Pith. sign in

REVIEW 3 major objections 4 minor 53 references

Contrastive Distillation of Emotion Knowledge from LLMs for Zero-Shot Emotion Recognition

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A compact BERT-sized model, trained only on GPT-4's free-form emotion descriptors, performs zero-shot emotion recognition across unseen label sets, approaching GPT-4's accuracy while being over 10,000 times smaller.

desk verdict Useful distillation recipe with real cross-dataset numbers, but the zero-shot label-generalization claim is undercut by likely overlap between GPT-4 descriptors and test labels. read the letter →

arxiv 2505.18040 v1 pith:GQD62BWY submitted 2025-05-23 cs.CL

Minxue Niu , Emily Mower Provost This is my paper · ORCID
classification cs.CL
keywords zero-shotemotionrecognitioncontrastivedistillationLLMknowledgedescriptorsmulti-labelclassificationembeddingGPT-4supervisionGoEmotions
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 tries to establish that a compact, BERT-sized model can inherit GPT-4's flexible emotion understanding without any human annotations or retraining, by learning from free-form emotion descriptors instead of fixed label sets. The authors prompt GPT-4 to describe the emotions in 43.4k GoEmotions texts, collect 2,173 unique emotional terms and phrases, and train a contrastive model to align each text with its descriptors in a shared emotion embedding space. At inference the fixed label set of a new dataset is embedded with the same label encoder, letting the model classify or regress on emotions it never encountered during training. If correct, this makes zero-shot emotion recognition practical on edge devices and removes the need for dataset-specific annotation and fine-tuning.

What carries the argument

The central mechanism is a multi-label extension of the CLIP contrastive objective operating on a shared emotion embedding space. A frozen BERT label encoder and a trainable BERT text encoder with a Q-Former projection produce normalized embeddings; the loss maximizes the sigmoid probability of matched text–descriptor pairs and minimizes it for unmatched pairs, averaging positive and negative terms separately to handle sparse alignments. Because labels are embedded rather than read from a fixed classification head, the inference-time label set can be swapped freely, and unseen labels are scored by cosine similarity to the text embedding. The pivotal data object is the GPT-4-generated descriptor set: 2,173 unique terms, deliberately retaining morphological and stylistic variability such as 'fear', 'fearful', and 'fear of abandonment'.

What would settle it

Train the same contrastive architecture on human GoEmotions labels whose vocabulary has been expanded to a comparable number of unique terms (e.g., by adding synonyms and morphological variants), then test zero-shot on SemEval and ISEAR. If the vocabulary-expanded human-label model matches or beats the GPT-4-distilled model, the benefit is merely descriptor diversity; if it still lags, GPT-4's specific emotion knowledge is doing the work.

Watch

Extended reading notes

Core claim

The central claim is that GPT-4's free-form emotion descriptors are a more generalizable supervision signal than fixed categorical human labels for learning a reusable emotion embedding. Trained only on GPT-4's descriptions of GoEmotions text, the distilled BERT model achieves zero-shot macro-F1 of 0.480 on SemEval multi-label classification, 0.479 on ISEAR single-label classification, and a Pearson correlation of 0.593 on EmoBank valence regression, approaching GPT-4's zero-shot performance while being over 10,000 times smaller. The same contrastive architecture trained on human GoEmotions labels performs better on seen classes but drops sharply on unseen classes (0.161 on SemEval and 0.215 on ISEAR), which the authors take as evidence that descriptor diversity, not architecture, drives zero-shot generalization. A nearest-neighbor analysis shows the learned emotion space retrieves emotionally aligned terms rather than part-of-speech matches, and this structure survives compression from 768 to 50 dimensions.

Load-bearing premise

The load-bearing premise is that GPT-4's free-form emotion descriptors are a richer and more generalizable supervision signal than human categorical labels; if GPT-4's vocabulary and interpretive style diverge from how target populations label emotions, zero-shot performance will degrade.

Editorial extensions

If this is right

  • A single compact model can serve multiple emotion recognition applications with different label schemas, granularities, and task formats without retraining.
  • Zero-shot emotion recognition no longer requires human annotations for new label sets; raw text plus LLM-generated descriptors suffice for training.
  • Deployment on edge devices becomes feasible, since the distilled model is BERT-sized rather than a trillion-parameter LLM.
  • The contrastive emotion space preserves emotion-specific similarity structure, making model outputs more interpretable through nearest-neighbor retrieval.
  • The approach's ceiling is tied to GPT-4's descriptive style; any shift in that style shifts downstream zero-shot performance.

