Pith. sign in

REVIEW 4 major objections 7 minor 46 references

Salience Adjustment for Context-Based Emotion Recognition

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a face's expressivity should set how much weight an emotion-recognition system gives to the face versus the situation, and that adding this salience adjustment improves human and automatic recognition alike.

desk verdict A clearly-written proof-of-concept for salience-weighted cue integration, but the performance gains rest on in-sample fitting and need independent validation. read the letter →

arxiv 2507.15878 v1 pith:6Z65WIEX submitted 2025-07-17 cs.CV cs.AI

Bin Han , Jonathan Gratch This is my paper
classification cs.CVcs.AI
keywords emotionrecognitionBayesiancueintegrationfacialexpressivitysalienceadjustmentcontext-basedperceptionvision-languagemodelsprisoner'sdilemmavalenceprediction
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 the visual salience of a facial expression — how much the face moves and shows emotion — controls how observers and machines should combine facial and situational evidence when judging emotion. On a corpus of prisoner's-dilemma reaction videos, the authors find that context-based human judgments track face-only judgments much more closely when faces are highly expressive, and track the situation about equally when faces are muted. They encode this as a salience adjustment: a single Expressivity Score is linearly mapped to a weight $w \in [0.5, 1.0]$ that becomes an exponent in a Bayesian fusion rule — the rule that combines face-only and situation-only probability estimates of emotion — $P(e|c,f) \propto P(e|f)^w P(e|c)^{1-w}/P(e)$. With that one change, the adjusted rule beats the unweighted rule in every evaluation the paper reports, whether the emotion distributions come from human annotations, from facial recognizers plus a language model, or from a vision-language model. The finding matters because the adjustment is inexpensive, rests on a psychological mechanism, and can be applied on top of any face-plus-context recognition pipeline.

What carries the argument

The load-bearing object is a single weighted fusion rule, Eq. 2: $P(e|c,f) \propto P(e|f)^w \cdot P(e|c)^{1-w} / P(e)$, where $w$ in $[0.5, 1.0]$ is the weight on the face channel. The weight comes from an Expressivity Score computed with OpenFace 2.0 action units (muscle-activation codes), dense optical-flow landmarks from ZFace, head-pose and gaze features — each standardized and equally weighted — which the authors validate against human expressivity ratings at $r = 0.61$. A linear mapping rescales the score into $w$, chosen to approximate the empirical relationship in Figure 1; the floor of $0.5$ keeps the situation channel always present, reproducing the equal-weight BCI baseline at minimum expressivity and approaching face-only integration at maximum expressivity. The same $w$ is then injected into a vision-language model prompt as an explicit instruction about whether to prioritize facial or contextual cues, which is how the mechanism carries over to a model that never computes the BCI product.

What would settle it

Hold out a random half of the 100 videos, fit the Expressivity-to-$w$ linear mapping using only the other half, and apply that frozen mapping to the held-out half: if salience-adjusted BCI no longer beats the unweighted rule on held-out MSE, RMSE, and KLD, the reported gains are an in-sample fit rather than a general mechanism. A complementary check is to run the same annotation protocol on a corpus whose expressions are not smile-dominated; if the 82% face-proximity effect for the high-expressivity tertile does not reappear, the hypothesis is corpus-specific.

Watch

Extended reading notes

Core claim

The paper's central claim is the Expression-Salience Hypothesis: when a face is visually expressive, observers weight facial evidence more heavily than situational context, and when it is muted, they weight the situation more heavily. The evidence for the hypothesis is Figure 1, built from 100 annotated videos in the USC Split-Steal corpus: for the most expressive tertile of faces, 82% of context-based human judgments lie closer to the face-only judgment than to the situation-only judgment, while for the least expressive tertile the two channels contribute nearly equally. The paper then claims that this perceptual mechanism transfers to automatic recognition. Replacing the BCI rule $P(e|c,f) \propto P(e|f)P(e|c)/P(e)$ with the salience-weighted form $P(e|c,f) \propto P(e|f)^w P(e|c)^{1-w}/P(e)$, with $w$ a linear rescaling of the Expressivity Score into $[0.5, 1.0]$, improves prediction of human context-based labels and improves fully automatic pipelines (EmoNet, Blueskeye, Facet, EAC, or a fine-tuned LSTM for the face channel, GPT-4 for the context channel) and a vision-language model prompted with the same salience weight. Across Tables I–IV, the adjusted version yields lower MSE, RMSE, and KL divergence and higher correlation than the unweighted baseline.

