acceptodds
Under review as a conference paper at ICLR 2027

Beyond Fixed Emotion Mappings: Learning Probabilistic Relations Across Emotion Granularities

Abstract

Emotion recognition is naturally hierarchical: fine-grained affective states such as grief or annoyance relate to broader categories such as sadness or anger. Existing emotion taxonomies specify such categorical relationships, but knowing the hierarchy does not uniquely determine how probability mass in a fine-grained predictive posterior should propagate into a coarse-grained distribution. We therefore formulate cross-granularity emotion prediction as learning a probabilistic relationship between emotion granularities while retaining known hierarchical structure. Our hierarchical consistency framework jointly optimizes distinct fine- and coarse-grained predictors together with a learnable column-stochastic transformation . The transformation maps the complete fine-grained posterior to an induced coarse posterior, which is coupled with a directly predicted coarse posterior through distribution-level consistency. Structural regularization anchors to a soft empirical mapping derived exclusively from the multi-label training data. Experiments on GoEmotions show that deterministic, empirical probabilistic, and learned probabilistic mappings yield nearly identical coarse-grained top-1 accuracy (–) while inducing meaningfully different predictive distributions. Relative to the training-derived empirical mapping, a learned transformation achieves a relative reduction in expected calibration error, while paired-bootstrap analysis supports reductions in negative log-likelihood for learned mappings. Importantly, the learned mappings preserve the dominant Ekman assignment of every fine-grained emotion. These results demonstrate that a known categorical hierarchy does not uniquely determine the probabilistic relationship across label granularities: mappings with nearly identical categorical behavior can exhibit systematically different predictive distributions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Emotion Recognition, Hierarchical Classification, Cross-Granularity Learning, Probabilistic Mapping, Hierarchical Consistency, Model Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1YPKOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether learning a fine-to-coarse emotion mapping can improve probability quality without substantially changing classification accuracy. The framework jointly trains fine- and coarse-grained predictors with a learnable mapping, using consistency and structural regularization. Experiments on GoEmotions report improvements in NLL and ECE, with little change in coarse-grained accuracy.

Critical strengths
  1. The paper addresses a useful question: whether learned mappings can improve predictive distributions even when classification accuracy remains largely unchanged.

  2. The evaluation includes NLL, Brier score, and ECE alongside accuracy and F1, which is appropriate for assessing probability quality.

Critical weaknesses
  1. The method combines familiar components but lacks comparisons with relevant transition models, multi-task baselines, and simpler calibration methods. Its novelty and the source of the gains are therefore difficult to assess.

  2. Experiments cover only one dataset, taxonomy, and backbone, leaving the generality of the findings unclear.

  3. The handling of multi-label examples needs clarification, particularly how co-occurrence counts are constructed and whether bootstrap sampling accounts for multiple rows from the same utterance.

  4. A frozen posterior at evaluation does not establish that the learned relations are stable. Since the mapping and encoders are trained jointly, the mapping may partly compensate for encoder errors.

  5. Figure 1 needs better resolution and clearer connections, and the hard-mapping description is unclear. Citation formatting errors, and missing implementation details also weaken the presentation.

Questions for authors
  1. How does the method compare with transition models, shared-encoder multi-task baselines, and standard calibration methods under comparable conditions?

  2. Do the gains hold on another dataset or taxonomy, or with a different backbone?

  3. How are fine-coarse co-occurrences counted for multi-label examples? Does bootstrap sampling operate on original utterances or expanded rows?

  4. Do the gains persist when the mapping is learned with the encoders held fixed? How stable is the mapping across encoder checkpoints?

Reviewer_wLfkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the cross-granularity probability propagation problem in fine-grained to coarse-grained sentiment prediction, distinguishing between the predefined hard mapping MhardM_{\mathrm{hard}}, the training set empirical mapping (MempM_{\mathrm{emp}}), and the learnable column random mapping MθM_\theta. It proposes the Hierarchical Consistency Framework: independent fine/coarse encoders, (p^c=Mθpf\hat p^c = M_\theta p^f), KL consistency constraints, and structural regularization towards (MempM_{\mathrm{emp}}). Experiments show that the coarse-grained top-1 performance of the three mappings is close, but there are differences in the probability quality metrics; learning the mappings mainly redistributes the non-dominant probability quality, rather than changing the dominant coarse-class assignment of the hard mapping.

