Pith. sign in

REVIEW 3 major objections 3 minor 37 references

EII-SCL: Harnessing Emotional Inertia for Multimodal Emotion Recognition in Conversation

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper aims to show that encoding 'emotional inertia'—the gradual persistence of a speaker's emotion over time—as a supervised contrastive loss improves multimodal emotion recognition in conversation, and that this works as a plug-in for

desk verdict The hard-negative idea is good and the loss is clean, but the adaptive inertia window is not trainable as written (the floor kills the gradient to Wq/Wk), so the central claim needs a fix before the empirical gains can be trusted. read the letter →

arxiv 2607.17366 v1 pith:3GEXJ43J submitted 2026-07-19 cs.MM cs.CLcs.HCeess.SP

Zilong Huang , Kong Aik Lee , Chong-Xin Gan , Zezhong Jin , Ruichen Zuo , Man-Wai Mak This is my paper · ORCID
classification cs.MMcs.CLcs.HCeess.SP
keywords emotionalinertiasupervisedcontrastivelearningmultimodalemotionrecognitionconversationalhardnegativesamplingtemporalwindowIEMOCAPMELD
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 proposes a plug-in training loss called EII-SCL that encodes the psychological prior of emotional inertia: a speaker's emotion changes gradually across utterances, so same-speaker utterances with different labels that are close in time are especially confusing and should be treated as hard negatives. The module computes an attention-weighted temporal window per utterance, labels samples inside that window as hard negatives, and adds a supervised contrastive loss to the usual cross-entropy loss. On IEMOCAP and MELD, adding the loss to two simple backbones (a transformer-based and a graph-based model) yields accuracy and weighted-F1 above previous state-of-the-art, e.g., 73.95% accuracy and 74.01 w-F1 on IEMOCAP. The central promise is that any MERC model can be improved by around 0.5–2 points without new data or architectural changes.

What carries the argument

The central mechanism is the attention-weighted inertia window ω_i = floor(Σ_j a_ij |i−j|), where attention weights are computed between the anchor utterance and all same-speaker utterances. This integer window defines which same-speaker, different-label utterances are treated as 'hard negatives' because their emotion is still transitioning under inertia. The loss then weights these hard negatives by (1−cos)/2 before feeding them into the contrastive denominator, making the repulsion force stronger for more similar (more confusing) pairs. This adaptivity expands the window in emotionally stable stretches and contracts it around abrupt shifts.

What would settle it

Train the same backbone with EII-SCL on a version of IEMOCAP where the utterance order of each speaker is randomly shuffled within each dialogue, destroying temporal inertia while keeping labels and speaker identities; if accuracy still improves by the same margin, the improvement is not due to emotional inertia. Alternatively, measure the hard-negative cosine-similarity gap on this shuffled version: if the gap disappears but the gain persists, the window is not the operative mechanism.

Watch

Extended reading notes

Core claim

The paper claims that emotional inertia can be turned into a contrastive objective. For each utterance, the positive set is same-speaker utterances with the same label; the easy-negative set includes different-speaker utterances and same-speaker different-label utterances outside the inertia window; the hard-negative set consists of same-speaker different-label utterances inside the window, where the window size is computed as the attention-weighted temporal distance to other utterances of the same speaker. Hard negatives are dynamically weighted by (1−cos)/2 to avoid over-penalizing similar states. Adding this loss to the cross-entropy loss improves both transformer-based and graph-based ba

Load-bearing premise

The load-bearing premise is that the learned attention-weighted window correctly identifies which same-speaker utterances with different emotion labels are 'hard negatives' because of emotional inertia; if this window is miscalibrated, the whole contrastive sampling collapses.

Editorial extensions

If this is right

  • Any existing MERC model can adopt EII-SCL by adding one auxiliary loss term; no new annotation or architecture change is required.
  • The same-speaker temporal locality of emotion transitions is a transferable prior: backbones that ignore speaker identity and time can gain discriminative power on confusing emotion pairs.
  • The reported gains on both IEMOCAP and MELD suggest the benefit is not dataset-specific, though the magnitude varies (about 0.5–2 points in accuracy/w-F1).
  • If the effect holds, the method offers a cheap way to build more emotion-discriminative representations in conversational AI systems such as dialogue agents and patient monitoring.