Load-bearing premise

The results rest on the assumption that the linear mapping from the Expressivity Score to the weight $w$ — chosen to approximate Figure 1, which was built from the same 100 videos used in every validation — describes how humans actually combine face and situation, and that the mapping works on new videos and new tasks.

Editorial extensions

If this is right

  • Salience adjustment improves BCI's fit to human context-based judgments for both valence and basic emotion: in the human-annotation validation, valence MSE drops from $0.199$ to $0.108$ and basic-emotion KLD drops from $0.308$ to $0.146$.
  • Fully automatic pipelines inherit the gain without retraining the underlying recognizers, since the adjustment is a post-processing step on their output distributions; e.g., LSTM+GPT basic-emotion KLD falls from $0.537$ to $0.347$ and Facet+GPT correlation rises from $0.238$ to $0.640$.
  • A vision-language model (gpt-4o-mini-2024-0718) prompted with the salience weight beats both its context-free and its context-based-without-salience versions on valence and basic emotion, so the mechanism is not tied to the Bayesian formulation.
  • The results support the Expression-Salience Hypothesis as a description of human perception: observers do not always weight face and situation equally, and the imbalance is predictable from facial movement.

Reading between the lines

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

  • The weight floor of $w = 0.5$ encodes an assumption the paper does not test: that the situation channel never outweighs the face channel. A natural extension is to allow $w < 0.5$ for suppressed or deliberately neutral expressions and test whether any corpus demands context-dominant fusion.
  • Because the Expressivity-to-$w$ mapping was fit on the same 100 videos used for all validation, the decisive check of generality is a corpus-level split that freezes $w$ on one half and evaluates on the other; the paper reports no such held-out fit and acknowledges that the 'proper' Bayesian treatment would require a larger corpus.
  • The same salience-weighting logic should transfer to other channels of multimodal emotion recognition — voice prosody or body movement whose intensity varies could carry the same kind of exponent — which the authors list as future work but do not demonstrate.
  • The corpus is smile-dominated, as the paper concedes; a corpus with diverse, rarer emotion categories would show whether the 82% face-proximity effect is a general expressivity effect or a smile-attention effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes a salience-adjusted extension of Bayesian Cue Integration (BCI) for context-based emotion recognition. Using 100 videos from the USC Split-Steal corpus, the authors first show that when facial expressions are more expressive, human observers' context-based judgments are closer to face-only annotations (Fig. 1). They then introduce Eq. 2, in which the face channel is weighted by w derived from an automatic Expressivity Score, rescaling it linearly to [0.5, 1.0]. Experiments compare Eq. 2 with unweighted BCI and with a GPT-4o-based VLM that receives a salience prompt, reporting improvements in MSE, RMSE, KLD, and correlation on valence and basic emotion tasks. The paper concludes that salience adjustment consistently improves human and automatic emotion recognition.

Significance. If the claimed effect is real, the paper contributes a simple, psychologically motivated mechanism for late fusion in emotion recognition and demonstrates it on a naturalistic social-decision task. The use of an established corpus, human annotations, multiple automatic face models, and a VLM is a strength, as is the explicit framing of the Expression-Salience Hypothesis. However, the evaluation is entirely in-sample: the salience mapping is fit to the same 100 videos and the same annotation set used as ground truth, and no held-out split, cross-validation, or uncertainty quantification is reported. The paper also contains point estimates that contradict the 'consistently improves' claim. The contribution is therefore plausible but not yet demonstrated as a generalizable method.