Reading between the lines

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

  • The same recipe could apply to other subjective annotation tasks, such as sentiment intensity, stance, or personality, where free-form descriptions may generalize better than fixed taxonomies.
  • A testable extension is mixing GPT-4 descriptors with human labels at different ratios to map the trade-off between familiarity with a target label set and generalization to unseen ones.
  • Because the label encoder remains frozen BERT, the model inherits BERT's linguistic and cultural biases; swapping in a multilingual or domain-adapted encoder is a natural experiment for measuring cross-cultural zero-shot shifts.
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 / 4 minor

Summary. The paper proposes a contrastive distillation framework that transfers emotion knowledge from GPT-4 into a compact BERT-based model. GPT-4 generates free-form, multi-label emotion descriptors for GoEmotions training text (2,173 unique terms). A CLIP-style contrastive loss aligns text and descriptor embeddings in a shared space. At inference the text encoder is paired with fixed label sets from target datasets, enabling zero-shot classification and regression. The approach is evaluated on GoEmotions (multi-label), SemEval (multi-label), ISEAR (single-label), and EmoBank (valence/activation regression), and it is compared against zero-shot baselines, GPT-4 zero-shot, and a fine-tuned BERT. The model reaches 0.480 macro-F1 on SemEval and 0.593 PCC on EmoBank valence, outperforming comparably sized baselines in several settings while remaining far smaller than GPT-4.

Significance. If the central claim holds, the paper provides a practical way to build a compact emotion recognition model that can adapt to new label spaces without human annotations, which would be valuable for edge deployment. The paper is transparent about its limitations, releases code and data, and includes a useful ablation contrasting GPT-4 supervision with human supervision. The contrastive formulation is simple and well-motivated. However, the main zero-shot generalization claim is not yet established at the label-vocabulary level, and the multi-label results depend on validation-set threshold tuning, so the current evidence is suggestive but not conclusive.

major comments (3)
  1. [§3.1, §5.2, Table 3] The central claim is zero-shot generalization to new emotion label spaces, but the paper never verifies that the evaluation labels were absent from the 2,173 GPT-4 descriptors used at training time. The datasets evaluated (SemEval, ISEAR, EmoBank) use common emotion terms such as 'joy', 'fear', 'guilt', 'shame', 'optimism', and 'trust', which very likely appear verbatim in GPT-4's free-form descriptors. If so, the model has directly seen those exact label strings in training, so the reported performance does not demonstrate generalization to genuinely unseen label vocabularies. This also undermines the Table 3 ablation: comparing a model trained on 2,173 labels to a model trained on 27 labels confounds vocabulary coverage with supervision richness. The authors should report the overlap between the 2,173 descriptors and every evaluation label set, and ideally add a held-out-label experiment in which the target emotion words are verifiably absent from the training descriptors (e.g., using synonyms or a novel taxonomy). Without such evidence, the headline claim of 'zero-shot prediction on different emotion classes' is not established.
  2. [§3.3, §4.3, Limitations] For multi-label classification, the paper tunes per-class thresholds on a validation set and acknowledges that this setup is 'not strictly zero-shot.' The abstract and introduction nevertheless emphasize zero-shot multi-label prediction, and the main numbers in Table 1 for GoEmotions and SemEval are obtained after this calibration. The authors report that relative rankings are preserved without calibration, but the actual magnitudes are threshold-dependent. This is a load-bearing presentation issue: either the multi-label results should be reported primarily in a strictly zero-shot form (e.g., using a fixed default threshold at 0.5 or a single global threshold), or the paper should quantify how much validation data and how many threshold searches are used and clearly state that multi-label performance is a few-shot, calibration-based result. The current text overstates the purity of the zero-shot setup.
  3. [§3.3 (regression conversion)] The regression experiments convert valence and activation into four fixed phrases ('Positivity', 'Negativity', 'High Activation', 'Low Activation') and compute predictions as differences of alignment scores. This is an ad-hoc representational choice that is not derived from the GPT-4 descriptors, and the paper does not state whether these exact phrases occur in the 2,173 training descriptors. If they do occur, the regression result is again susceptible to the vocabulary-overlap issue; if they do not, it would provide valuable evidence of genuine generalization. The authors should report whether these four terms appear in the training descriptor set and, if they do, replace them with non-overlapping terms to make the regression evaluation truly zero-shot.
minor comments (4)
  1. [Abstract and §4.2] The paper claims GPT-4 is 'over 10,000 times larger' than BERT-base and that the model is 'over 10,000 times smaller,' but using the authors' own numbers (GPT-4 >100B, BERT 110M) the ratio is roughly 900x, not 10,000x. This should be corrected for accuracy.
  2. [§3.3] Typo: 'adapter' should be 'adapt' in the sentence 'the model can flexibly adapter to different emotion label spaces.'
  3. [Table 3] The table row labels 'seen classes' and 'unseen classes' are not immediately interpretable; please clarify that these refer to the human-supervised model's training label set, and specify the exact number of seen/unseen classes per dataset in a caption or footnote.
  4. [§5.3, Table 4] The nearest-neighbor analysis is qualitative only. Consider reporting a quantitative metric (e.g., agreement with a human-rated emotion similarity or retrieval accuracy on a small labeled evaluation set) to support the claim that the learned space is more emotionally aligned than BERT.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: test-time human labels are independent of GPT-4 descriptor supervision, and reported metrics are not fitted constants.