Reading between the lines

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

  • A natural extension is to apply the same inertia window to other sequential prediction tasks with a notion of agent identity, such as stance detection in debates or intent tracking in dialogue; the hard-negative idea only requires a consistent 'speaker' attribute.
  • The attention-based window is trained jointly with the classification loss, so early training iterations may mis-estimate the window and distort hard-negative sets; a warm-up phase or a fixed, validated window for the first epochs might stabilize the method.
  • The uncommented drop in sad-class F1 for the graph-based backbone (74.14→65.83) suggests the hard-negative construction may be biased against low-frequency emotions; a class-balanced or adaptive sampling variant could recover this while preserving overall gains.
  • A more direct test of the inertia hypothesis would be to measure whether the model's learned attention weights align with human-annotated emotion-shift boundaries; the paper currently validates only the cosine-similarity gap between hard and easy negatives.
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 / 3 minor

Summary. The paper proposes EII-SCL, a supervised contrastive learning module for multimodal emotion recognition in conversation (MERC). The module constructs positive, easy-negative, and hard-negative sample sets for each utterance using speaker identity, emotion labels, and a supposedly adaptive temporal 'inertia window' derived from attention weights (Eqs. 6–9). The contrastive loss (Eqs. 10–11) is added to the backbone cross-entropy loss (Eq. 12) and is evaluated on IEMOCAP and MELD using two backbone MERC models (MM-Transformer and MM-DialogueGCN). The paper reports consistent improvements over baseline methods and performs analyses of the inertia window and ambiguous emotion pairs.

Significance. If the proposed mechanism worked as described, EII-SCL would be a lightweight, annotation-free plug-in that improves MERC performance through a psychologically motivated inductive bias. The paper also provides an empirical analysis of hard-negative similarity that could be valuable for understanding emotional transitions. However, the central adaptive-window mechanism is undermined by a non-differentiability problem, and the experimental evidence lacks statistical grounding. The significance of the empirical gains is therefore questionable, and the main claimed contribution — the learned dynamic inertia window — is not supported.

major comments (3)
  1. [§2.4, Eq. (7)] The inertia window ω_i = floor(Σ_j a_ij |i−j|) is non-differentiable with respect to the attention weights a_ij. Since L_eii (Eq. 11) depends on Wq and Wk only through the discrete set memberships in Eqs. (8)–(9), the gradient through the floor is zero almost everywhere; under standard backpropagation, Wq and Wk receive no gradient from L_eii. The cross-entropy loss L_CE does not involve them either. Thus the 'learnable' query/key projections remain at their random initialization, and the dynamic window is not adapted to the data. The §3.5 claim that the dynamic window 'consistently provides more informative hard-negative samples' is therefore unsupported. The authors must specify a differentiable relaxation (e.g., soft window, Gumbel-softmax, or straight-through estimator) or otherwise demonstrate that Wq/Wk are actually trained.
  2. [Table 1, §3.3] Experimental results are reported for a single run with no seed variation or error bars. On MELD, the improvements from EII-SCL are small (e.g., MM-TransFormer accuracy 67.64→68.19, w-F1 66.58→67.33) and are within typical run-to-run variance. Additionally, class-level regressions are unacknowledged: Sad drops from 74.14 to 65.83 for MM-DialogueGCN+EII-SCL, and Angry drops from 71.18 to 69.10 for MM-Transformer+EII-SCL. These large per-class drops undermine the claim of consistent improvement and need to be discussed or addressed.
  3. [§3.4, Fig. 2] The validation of the emotional inertia hypothesis uses the same fused features that are produced by the model trained with EII-SCL. If the dynamic window is not actually learned (see preceding comment), the analysis merely shows that utterances inside a random window are more similar, which is a trivial property of temporally adjacent utterances. If the window is learned, the analysis is partially self-referential because the loss itself reshapes the features used to compute the similarity. A more direct validation would use fixed windows and features from a backbone without EII-SCL, or a held-out set of features.
minor comments (3)
  1. [§2.4, Eq. (7)] The floor notation is used without comment on its differentiability. If a straight-through estimator or other approximation is used, it should be stated explicitly; if not, the mechanism should be revised.
  2. [Table 1] The caption says best and second-best are bolded and underlined, but the table does not consistently show underlining for second-best entries. Please check formatting.
  3. [§3.2] The hyperparameters τ=0.07 and α=0.02 are given, but there is no sensitivity analysis or ablation for these values. Given the small performance gains, such an analysis would help establish robustness.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EII-SCL's headline result is held-out benchmark performance against external baselines; the auxiliary loss and window construction are not equivalent to the inputs by construction.