major comments (4)
  1. [Section III-B and III-C] The salience weight w is a linear rescaling of the Expressivity Score chosen to approximate Figure 1, which is computed from the same 100 videos and the same human context-based annotations that are used as ground truth in all evaluation tables. Consequently, the improvement of Eq. 2 over Eq. 1 in Table I is at least partially in-sample, and the fixed w used in Tables II-IV is tuned on the evaluation corpus. Please separate fitting and evaluation (e.g., cross-validation across videos or an independent corpus) and report the performance of a version of w learned only on the training portion.
  2. [Section III-B] The exact mapping from Expressivity Score to w is not specified: the text says 'rescale it to the range [0.5, 1.0] using a linear mapping' but gives no formula for standardization, slope/intercept, or clipping. Without this, Eq. 2 is not reproducible and the number of free parameters in the model cannot be assessed.
  3. [Tables II and III and Section III-D] The statement that salience adjustment 'consistently improves performance' is contradicted by the reported point estimates: in Table II, Blueskeye+GPT correlation decreases from 0.653 to 0.620, and in Table III, LSTM correlation is unchanged at 0.725 despite improvements in KLD/RMSE. Please report per-metric significance tests or qualify the claim.
  4. [Tables I-IV] No confidence intervals, standard errors, significance tests, or cross-validation are reported for any metric. With N=100 and correlation differences on the order of 0.02-0.03 (e.g., LSTM 0.725 vs 0.725), the reported improvements cannot be distinguished from noise. Please add uncertainty quantification or paired tests, at least for the headline comparisons.
minor comments (7)
  1. [Section II-A] The sentence 'As BCI treats emotions judgments as a probability distribution over possible labels. This was estimated using the 20 annotations per video' is a grammatical fragment; revise.
  2. [Section II-C] There is a typo: 'clear support for for the Expression-Salience Hypothesis'.
  3. [Section III] There is a typo: 'large langue models' should be 'large language models'.
  4. [Section IV] There is a typo: 'rather that using specialized models' should be 'rather than using specialized models'.
  5. [Section III-C] The fine-tuned LSTM is mentioned as following [20], but the training data split and hyperparameters are not described; please clarify whether any portion of the 100 videos was used for training it.
  6. [Section II-B] Expressivity validation uses only 24 of the 100 videos; please state whether the automatic Expressivity Score was applied to all 100 and whether the w mapping was fit on the full set or the 24-video subset.
  7. [Figure 2] The prompt template referenced as Figure 2 is not visible in the manuscript text; please include the full prompt in the text or an appendix to support reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

The salience weight w in Eq. 2 is fit to Figure 1, which is computed from the same 100-video human context-based judgments used as ground truth in Tables I-IV; Tables I-IV therefore report in-sample fit rather than validated prediction.

  1. fitted input called prediction [Section III-B (Eq. 2) and Section III-C (Table I)]
    "We calculate the weight w based on the Expressivity Score (Section II) and rescale it to the range [0.5, 1.0] using a linear mapping, approximating the findings in Figure 1. ... We first evaluate salience adjustment by seeing if it improves BCI’s fit to the human annotations: i.e., using the human annotations to estimate P(e|f), P(e|c) and P(e|c,f), does Equation (2) better predict the distribution of P(e|c,f) than Equation (1). Table I verifies that salience adjustment improves the accuracy of BCI for both valence and basic emotion prediction, averaging across the 100 videos."

    Figure 1 is constructed from the same P(e|c,f) human context-based judgments on the same 100 videos that Table I treats as ground truth. The only new ingredient in Eq. 2 is w, and w is explicitly chosen to "approximate" the pattern observed in Figure 1 (high expressivity -> judgments closer to face; low expressivity -> equal weighting). Thus Eq. 2 is not derived from an independent theory or a separate training set; it is a parameterization of the very target distribution it is then said to "predict." Comparing Eq. 2 to Eq. 1 on those same P(e|c,f) labels therefore measures in-sample fit of a curve that was shaped by those labels.