full rationale

The derivation chain is: (1) prompt GPT-4 to produce free-form emotion descriptors for GoEmotions text (Section 3.1); (2) train a BERT-based text encoder and a frozen-BERT label encoder with a multi-label contrastive sigmoid loss to align texts with those descriptors (Section 3.2); (3) at inference, replace the descriptor set with each target dataset's human label set and score text-label similarity (Section 3.3). The test-time labels for SemEval, ISEAR, and EmoBank are human annotations that were never used as supervision, and the reported macro-F1 and PCC numbers are measured against those independent labels. The per-class thresholds in multi-label evaluation are calibrated on validation splits, but this is disclosed, applied equally to baselines, and does not make test metrics follow from the training loss by construction. The only self-citation is the adopted GPT-4 prompting strategy from Niu et al. (2024); this supplies a design choice, not an unverified uniqueness theorem or a fitted parameter, so it does not make the central result circular. The vocabulary-overlap concern (test labels such as 'joy' may also appear among the 2,173 GPT-4 descriptors) weakens the 'unseen label' interpretation, but it is an experimental confound rather than a definitional equivalence. No equation in the paper reduces its predictions to its training inputs by construction, and the paper is self-contained against external benchmarks. Therefore no circular step is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central empirical claim relies on the quality of GPT-4 supervision, BERT's semantic knowledge of emotion terms, and the regression mapping. Model weights are standard trained parameters; the only data-fitted constants beyond training are per-class thresholds for multi-label tasks.

free parameters (3)
  • temperature tau = 0.1
    Fixed hyperparameter in the contrastive alignment score; chosen by hand, not tuned per dataset.
  • emotion space dimension d = 768 (ablated to 200, 100, 50)
    Chosen to match baseline BERT dimension; ablation shows small sensitivity.
  • per-class multi-label thresholds = searched in [0,1] in steps of 0.05 on validation sets
    Calibrated per emotion class for GoEmotions and SemEval multi-label predictions; the paper acknowledges this is a weak-supervision step that affects F1 scores.
assumptions (4)
  • domain assumption GPT-4-generated emotion descriptors are a richer and more generalizable supervision signal than human categorical labels.
    Invoked in Section 3.1 and Section 5.2; if false, the method's advantage over fixed-label training disappears.
  • domain assumption Frozen BERT embeddings of emotion words, after a linear projection, can align with learned text embeddings for unseen labels.
    Invoked in Section 3.2; the label encoder is frozen BERT, so zero-shot generalization depends on BERT's semantic coverage of novel descriptors.
  • ad hoc to paper The four discrete categories Positivity, Negativity, High Activation, and Low Activation can represent valence and activation for regression.
    Introduced in Section 3.3 to map continuous emotion dimensions onto the contrastive space; a lossy mapping would make regression scores less meaningful.
  • standard math Normalized dot-product similarity in the learned space is a valid measure of emotion alignment.
    Standard CLIP-style assumption used in the alignment matrix and inference in Sections 3.2 and 3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Distillation of Emotion Knowledge from LLMs for Zero-Shot Emotion Recognition." pith.science (2026). https://pith.science/paper/GQD62BWY

@misc{pith2026250518040,
  author       = {Pith},
  title        = {Pith review of: Contrastive Distillation of Emotion Knowledge from LLMs for Zero-Shot Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GQD62BWY}},
  note         = {Machine review of arXiv:2505.18040}
}
read the original abstract