full rationale

The central claim is empirical: adding L_eii (Eq. 12) to two backbone MERC models improves Acc/w-F1 on IEMOCAP and MELD relative to published baselines (Table 1). That comparison is external and not derived from the method's own equations. The inertia window (Eqs. 6-8) and the hard/easy negative sets (Eq. 9) are operational definitions used to build the auxiliary loss; they are not fitted to the test labels or to the reported metrics. The Sec. 3.4 similarity analysis is a dataset-property sanity check using fixed windows, not a prediction made from the fitted model, so it does not make the argument circular. Self-citations [6-10] from the same author group appear only as related-work decoration for representation learning and are not load-bearing for the EII-SCL derivation. I note one correctness risk, not circularity: because Eq. 7 applies floor to the attention-weighted distance, gradients w.r.t. Wq/Wk vanish almost everywhere through omega_i, so the 'learnable' dynamic window is not actually trained as claimed unless a straight-through/soft mechanism is used; however, this affects whether the mechanism works as described, not whether the benchmark outcome is circular. The paper's headline claim remains self-contained against external baselines, so the circularity score is 0.

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

The central claim rests on the emotional-inertia prior borrowed from psychology (refs [15,16]), on the transfer from similarity-based hard-negative sampling to classification accuracy, and on a hand-tuned loss. Free parameters: temperature τ=0.07, loss weight α=0.02, learned attention projections Wq/Wk (Eq. 6), the hand-designed repulsion weight (Eq. 10), and the integer floor in Eq. (7). No new entities are invented: the inertia window is a sampling rule, not a postulated object.

free parameters (5)
  • temperature τ = 0.07
    Sharpness of the contrastive softmax in Eq. (10); fixed by hand, no sweep with error bars.
  • loss balance α = 0.02
    Weights EII loss vs cross-entropy in Eq. (12); chosen by hand, reported once.
  • query/key projections Wq, Wk = learned
    Eq. (6) attention weights defining the inertia window; learned jointly with the backbone, so the window is optimized rather than independently validated.
  • dynamic weight form (1−cos)/2 = fixed functional form
    Eq. (10) repulsion schedule on hard negatives; hand-designed, not derived or ablated against alternatives.
  • inertia window floor ⌊·⌋ = integer rounding
    Eq. (7) converts attention-weighted distance into a hard integer window; the choice of floor vs round vs threshold is unexamined.
assumptions (4)
  • domain assumption Emotional states of a speaker persist and transition gradually (emotional inertia)
    Borrowed from psychology (refs [15,16]) and treated as a prior in §1; the paper's own §3.4 analysis is offered as support but may use the same features the loss reshapes.
  • domain assumption Same-speaker utterances with different labels inside the inertia window are 'hard' negatives whose repulsion helps classification
    Core sample-construction premise of §2.4 (Eq. 9); the similarity gap is measured in §3.4, but transfer to classification accuracy is asserted, not derived.
  • domain assumption Fused multimodal embeddings are a valid space for similarity-based contrastive sampling
    The loss operates on fused features F from Eq. (4); assumes cosine similarity there reflects emotional affinity.
  • standard math Standard building blocks work as published (Bi-GRU, Transformer, DialogueGCN, RoBERTa/wav2vec2/CLIP)
    §3.2 uses public encoders and DialogueGCN [22] unchanged; correctness of these backbones is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EII-SCL: Harnessing Emotional Inertia for Multimodal Emotion Recognition in Conversation." pith.science (2026). https://pith.science/paper/3GEXJ43J

@misc{pith2026260717366,
  author       = {Pith},
  title        = {Pith review of: EII-SCL: Harnessing Emotional Inertia for Multimodal Emotion Recognition in Conversation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GEXJ43J}},
  note         = {Machine review of arXiv:2607.17366}
}
read the original abstract

Multimodal emotion recognition in conversation (MERC) achieves accurate predictions by integrating multimodal and contextual information in dialogues. While current MERC approaches focus on modeling complex contextual dependencies in conversation, they often overlook the impact of contextual emotional inertia in emotion shift, leading to sub-optimal performance. To address this issue, we propose a novel Emotional Inertia-Informed Supervised Contrastive Learning module (EII-SCL) that informs the contrastive objective by constructing inertia-affected samples within temporal windows, effectively leveraging emotional inertia as a prior while enabling seamless integration with existing MERC models without requiring additional data. Extensive experiments on IEMOCAP and MELD show that our approach consistently outperforms state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2607.17366 by the authors.