  2. fitted input called prediction [Section III-D and Tables II-III]
    "We next validate salience adjustment using automatic methods to estimate the emotion distributions and compare the automatic context-based prediction against the emotion distribution of the human context-based labels. ... The results show that salience adjustment consistently improves performance in both human and automatic emotion recognition tasks."

    The automatic experiments reuse the same corpus-derived w: it is the same linear rescaling "approximating the findings in Figure 1," and Figure 1 was computed from the same 100 videos' human context-based labels that Tables II-IV use as the evaluation target. No held-out split, cross-validation, or independent corpus is introduced when moving from Table I to Tables II-IV. Consequently the claim that salience adjustment "consistently improves" automatic recognition is still evaluated on the videos used to set w; the automatic models add independent feature extractors, but the salience parameter itself is a fixed in-sample statistic of the test set.

full rationale

The circularity is concentrated in how the salience weight is obtained and evaluated. Eq. 2's w is not predicted from a first-principles Bayesian model; it is a linear rescaling of the Expressivity Score chosen to approximate Figure 1, and Figure 1 is a summary of the same human context-based judgments P(e|c,f) on the same 100 videos. Table I then reports that Eq. 2 fits those exact judgments better than Eq. 1, so the headline "salience adjustment improves BCI" is an in-sample curve fit. Tables II-IV carry the same w into automatic pipelines while still scoring against the same human context-based labels and still without any held-out separation, so the "consistently improves" claim inherits the same leakage. The paper does not hide this: the conclusion calls the approach "a heuristic modification of BCI" and notes that a "proper" Bayesian treatment would require a larger corpus. The self-citations to [20] and [27] are methodological (dataset, protocol, expressivity measure) and are not the mechanism that forces the result; the forcing mechanism is the Figure-1-to-w-to-ground-truth chain. Because the effect could in principle have failed per-video even after matching Fig. 1's aggregate tertiles, and because the automatic channels are external models, the circularity is partial rather than total, hence a 6 rather than 8-10.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central mechanism rests on one fitted salience mapping and several domain assumptions inherited from BCI and the annotation protocol. No new physical or conceptual entities are postulated; the invented_entities list is empty.

free parameters (3)
  • Salience weight mapping w(E) = linear rescale of Expressivity Score to [0.5, 1.0]
    Chosen to approximate Figure 1, which is computed from the same 100 videos' context-based human judgments used as ground truth in Tables I-III. This is the main parameter of Eq. 2.
  • VLM frame count = 4 frames
    Picked by testing 2-6 frames on the evaluation set (Section IV-A), so this hyperparameter is tuned to the test data.
  • Expressivity feature weighting = equal weights across standardized AUs, gaze, head pose, optical flow
    Ad hoc equal weighting in Section II-B; affects w but is not fitted against emotion labels.
assumptions (5)
  • domain assumption BCI (Eq. 1) is a valid model of human context-based emotion judgment
    The paper uses Eq. 1 as the baseline and builds Eq. 2 from it; if BCI is not the right base model, the salience adjustment operates on the wrong foundation.
  • standard math P(e) can be treated as uniform and dropped
    Eq. 1 and Eq. 2 use proportionality and never estimate the prior P(e). This is standard Bayes normalization but is not stated explicitly.
  • domain assumption Expressivity score is a valid single-axis measure of facial expressiveness
    Validated on only 24 videos with correlation 0.61 in Section II-B; the score is then used for all 100 videos.
  • ad hoc to paper Linear mapping from expressivity to w approximates human cue integration
    Section III-B states the linear rescaling approximates Figure 1; no theoretical derivation is given, and the authors themselves call the approach a heuristic.
  • domain assumption Human annotator ratings form reliable probability distributions
    20 ratings per video are treated as distributions; no inter-annotator agreement or reliability metric is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Salience Adjustment for Context-Based Emotion Recognition." pith.science (2026). https://pith.science/paper/6Z65WIEX