Critical strengths
  1. The distinction of the problem is valuable. The paper clearly distinguishes between "hierarchical structure of categories" and "how probability quality propagates across granularity levels", which is more detailed than a simple hierarchical classification.
  2. The experimental control is cleaner. Table 2/3 replaces the mapping (MM) on the same frozen fine posterior (pf(x)p^f(x)), effectively isolating the influence of the mapping itself.
  3. The evaluation dimensions are comprehensive. It reports Accuracy, Macro-F1, NLL, Brier, ECE, column entropy, and (L1L_1) distance simultaneously, and conducts paired bootstrap.
  4. The structural regularization analysis is reasonable. Tables 2 and 6 support that as (γ\gamma) increases, (MθM_\theta) approaches (MempM_{\mathrm{emp}}) and the column entropy increases.
  5. The discussion is restrained. The authors do not claim that learning the mapping overturns the hard hierarchy, but rather admit that it retains the dominant coarse categories and adjusts the non-dominant probability quality.
  6. The reproducibility is acceptable. The data division, hyperparameters, algorithms, and bootstrap settings are clear; (MempM_{\mathrm{emp}}) is only constructed using the training set to avoid obvious leakage.
Critical weaknesses
  1. The caption of Table 1 is ambiguous. The caption only states "GoEmotions test set", but there are 27 types of Fine classifiers and 6 types of coarse paths presented in the same table; readers cannot determine whether Direct coarse is evaluated in the 6-class space or mapped back to the 27-class evaluation.
  2. The juxtaposition of cross-granularity results is misleading. The Fine 27 types and Coarse 6 types cannot be directly compared; in Section 4.2, the description "6 types are higher than 27 types" can be understood as having lower ambiguity for the coarse task, but the paragraph does not clearly state that "0.7834 vs 0.7837" is compared within the same 6-class coarse space.
  3. The independent value of the transfer matrix is insufficient. In Table 1, the Accuracy/Macro-F1 of Direct coarse is 0.7837/0.6643, Hard is 0.7826/0.6582, and Learned is 0.7834/0.6551; Hard/Learned does not exceed Direct coarse. The paper does not explain the application scenarios and necessity of additional learning (MθM_\theta).
  4. Format and layout issues. Figure 1 is not a vector graph, and the text and symbols become blurry when enlarged; Algorithm 1 is inserted in the middle of the references, and the reading order is unreasonable;
  5. There is a closed-loop risk in the derivation of hard coarse labels and consistency loss. (ycy^c) is determinedly derived from (yfy^f) through the predefined Ekman mapping, and coarse supervision is not independent; while (Lcons=DKL(pc∥p^c)\mathcal L_{\mathrm{cons}}=D_{\mathrm{KL}}(p^c\|\hat p^c)) pulls (pc^\hat{p^c}) towards the (pcp^c) supervised by the hard derived labels, which may limit the deviation of (MθM_\theta) from the hard structure. The paper does not discuss this risk, nor does it have an ablation study with (λ=0\lambda=0).
Questions for authors

Same to weaknesses.

Reviewer_QajJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a hierarchical emotion classification framework that learns probabilistic relationships between fine-grained and coarse-grained emotion categories. The framework combines two independent BERT encoders, a learnable fine-to-coarse transition matrix, distribution-level consistency, and structural regularization. Experiments are conducted exclusively on GoEmotions using 27 fine-grained emotions and six Ekman categories.