Figure 1
Figure 1. The overall architecture of the proposed EII-SCL approach. The backbone MERC model completes emotion prediction through contextual features extraction, multimodal feature fusion, and emotion classification. EII-SCL works with the backbone network to learn more subtle and discriminative feature representations through a carefully designed inertia-informed sampling strategy. The thickness of the lines indicates the in… view at source ↗
Figure 3
Figure 3. Impact of the dynamic and fixed inertia window size on ERC performance. The dynamic window adjusts to capture emotional persistence for each utterance. 3.6. Performance on Ambiguous Emotion Pairs The results in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 7 linked inside Pith

  1. [1]

    Different from traditional emotion recog- nition, ERC analyzes dialogue turns and assigns one emotion label per utterance in the dialogue [2, 3]

    Introduction Emotion recognition in conversation(ERC) is a key technique in applications such as human-computer interaction and intelli- gent medical care [1]. Different from traditional emotion recog- nition, ERC analyzes dialogue turns and assigns one emotion label per utterance in the dialogue [2, 3]. The core of ERC lies in modeling contextual depende...

  2. [2]

    Methodology 2.1. Task Definition Consider a dialogue consisting ofktemporally ordered utter- ances, denoted asU={u i}k i=1, where individual utterances have corresponding emotion labels{y i}k i=1 and speaker labels {si}k i=1. The emotion labely i belongs to a predefined set of emotionsY. Each utteranceu i is accompanied by multimodal data, including a vid...

    arXiv 2026

  3. [3]

    frustrated

    Experiments 3.1. Datasets and Evaluation Metric IEMOCAP[25] is a widely used dataset for emotion recog- nition in conversation. For partitioning the data, we utilized the commonly used LOSO (Leave-One-Session-Out) strategy. MELD[26] is a multi-modal, multi-speaker conversational dataset derived from the TV series Friends. To ensure a fair comparison, we f...

    arXiv 2019

  4. [4]

    Conclusions We have proposed the emotional inertia-informed supervised contrastive learning (EII-SCL), which leverages emotional iner- tia—the psychological phenomenon in which emotional states tend to resist change over time—to enhance multimodal emo- tion recognition in conversation. Specifically, we select hard- negative samples within speaker-specific...

  5. [5]

    Acknowledgment The work presented in this article is supported by the Research Platform for Advanced Audio and Speech Signal Processing (P0049192) funded by Innovation Technology Co. Ltd

  6. [6]

    All scientific content, experiments and analyses were produced and verified by the authors

    Use of Generative AI Disclosure Generative AI tools were used only for language polishing and formatting assistance. All scientific content, experiments and analyses were produced and verified by the authors

  7. [7]

    UNet-based fusion and exponential moving average adaptation for noise-robust speaker recognition,

    C.-X. Gan, P. Bell, M.-W. Mak, Z. Li, Z. Jin, Z. Huang, and K. A. Lee, “UNet-based fusion and exponential moving average adaptation for noise-robust speaker recognition,”arXiv preprint arXiv:2604.25624, 2026

    arXiv 2026

  8. [8]

    Current state of text sentiment analysis from opinion to emotion mining,

    A. Yadollahi, A. G. Shahraki, and O. R. Zaiane, “Current state of text sentiment analysis from opinion to emotion mining,”ACM Computing Surveys (CSUR), vol. 50, no. 2, pp. 1–33, 2017

    2017

Show all 37 references
  1. [9]

    Emotion recognition in conversation: Research challenges, datasets, and recent advances,

    S. Poria, N. Majumder, R. Mihalcea, and E. Hovy, “Emotion recognition in conversation: Research challenges, datasets, and recent advances,”IEEE Access, vol. 7, pp. 100 943–100 953, 2019

    2019

  2. [10]

    MM-NodeFormer: Node transformer multimodal fusion for emotion recognition in conver- sation,

    Z. Huang, M.-W. Mak, and K. A. Lee, “MM-NodeFormer: Node transformer multimodal fusion for emotion recognition in conver- sation,” inProc. Interspeech, 2024, pp. 4069–4073

    2024

  3. [11]

    MMGCN: Multimodal fusion via deep graph convolution network for emotion recognition in conversation,

    J. Hu, Y . Liu, J. Zhao, and Q. Jin, “MMGCN: Multimodal fusion via deep graph convolution network for emotion recognition in conversation,”arXiv preprint arXiv:2107.06779, 2021