@misc{pith2026250715878,
  author       = {Pith},
  title        = {Pith review of: Salience Adjustment for Context-Based Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6Z65WIEX}},
  note         = {Machine review of arXiv:2507.15878}
}
read the original abstract

Emotion recognition in dynamic social contexts requires an understanding of the complex interaction between facial expressions and situational cues. This paper presents a salience-adjusted framework for context-aware emotion recognition with Bayesian Cue Integration (BCI) and Visual-Language Models (VLMs) to dynamically weight facial and contextual information based on the expressivity of facial cues. We evaluate this approach using human annotations and automatic emotion recognition systems in prisoner's dilemma scenarios, which are designed to evoke emotional reactions. Our findings demonstrate that incorporating salience adjustment enhances emotion recognition performance, offering promising directions for future research to extend this framework to broader social contexts and multimodal applications.

Figures

Figures reproduced from arXiv: 2507.15878 by the authors.

Figure 1
Figure 1. Proportion of context-based annotations that are best approximated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Salience Adjustment Prompt Template for GPT Visual Reasoning. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 42 canonical work pages

  1. [1]

    Ambadar, J

    Z. Ambadar, J. F. Cohn, and L. I. Reed. All smiles are not created equal: Morphology and timing of smiles perceived as amused, polite, and embarrassed/nervous.Journal of nonverbal behavior, 33:17–34, 2009

    work page 2009

  2. [2]

    M. E. Ansfield. Smiling when distressed: When a smile is a frown turned upside down.Personality and social psychology bulletin, 33(6):763–775, 2007

    work page 2007

  3. [3]

    Anzellotti, S

    S. Anzellotti, S. D. Houlihan, S. Liburd Jr, and R. Saxe. Leveraging facial expressions and contextual information to investigate opaque representations of emotions.Emotion, 21(1):96, 2021

    work page 2021

  4. [4]

    Aviezer, R

    H. Aviezer, R. R. Hassin, J. Ryan, C. Grady, J. Susskind, A. Anderson, M. Moscovitch, and S. Bentin. Angry, disgusted, or afraid? studies on the malleability of emotion perception.Psychological science, 19(7):724–732, 2008

    work page 2008

  5. [5]

    Baltrusaitis, A

    T. Baltrusaitis, A. Zadeh, Y . C. Lim, and L.-P. Morency. Openface 2.0: Facial behavior analysis toolkit. In2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 59–66. IEEE, 2018

    work page 2018

  6. [6]

    L. F. Barrett, B. Mesquita, and M. Gendron. Context in emotion perception.Current directions in psychological science, 20(5):286– 290, 2011

    work page 2011

  7. [7]

    Blagrove and D

    E. Blagrove and D. G. Watson. Ignoring real faces: Effects of valence, threat, and salience.Attention, Perception, & Psychophysics, 76:725– 745, 2014

    work page 2014

  8. [8]

    R. T. Boone and R. Buck. Emotional expressivity and trustworthiness: The role of nonverbal behavior in the evolution of cooperation.Journal of Nonverbal Behavior, 27:163–182, 2003

    work page 2003

Show all 46 references
  1. [9]

    M. M. Bradley and P. J. Lang. Measuring emotion: the self-assessment manikin and the semantic differential.Journal of behavior therapy and experimental psychiatry, 25(1):49–59, 1994

    1994

  2. [10]

    M. G. Calvo and L. Nummenmaa. Detection of emotional faces: salient physical features guide effective visual search.Journal of Experimental Psychology: General, 137(3):471, 2008

    2008

  3. [11]

    M. G. Calvo, L. Nummenmaa, and P. Avero. Visual search of emotional faces: Eye-movement assessment of component processes. Experimental Psychology, 55(6):359–370, 2008

    2008

  4. [12]

    J. A. Coan and J. M. Gottman. The specific affect coding system (spaff).Handbook of emotion elicitation and assessment, 267:285, 2007

    2007

  5. [13]