Critical strengths
  1. The paper addresses a relevant problem concerning probabilistic relationships between fine-grained and coarse-grained emotion categories.
  2. The proposed framework has a clear mathematical formulation and explicitly models cross-granularity probability distributions.
  3. The authors evaluate probability quality using NLL, Brier score, and ECE, in addition to Accuracy and Macro-F1.
  4. The paper includes sensitivity analyses of regularization coefficients and paired-bootstrap confidence intervals for NLL and Brier score.
Critical weaknesses
  1. Learning probabilistic transition matrices between hierarchical label spaces has already been investigated, particularly in Label Hierarchy Transition (Wang et al., IEEE TPAMI 2026). The proposed framework primarily combines a global learnable transition matrix with standard classification, KL-divergence consistency, and structural regularization. Its methodological distinction and advantages over existing approaches are insufficiently demonstrated.
  2. The experimental evaluation compares deterministic, empirical, and learned mappings but does not include competing hierarchical classification methods, including the closely related Label Hierarchy Transition framework. Consequently, the proposed method is not evaluated against established alternatives.
  3. The learned mapping with γ = 0.5 improves coarse-grained Accuracy by only 0.08 percentage points over the deterministic mapping, while Macro-F1 decreases from 65.82% to 65.51%. Compared with the empirical mapping, Accuracy improves by only 0.01 points, and Macro-F1 remains slightly lower. These results provide little evidence of improved emotion recognition.
  4. The main positive result is improved probability calibration, particularly the reduction in ECE from 0.0342 to 0.0254 relative to the empirical mapping. However, the paper does not compare against standard post-hoc calibration techniques such as temperature scaling, vector scaling, or Dirichlet calibration, nor against recent methods specifically developed for emotion classification.
  5. All experiments use only GoEmotions, one BERT-base architecture, and a single Ekman-based hierarchy. No additional datasets, alternative label taxonomies, or modern language-model architectures are evaluated. This substantially limits the evidence for the proposed framework's general applicability.
  6. GoEmotions originally supports multiple simultaneous emotion labels, but the authors expand multi-label instances into separate single-label examples and optimize softmax-based classification objectives. This transformation changes the original task formulation and may introduce conflicting supervision for identical texts, raising questions about the interpretation and calibration of the resulting probability distributions.
  7. The empirical and learned mappings produce almost identical categorical predictions. Appendix Table 7 reports only two disagreements among 9,192 test instances for the empirical and learned mapping with γ = 1.0. The observed differences in probability quality are small, and the paper does not establish whether they justify learning an additional transformation.
  8. The framework employs two independently parameterized BERT encoders for fine- and coarse-grained prediction. The authors do not compare this design with a shared encoder and two classification heads or analyze its computational overhead, training time, and inference efficiency.
  9. Although paired-bootstrap confidence intervals are reported for NLL and Brier score, the experiments use a single random seed. Bootstrap resampling of fixed predictions does not capture variability across independent training runs. The robustness of the reported ECE improvements is also insufficiently established.
  10. The learned transition matrix is intended to capture probabilistic relationships between emotion categories, but the paper does not provide a detailed visualization or semantic analysis of the learned mappings. Furthermore, no downstream application demonstrates the practical value of the relatively small calibration improvements.
  11. The paper attributes improvements in probability quality to learned cross-granularity relationships, but the experiments do not establish whether these gains arise from meaningful hierarchical modeling or simply from additional probability smoothing and regularization.
Questions for authors
  1. What is the main methodological novelty relative to existing approaches that learn probabilistic transitions between hierarchical label spaces, particularly Wang et al. (IEEE TPAMI 2026)? Can the authors demonstrate the advantages of their framework through direct comparisons under matched experimental conditions?
  2. Given that the learned mapping provides virtually no improvement in Accuracy or Macro-F1 over the empirical mapping, what justifies its additional complexity? Can the authors compare their approach with standard post-hoc calibration methods and demonstrate whether the reported ECE improvement provides a meaningful practical benefit?
  3. Why is the original multi-label GoEmotions annotation converted into separate single-label instances? How does this transformation affect conflicting labels, learned fine-to-coarse relationships, and the interpretation of predictive probabilities? Would the conclusions remain valid under the original multi-label formulation?
  4. Why is the evaluation restricted to GoEmotions and BERT-base? Can the authors demonstrate that the proposed method generalizes across different datasets, emotion taxonomies, and model architectures, and report variability across independent training runs?
  5. Why are two independent BERT encoders necessary instead of a shared encoder with separate prediction heads? Can the authors isolate the contributions of the learnable transition matrix, consistency loss, and structural regularization while reporting their computational costs and comparing them with simpler alternatives?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.