic review of history, datasets, multimodal fusion methods, applications, challenges and future directions. Information Fusion, 91: 0 424--444, March 2023. doi:10.1016/j.inffus.2022.09.025

  6. [6]

    Filtering of brand-related microblogs using social-smooth multiview embedding

    Yue Gao, Yi Zhen, Haojie Li, and Tat-Seng Chua. Filtering of brand-related microblogs using social-smooth multiview embedding. IEEE Transactions on Multimedia, 18 0 (10): 0 2115--2126, October 2016. doi:10.1109/TMM.2016.2581483

  7. [7]

    Gencoglu and M

    O. Gencoglu and M. Gruber. Causal modeling of twitter activity during covid-19. Computation, 8 0 (4): 0 85, September 2020. doi:10.3390/computation8040085

  8. [8]

    Hu, T.-E

    G. Hu, T.-E. Lin, Y. Zhao, G. Lu, Y. Wu, and Y. Li. Unimse: Towards unified multimodal sentiment analysis and emotion recognition. arXiv:2211.11256. Available: http://arxiv.org/abs/2211.11256, November 2022. Accessed: Aug. 09, 2024

Show all 33 references
  1. [9]

    Jiang, X.-H

    W.-B. Jiang, X.-H. Liu, W.-L. Zheng, and B.-L. Lu. Multimodal adaptive emotion transformer with flexible modality inputs on a novel dataset with continuous labels. In Proceedings of the 31st ACM International Conference on Multimedia, pp.\ 5975--5984, Ottawa, ON, Canada, 2023....

  2. [10]

    Multimodal pear chain-of-thought reasoning for multimodal sentiment analysis

    Yan Li, Xiangyuan Lan, Haifeng Chen, Ke Lu, and Dongmei Jiang. Multimodal pear chain-of-thought reasoning for multimodal sentiment analysis. ACM Transactions on Multimedia Computing, Communications and Applications, 2024

  3. [11]

    Y. Liu, W. Chen, G. Li, and L. Lin. Causalvlr: A toolbox and benchmark for visual-linguistic causal reasoning. arXiv:2306.17462. Available: http://arxiv.org/abs/2306.17462, December 2023 a . Accessed: Aug. 09, 2024

  4. [12]

    Y. Liu, G. Li, and L. Lin. Cross-modal causal relational reasoning for event-level visual question answering. arXiv:2207.12647. Available: http://arxiv.org/abs/2207.12647, June 2023 b . Accessed: Aug. 09, 2024

  5. [13]

    Louizos, U

    C. Louizos, U. Shalit, J. Mooij, D. Sontag, R. Zemel, and M. Welling. Causal effect inference with deep latent-variable models. arXiv:1705.08821. Available: http://arxiv.org/abs/1705.08821, November 2017. Accessed: Aug. 09, 2024

  6. [14]

    Counterfactual vqa: A cause-effect look at language bias

    Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. arXiv preprint arXiv:2006.04315, 2021. URL https://arxiv.org/abs/2006.04315. Accessed: Oct. 01, 2024

  7. [15]

    Radford and et al

    A. Radford and et al. Learning transferable visual models from natural language supervision. arXiv:2103.00020. Available: http://arxiv.org/abs/2103.00020, February 2021. Accessed: Aug. 09, 2024

  8. [16]

    Rajan, A

    V. Rajan, A. Brutti, and A. Cavallaro. Is cross-attention preferable to self-attention for multi-modal emotion recognition? In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.\ 4693--4697, Singapore, Singapore, 2022. IEE...

  9. [18]

    Shi and S.-L

    T. Shi and S.-L. Huang. Multiemo: An attention-based correlation-aware multimodal fusion framework for emotion recognition in conversations, 2023

  10. [19]

    Cubemlp: An mlp-based model for multimodal sentiment analysis and depression estimation

    Hao Sun, Hongyi Wang, Jiaqing Liu, Yen-Wei Chen, and Lanfen Lin. Cubemlp: An mlp-based model for multimodal sentiment analysis and depression estimation. In Proceedings of the 30th ACM International Conference on Multimedia, pp.\ 1--8, Lisboa, Portugal, 2022. ACM. doi:10.1145/...

  11. [20]

    K. Tang, Y. Niu, J. Huang, J. Shi, and H. Zhang. Unbiased scene graph generation from biased training. arXiv:2002.11949. Available: http://arxiv.org/abs/2002.11949, March 2020. Accessed: Aug. 09, 2024

  12. [21]

    T. Wang, J. Huang, H. Zhang, and Q. Sun. Visual commonsense r-cnn. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 10757--10767, Seattle, WA, USA, 2020. IEEE. doi:10.1109/CVPR42600.2020.01077

  13. [22]

    T. Wang, C. Zhou, Q. Sun, and H. Zhang. Causal attention for unbiased visual recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 3071--3080, Montreal, QC, Canada, 2021. IEEE. doi:10.1109/ICCV48922.2021.00308

  14. [23]

    Y. Wang, M. Huang, X. Zhu, and L. Zhao. Attention-based lstm for aspect-level sentiment classification. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp.\ 606--615, Austin, Texas, 2016. Association for Computational Linguistics. doi...

  15. [24]

    Borrowing human senses: Comment-aware self-training for social media multimodal classification

    Chunpu Xu and Jing Li. Borrowing human senses: Comment-aware self-training for social media multimodal classification. arXiv, March 2023. Accessed: Aug. 09, 2024

  16. [25]

    Xu and W

    N. Xu and W. Mao. Multisentinet: A deep semantic network for multimodal sentiment analysis. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pp.\ 2399--2402, Singapore, Singapore, 2017. ACM. doi:10.1145/3132847.3133142

  17. [26]

    X. Yang, H. Zhang, G. Qi, and J. Cai. Causal attention for vision-language tasks. arXiv:2103.03493. Available: http://arxiv.org/abs/2103.03493, March 2021. Accessed: Aug. 09, 2024

  18. [27]

    C. Zang, H. Wang, M. Pei, and W. Liang. Discovering the real association: Multimodal causal reasoning in video question answering. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 19027--19036, Vancouver, BC, Canada, 2023. IEEE. doi:10.1109/C...

  19. [28]

    Multimodal sentiment analysis

    Yeexiao Zheng. Multimodal sentiment analysis. https://github.com/YeexiaoZheng/Multimodal-Sentiment-Analysis, 2022. Accessed: Sept. 01, 2024

  20. [29]

    T. Zhu, L. Li, J. Yang, S. Zhao, H. Liu, and J. Qian. Multimodal sentiment analysis with image-text interaction network. IEEE Trans. Multimedia, 25: 0 3375--3385, 2023 a . doi:10.1109/TMM.2022.3160060

  21. [30]

    T. Zhu, L. Li, J. Yang, S. Zhao, and X. Xiao. Multimodal emotion classification with multi-level semantic reasoning network. IEEE Trans. Multimedia, 25: 0 6868--6880, 2023 b . doi:10.1109/TMM.2022.3214989

  22. [31]

    write newline

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

  23. [32]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  24. [33]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  25. [34]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.