    C. M. De Melo, P. J. Carnevale, S. J. Read, and J. Gratch. Reading people’s minds from emotion expressions in interdependent decision making.Journal of personality and social psychology, 106(1):73, 2014

    2014

  6. [14]

    C. M. de Melo, J. Gratch, and P. J. Carnevale. Humans versus com- puters: Impact of emotion expressions on people’s decision making. IEEE Transactions on Affective Computing, 6(2):127–136, 2014

    2014

  7. [15]

    C. M. de Melo, K. Terada, and F. C. Santos. Emotion expressions shape human social norms and reputations.Iscience, 24(3), 2021

    2021

  8. [16]

    M. O. Ernst and M. S. Banks. Humans integrate visual and haptic information in a statistically optimal fashion.Nature, 415(6870):429– 433, 2002

    2002

  9. [17]

    E. Fox, V . Lester, R. Russo, R. Bowles, A. Pichler, and K. Dutton. Fa- cial expressions of emotion: Are angry faces detected more efficiently? Cognition & emotion, 14(1):61–92, 2000

    2000

  10. [18]

    J. Ge, H. Luo, S. Qian, Y . Gan, J. Fu, and S. Zhang. Chain of thought prompt tuning in vision language models.arXiv preprint arXiv:2304.07919, 2023

    2023 arXiv

  11. [19]

    S. Goel, J. Jara-Ettinger, D. C. Ong, and M. Gendron. Face and context integration in emotion inference is limited and variable across categories and individuals.Nature Communications, 15(1):2443, 2024

    2024

  12. [20]

    B. Han, C. Yau, S. Lei, and J. Gratch. Knowledge-based emo- tion recognition using large language models.arXiv preprint arXiv:2408.04123, 2024

    2024 arXiv

  13. [21]

    Heesen, M

    R. Heesen, M. A. Szenteczki, Y . Kim, M. E. Kret, A. P. Atkinson, Z. Upton, and Z. Clay. Impact of social context on human facial and gestural emotion expressions.iScience, 27(11), 2024

    2024

  14. [22]

    Hladk `y, R

    M. Hladk `y, R. R. Guerra, X. L. Cang, K. E. MacLean, P. Gebhard, and T. Schneeberger. Modeling the ‘kiss my ass’-smile: Appearance and functions of smiles in negative social situations. In12th International Conference on Affective Computing and Intelligent Interaction (ACII)....

    2024

  15. [23]

    L. A. Jeni, J. F. Cohn, and T. Kanade. Dense 3d face alignment from 2d videos in real-time. In2015 11th IEEE international conference and workshops on automatic face and gesture recognition (FG), volume 1, pages 1–8. IEEE, 2015

    2015

  16. [24]

    Kosti, J

    R. Kosti, J. M. Alvarez, A. Recasens, and A. Lapedriza. Context based emotion recognition using emotic dataset.IEEE transactions on pattern analysis and machine intelligence, 42(11):2755–2766, 2019

    2019

  17. [25]

    R. E. Kraut and R. E. Johnston. Social and emotional messages of smiling: an ethological approach.Journal of personality and social psychology, 37(9):1539, 1979

    1979

  18. [26]

    J. Lee, S. Kim, S. Kim, J. Park, and K. Sohn. Context-aware emotion recognition networks. InProceedings of the IEEE/CVF international conference on computer vision, pages 10143–10152, 2019

    2019

  19. [27]

    Lei and J

    S. Lei and J. Gratch. Emotional expressivity is a reliable signal of surprise.IEEE Transactions on Affective Computing, 14(4):2913– 2924, 2023

    2023

  20. [28]

    Z. Lian, L. Sun, H. Sun, K. Chen, Z. Wen, H. Gu, B. Liu, and J. Tao. Gpt-4v with emotion: A zero-shot benchmark for generalized emotion recognition.Information Fusion, 108:102367, 2024

    2024

  21. [29]

    Littlewort, J

    G. Littlewort, J. Whitehill, T. Wu, I. Fasel, M. Frank, and J. Movellan. The computer expression recognition toolbox (cert). In2011 IEEE International Conference on Automatic Face & Gesture Recognition (FG), pages 298–305, 2011