    2021 arXiv

  4. [12]

    Adaptive graph learning for multimodal conversational emotion detection,

    G. Tu, T. Xie, B. Liang, H. Wang, and R. Xu, “Adaptive graph learning for multimodal conversational emotion detection,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 17, 2024. [Online]. Available: https://ojs.aaai.org/ index.php/AAAI/article/view/29876

    2024

  5. [13]

    Distilling attention knowledge for speaker verification,

    Z. Jin, S. Liu, Z. Li, C.-X. Gan, Z. Huang, M.-W. Mak, and K. A. Lee, “Distilling attention knowledge for speaker verification,” in ICASSP 2026 - 2026 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2026, pp. 16 447– 16 451

    2026

  6. [14]

    A transformer-based model with self-distillation for multimodal emotion recognition in conversations,

    H. Ma, J. Wang, H. Lin, B. Zhang, Y . Zhang, and B. Xu, “A transformer-based model with self-distillation for multimodal emotion recognition in conversations,”IEEE Transactions on Multimedia, vol. 26, pp. 776–788, 2023

    2023

  7. [15]

    Curriculum learning meets directed acyclic graph for multimodal emotion recogni- tion,

    C.-B. Nguyen, D.-T. Le, Q. T. Haet al., “Curriculum learning meets directed acyclic graph for multimodal emotion recogni- tion,” inProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evalua- tion (LREC-COLING 2024), 2024,...

    2024

  8. [16]

    De- noising student features with diffusion models for knowledge dis- tillation in speaker verification,

    Z. Jin, Y . Tu, Z. Li, Z. Huang, C.-X. Gan, and M.-W. Mak, “De- noising student features with diffusion models for knowledge dis- tillation in speaker verification,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2025, pp. 1–5

    2025

  9. [17]

    BA V-MossFormer2: En- hanced MossFormer2 for binaural audio-visual speech enhance- ment,

    W. Ren, K. Li, R. Chao, J. Li, Z. Huang, S. Ahmed, Y .-J. Li, K.-H. Hung, S.-S. Wang, H.-M. Wanget al., “BA V-MossFormer2: En- hanced MossFormer2 for binaural audio-visual speech enhance- ment,” inProc. AVSEC 2025, 2025, pp. 79–80

    2025

  10. [18]

    COLD Fusion: Calibrated and or- dinal latent distribution fusion for uncertainty-aware multimodal emotion recognition,

    M. K. Tellamekala, S. Amiriparian, B. W. Schuller, E. Andr ´e, T. Giesbrecht, and M. Valstar, “COLD Fusion: Calibrated and or- dinal latent distribution fusion for uncertainty-aware multimodal emotion recognition,”IEEE Transactions on Pattern Analysis and Machine Intelligence,...

    2024

  11. [19]

    Feature-enhanced multi- modal interaction model for emotion recognition in conversation,

    Y . Fu, X. Yan, W. Chen, and J. Zhang, “Feature-enhanced multi- modal interaction model for emotion recognition in conversation,” Knowledge-Based Systems, vol. 309, p. 112876, 2025

    2025

  12. [20]

    IDIR: Identifying and Distilling Informative Relations for Speaker Verification,

    C.-X. Gan, Z. Li, Z. Jin, Z. Huang, M.-W. Mak, and K. A. Lee, “IDIR: Identifying and Distilling Informative Relations for Speaker Verification,” inInterspeech 2025, 2025, pp. 5758–5762

    2025

  13. [21]

    ,fk] = FeatureFusion[R(t),R (a),R (v)],(4) wheref i ∈R Df denotes the fused multimodal embedding of utteranceu i

    or Graph-based [22]): F= [f 1, . . . ,fk] = FeatureFusion[R(t),R (a),R (v)],(4) wheref i ∈R Df denotes the fused multimodal embedding of utteranceu i. Each fused embeddingf i is fed into a classifier to obtain emotion probabilitiesp i, optimized using the standard cross-entrop...