The ability to handle various emotion labels without dedicated training is crucial for building adaptable Emotion Recognition (ER) systems. Conventional ER models rely on training using fixed label sets and struggle to generalize beyond them. On the other hand, Large Language Models (LLMs) have shown strong zero-shot ER performance across diverse label spaces, but their scale limits their use on edge devices. In this work, we propose a contrastive distillation framework that transfers rich emotional knowledge from LLMs into a compact model without the use of human annotations. We use GPT-4 to generate descriptive emotion annotations, offering rich supervision beyond fixed label sets. By aligning text samples with emotion descriptors in a shared embedding space, our method enables zero-shot prediction on different emotion classes, granularity, and label schema. The distilled model is effective across multiple datasets and label spaces, outperforming strong baselines of similar size and approaching GPT-4's zero-shot performance, while being over 10,000 times smaller.

Figures

Figures reproduced from arXiv: 2505.18040 by the authors.

Figure 1
Figure 1. Our model is trained with rich emotion de [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our contrastive distillation model structure. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 42 canonical work pages

  1. [1]

    Abdullah Al Maruf, Fahima Khanam, Md Mahmudul Haque, Zakaria Masud Jiyad, Muhammad Firoz Mridha, and Zeyar Aung. 2024. Challenges and opportunities of text-based emotion detection: a survey. IEEE access, 12:18416--18450

    2024

  2. [2]

    o l \"u c \

    Iqra Ameer, Necva B \"o l \"u c \"u , Muhammad Hammad Fahim Siddiqui, Burcu Can, Grigori Sidorov, and Alexander Gelbukh. 2023. Multi-label emotion classification in texts using transfer learning. Expert Systems with Applications, 213:118534

    work page 2023

  3. [3]

    Patrick Barei , Roman Klinger, and Jeremy Barnes. 2024. English prompts are better for nli-based zero-shot emotion classification than target-language prompts. In Companion Proceedings of the ACM Web Conference 2024, pages 1318--1326

    work page 2024

  4. [4]

    Ankita Bhaumik and Tomek Strzalkowski. 2024 a . Towards a generative approach for emotion detection and reasoning. arXiv [cs.CL]

    work page 2024

  5. [5]

    Ankita Bhaumik and Tomek Strzalkowski. 2024 b . Towards a generative approach for emotion detection and reasoning. arXiv preprint arXiv:2408.04906

    work page Pith review arXiv 2024

  6. [6]

    Su Bo-Hao, Shreya G Upadhyay, and Lee Chi-Chun. 2025. Toward zero-shot speech emotion recognition using llms in the absence of target data. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5. IEEE

    work page 2025

  7. [7]

    Sven Buechel and Udo Hahn. 2017. E mo B ank : Studying the impact of annotation perspective and representation format on dimensional emotion analysis. In Proceedings of the 15th Conference of EACL, pages 578--585, Valencia, Spain

    work page 2017

  8. [8]

    Sven Buechel, Luise Modersohn, and Udo Hahn. 2021. Towards label-agnostic emotion embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9231--9249

    work page 2021

Show all 53 references
  1. [9]

    Georgios Chochlakis, Gireesh Mahajan, Sabyasachee Baruah, Keith Burghardt, Kristina Lerman, and Shrikanth Narayanan. 2023. Using emotion embeddings to transfer knowledge between emotions, languages, and annotation formats. In ICASSP 2023-2023 IEEE International Conference on A...

    2023

  2. [10]

    Roddy Cowie and Randolph R Cornelius. 2003. Describing the emotional states that are expressed in speech. Speech communication, 40(1-2):5--32

    2003

  3. [11]

    Roddy Cowie, Ellen Douglas-Cowie, Nicolas Tsapatsoulis, George Votsis, Stefanos Kollias, Winfried Fellenz, and John G Taylor. 2001. Emotion recognition in human-computer interaction. IEEE Signal processing magazine, 18(1):32--80

    2001

  4. [12]

    Flor Miriam Plaza Del Arco, Mar \' a-Teresa Mart \' n-Valdivia, and Roman Klinger. 2022. Natural language inference prompts for zero-shot emotion classification in text across corpora. In Proceedings of the 29th International Conference on Computational Linguistics, pages 6805--6817

    2022

  5. [13]

    Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. 2020. Goemotions: A dataset of fine-grained emotions. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4040--4054

    2020

  6. [14]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human...