    2011

  22. [30]

    Masuda, P

    T. Masuda, P. C. Ellsworth, B. Mesquita, J. Leu, S. Tanida, and E. Van de Veerdonk. Placing the face in context: cultural differences in the perception of facial emotion.Journal of personality and social psychology, 94(3):365, 2008

    2008

  23. [31]

    Mortillaro, B

    M. Mortillaro, B. Meuleman, and K. R. Scherer. Advocating a com- ponential appraisal model to guide emotion recognition.International Journal of Synthetic Emotions (IJSE), 3(1):18–32, 2012

    2012

  24. [32]

    P. M. Niedenthal, M. Brauer, L. Robin, and ˚A. H. Innes-Ker. Adult attachment and the perception of facial expression of emotion.Journal of personality and social psychology, 82(3):419, 2002

    2002

  25. [33]

    D. C. Ong, J. Zaki, and N. D. Goodman. Affective cognition: Exploring lay theories of emotion.Cognition, 143:141–162, 10 2015

    2015

  26. [34]

    Poria, D

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

    2018 arXiv

  27. [35]

    Rapoport and A

    A. Rapoport and A. M. Chammah.Prisoner’s dilemma: A study in conflict and cooperation, volume 165. University of Michigan press, 1965

    1965

  28. [36]

    Saxe and S

    R. Saxe and S. D. Houlihan. Formalizing emotion concepts within a bayesian model of theory of mind.Current opinion in Psychology, 17:15–21, 2017

    2017

  29. [37]

    Stratou, J

    G. Stratou, J. Van Der Schalk, R. Hoegen, and J. Gratch. Refactoring facial expressions: An automatic analysis of natural occurring facial expressions in iterative social dilemma. In2017 Seventh International Conference on Affective Computing and Intelligent Interaction (ACII)...

    2017

  30. [38]

    J. B. Tenenbaum, C. Kemp, T. L. Griffiths, and N. D. Goodman. How to grow a mind: Statistics, structure, and abstraction.science, 331(6022):1279–1285, 2011

    2011

  31. [39]

    Toisoul, J

    A. Toisoul, J. Kossaifi, A. Bulat, G. Tzimiropoulos, and M. Pantic. Estimation of continuous valence and arousal levels from faces in naturalistic conditions.Nature Machine Intelligence, 3(1):42–50, 2021

    2021

  32. [40]

    Troiano, L

    E. Troiano, L. Oberl ¨ander, and R. Klinger. Dimensional modeling of emotions in text with appraisal theories: Corpus creation, annotation reliability, and prediction.Computational Linguistics, 49(1):1–72, 03 2023

    2023

  33. [41]

    G. A. Van Kleef, C. K. De Dreu, and A. S. Manstead. The inter- personal effects of emotions in negotiations: a motivated information processing approach.Journal of personality and social psychology, 87(4):510, 2004

    2004

  34. [42]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  35. [43]

    J. Zaki. Cue integration: A common framework for social cognition and physical perception.Perspectives on Psychological Science, 8(3):296–312, 2013

    2013

  36. [44]

    Zhang, B

    R. Zhang, B. Zhang, Y . Li, H. Zhang, Z. Sun, Z. Gan, Y . Yang, R. Pang, and Y . Yang. Improve vision language model chain-of- thought reasoning.arXiv preprint arXiv:2410.16198, 2024

    2024 arXiv

  37. [45]

    Zhang, C

    Y . Zhang, C. Wang, X. Ling, and W. Deng. Learn from all: Erasing attention consistency for noisy label facial expression recognition. In European Conference on Computer Vision, pages 418–434. Springer, 2022

    2022

  38. [46]

    S. Zhao, G. Ding, Y . Gao, and J. Han. Approximating discrete probability distribution of image emotions by multi-modal features fusion.Transfer, 1000(1):4669–4675, 2017

    2017

Pith tools

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