  14. [22]

    Emotional inertia and psychological maladjustment,

    P. Kuppens, N. Allen, and L. Sheeber, “Emotional inertia and psychological maladjustment,”Psychological science, vol. 21, pp. 984–91, 07 2010

    2010

  15. [23]

    Emotional inertia and external events: The roles of exposure, reactivity, and recovery,

    P. Koval, A. Brose, M. Pe, M. Houben, Y . Erbas, D. Cham- pagne, and P. Kuppens, “Emotional inertia and external events: The roles of exposure, reactivity, and recovery,”Emotion (Wash- ington, D.C.), vol. 15, 04 2015

    2015

  16. [24]

    Emotion recognition in conversations with emotion shift detection based on multi-task learning,

    Q. Gao, B. Cao, X. Guan, T. Gu, X. Bao, J. Wu, B. Liu, and J. Cao, “Emotion recognition in conversations with emotion shift detection based on multi-task learning,”Knowledge-Based Sys- tems, vol. 248, p. 108861, 04 2022

    2022

  17. [25]

    Hybrid curriculum learn- ing for emotion recognition in conversation,

    L. Yang, Y . Shen, Y . Mao, and L. Cai, “Hybrid curriculum learn- ing for emotion recognition in conversation,” inProceedings of the AAAI conference on artificial intelligence, vol. 36, 2022, pp. 11 595–11 603

    2022

  18. [26]

    The Emotion is Not One-hot Encoding: Learning with Grayscale Label for Emotion Recognition in Conversation,

    Joosung Lee, “The Emotion is Not One-hot Encoding: Learning with Grayscale Label for Emotion Recognition in Conversation,” inProc. Interspeech, 2022, pp. 141–145

    2022

  19. [27]

    Beyond single emotion: Multi-label ap- proach to conversational emotion recognition,

    Y . Kang and Y .-S. Cho, “Beyond single emotion: Multi-label ap- proach to conversational emotion recognition,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025, pp. 24 321–24 329

    2025

  20. [28]

    M2FNET: Multi-modal fusion network for emotion recognition in conversation,

    V . Chudasama, P. Kar, A. Gudmalwar, N. Shah, P. Wasnik, and N. Onoe, “M2FNET: Multi-modal fusion network for emotion recognition in conversation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4652–4661

    2022

  21. [29]

    DialogueGCN: A graph convolutional neural net- work for emotion recognition in conversation,

    D. Ghosal, N. Majumder, S. Poria, N. Chhaya, and A. Gel- bukh, “DialogueGCN: A graph convolutional neural net- work for emotion recognition in conversation,”arXiv preprint arXiv:1908.11540, 2019

    1908 arXiv

  22. [30]

    CFN-ESA: A cross-modal fusion network with emotion-shift awareness for dialogue emo- tion recognition,

    J. Li, Y . Liu, X. Wang, and Z. Zeng, “CFN-ESA: A cross-modal fusion network with emotion-shift awareness for dialogue emo- tion recognition,”arXiv preprint arXiv:2307.15432, 2023

    2023 arXiv

  23. [31]

    DER-GCN: Dialog and event relation-aware graph convolutional neural network for mul- timodal dialog emotion recognition,

    W. Ai, Y . Shou, T. Meng, and K. Li, “DER-GCN: Dialog and event relation-aware graph convolutional neural network for mul- timodal dialog emotion recognition,”IEEE Transactions on Neu- ral Networks and Learning Systems, vol. 36, no. 3, pp. 4908– 4921, 2025

    2025

  24. [32]

    IEMOCAP: Interactive emotional dyadic motion capture database,

    C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan, “IEMOCAP: Interactive emotional dyadic motion capture database,”Language Resources and Evaluation, vol. 42, pp. 335–359, 2008

    2008

  25. [33]

    MELD: A multimodal multi-party dataset for emotion recognition in conversations,

    S. Poria, D. Hazarika, N. Majumder, G. Naik, E. Cam- bria, and R. Mihalcea, “MELD: A multimodal multi-party dataset for emotion recognition in conversations,”arXiv preprint arXiv:1810.02508, 2018

    2018 arXiv

  26. [34]

    MM-DFN: Mul- timodal dynamic fusion network for emotion recognition in con- versations,

    D. Hu, X. Hou, L. Wei, L. Jiang, and Y . Mo, “MM-DFN: Mul- timodal dynamic fusion network for emotion recognition in con- versations,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing, 2022, pp. 7037–7041

    2022

  27. [35]

    RoBERTa: A robustly optimized bert pretraining approach,

    Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov, “RoBERTa: A robustly optimized bert pretraining approach,”arXiv preprint arXiv:1907.11692, 2019

    1907 arXiv

  28. [36]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in Neural Information Processing Systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  29. [37]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” in Proc. International Conference on Machine Learning, 2021, pp. 8748–8763

    2021

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.