    2019

  7. [15]

    Kexin Feng and Theodora Chaspari. 2020. A review of generalizable transfer learning in automatic emotion recognition. Frontiers in Computer Science, 2:9

    2020

  8. [16]

    Yuan Gao, Longbiao Wang, Jiaxing Liu, Jianwu Dang, and Shogo Okada. 2023. Adversarial domain generalized transformer for cross-corpus speech emotion recognition. IEEE Transactions on Affective Computing, 15(2):697--708

    2023

  9. [17]

    Katie Hoemann, Evan Warfel, Caitlin Mills, Laura Allen, Peter Kuppens, and Jolie B Wormwood. 2024. Using freely generated labels instead of rating scales to assess emotion in everyday life. Assessment, page 10731911241283623

    2024

  10. [18]

    Chul Min Lee and Shrikanth S Narayanan. 2005. Toward detecting emotions in spoken dialogs. IEEE transactions on speech and audio processing, 13(2):293--303

    2005

  11. [19]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730--19742. PMLR

    2023

  12. [20]

    Zheng Lian, Licai Sun, Haiyang Sun, Kang Chen, Zhuofan Wen, Hao Gu, Bin Liu, and Jianhua Tao. 2024. Gpt-4v with emotion: A zero-shot benchmark for generalized emotion recognition. Information Fusion, 108:102367

    2024

  13. [21]

    Yuanyuan Liu, Ke Wang, Lin Wei, Jingying Chen, Yibing Zhan, Dapeng Tao, and Zhe Chen. 2024 a . Affective computing for healthcare: Recent trends, applications, challenges, and beyond. arXiv preprint arXiv:2402.13589

    2024 arXiv

  14. [22]

    Zhiwei Liu, Kailai Yang, Qianqian Xie, Tianlin Zhang, and Sophia Ananiadou. 2024 b . Emollms: A series of emotional large language models and annotation tools for comprehensive affective analysis. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data...

    2024

  15. [23]

    Rui Mao, Qian Liu, Kai He, Wei Li, and Erik Cambria. 2023. The biases of pre-trained language models: An empirical study on prompt-based sentiment analysis and emotion detection. IEEE Transactions on Affective Computing, 14(3):1743--1753

    2023

  16. [24]

    Saif Mohammad, Felipe Bravo-Marquez, Mohammad Salameh, and Svetlana Kiritchenko. 2018. SemEval -2018 task 1: Affect in tweets. In Proceedings of The 12th International Workshop on Semantic Evaluation, pages 1--17, Stroudsburg, PA, USA. Association for Computational Linguistics

    2018

  17. [25]

    Tong Mu, Alec Helyar, Johannes Heidecke, Joshua Achiam, Andrea Vallone, Ian Kivlichan, Molly Lin, Alex Beutel, John Schulman, and Lilian Weng. 2024. Rule based rewards for language model safety. arXiv preprint arXiv:2411.01111

    2024 arXiv

  18. [26]

    Minxue Niu, Mimansa Jaiswal, and Emily Mower Provost. 2024. From text to emotion: Unveiling the emotion annotation capabilities of llms. In Proc. Interspeech 2024, pages 2650--2654

    2024

  19. [27]

    Justin Olah, Sabyasachee Baruah, Digbalay Bose, and Shrikanth Narayanan. 2021. Cross domain emotion recognition using few shot knowledge transfer. arXiv [cs.CL]

    2021

  20. [28]

    Miros aw P aza, Robert Kaza a, Zbigniew Koruba, Marcin Koz owski, Ma gorzata Luci \'n ska, Kamil Sitek, and Jaros aw Spyrka. 2022. Emotion recognition method for call/contact centre systems. Applied Sciences, 12(21):10951

    2022

  21. [29]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural language supervision. In International conference on machine...

    2021

  22. [30]

    Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370--5381

    2019

  23. [31]

    Abdur Rasool, Saba Aslam, Naeem Hussain, Sharjeel Imtiaz, and Waqar Riaz. 2025. nbert: Harnessing nlp for emotion recognition in psychotherapy to transform mental health care. Information, 16(4):301

    2025

  24. [32]

    Klaus R Scherer, Elizabeth Clark-Polner, and Marcello Mortillaro. 2011. In the eye of the beholder? universality and cultural specificity in the expression and perception of emotion. International Journal of Psychology, 46(6):401--435

    2011

  25. [33]

    Eimear Stanley, Eric DeMattos, Anita Klementiev, Piotr Ozimek, Georgia Clarke, Michael Berger, and Dimitri Palaz. 2023. Emotion label encoding using word embeddings for speech emotion recognition. In Proceedings of the INTERSPEECH, pages 2418--2422

    2023

  26. [34]

    Ala N Tak and Jonathan Gratch. 2023. Is GPT a computational model of emotion? In 2023 11th International Conference on Affective Computing and Intelligent Interaction ( ACII ) , pages 1--8. IEEE

    2023

  27. [35]

    Ala N Tak and Jonathan Gratch. 2024. Gpt-4 emulates average-human emotional cognition from a third-person perspective. arXiv preprint arXiv:2408.13718

    2024 arXiv

  28. [36]

    Senait Gebremichael Tesfagergish, Jurgita Kapo c i \=u t \.e -Dzikien \.e , and Robertas Dama s evi c ius. 2022. Zero-shot emotion detection for semi-supervised sentiment analysis using sentence transformers and ensemble learning. Applied Sciences, 12(17):8662

    2022

  29. [37]

    Marcel Trotzek, Sven Koitka, and Christoph M Friedrich. 2018. Utilizing neural networks and linguistic metadata for early detection of depression indications in text sequences. IEEE Transactions on Knowledge and Data Engineering, 32(3):588--601

    2018

  30. [38]

    Manh Tu Vu, Marie Beurton-Aimar, and Serge Marchand. 2021. Multitask multi-database emotion recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3637--3644

    2021

  31. [39]

    Johannes Wagner, Andreas Triantafyllopoulos, Hagen Wierstorf, Maximilian Schmitt, Felix Burkhardt, Florian Eyben, and Bj \"o rn W Schuller. 2023. Dawn of the transformer era in speech emotion recognition: closing the valence gap. IEEE Transactions on Pattern Analysis and Machi...

    2023

  32. [40]

    Harald G Wallbott and Klaus R Scherer. 1986. How universal and specific is emotional experience? evidence from 27 countries on five continents. Social science information, 25(4):763--795

    1986

  33. [41]

    Xuena Wang, Xueting Li, Zi Yin, Yue Wu, and Jia Liu. 2023. Emotional intelligence of large language models. Journal of Pacific Rim Psychology, 17

    2023

  34. [42]

    Congying Xia, Chen Xing, Jiangshu Du, Xinyi Yang, Yihao Feng, Ran Xu, Wenpeng Yin, and Caiming Xiong. 2024. Fofo: A benchmark to evaluate llms' format-following capability. In 62nd Annual Meeting of the Association for Computational Linguistics, pages 680--699

    2024

  35. [43]

    Tian Xu, Jennifer White, Sinan Kalkan, and Hatice Gunes. 2020. Investigating bias and fairness in facial expression recognition. In Computer Vision--ECCV 2020 Workshops: Glasgow, UK, August 23--28, 2020, Proceedings, Part VI 16, pages 506--523. Springer

    2020

  36. [44]

    Yifan Yao, Jinhao Duan, Kaidi Xu, Yuanfang Cai, Zhibo Sun, and Yue Zhang. 2024. A survey on large language model (llm) security and privacy: The good, the bad, and the ugly. High-Confidence Computing, page 100211

    2024

  37. [45]

    Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. In 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing, ...

    2019

  38. [46]

    u ksel Yurtay, H \

    Y \"u ksel Yurtay, H \"u seyin Demirci, H \"u seyin Tiryaki, and Tekin Altun. 2024. Emotion recognition on call center voice data. Applied Sciences, 14(20):9458

    2024

  39. [47]

    Chi Zhan, Dongyu She, Sicheng Zhao, Ming-Ming Cheng, and Jufeng Yang. 2019. Zero-shot emotion recognition via affective structural embedding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1151--1160

    2019

  40. [48]

    Guanhong Zhang, Sophia Ananiadou, and 1 others. 2022. Examining and mitigating gender bias in text emotion detection task. Neurocomputing, 493:422--434

    2022

  41. [49]

    Wenxuan Zhang, Yue Deng, Bing Liu, Sinno Pan, and Lidong Bing. 2024. Sentiment analysis in the era of large language models: A reality check. In Findings of the Association for Computational Linguistics: NAACL, pages 3881--3906

    2024

  42. [50]

    Weixiang Zhao, Yanyan Zhao, Xin Lu, Shilong Wang, Yanpeng Tong, and Bing Qin. 2023. Is chatgpt equipped with emotional dialogue capabilities? arXiv preprint arXiv:2304.09582

    2023 arXiv

  43. [51]

    Yue Zheng, Yuhao Chen, Bin Qian, Xiufang Shi, Yuanchao Shu, and Jiming Chen. 2025. A review on edge large language models: Design, execution, and applications. ACM Computing Surveys, 57(8):1--35

    2025

  44. [52]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  45. [53]

    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...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.