Quantifying the Affective Gap: A Zero-Shot Evaluation of LLMs on Fine-Grained Emotion Taxonomies

# A Zero-Shot Evaluation of LLMs on Fine-Grained Emotion Taxonomies
Source: [https://arxiv.org/html/2607.00968](https://arxiv.org/html/2607.00968)
## Quantifying the Affective Gap: A Zero\-Shot Evaluation of LLMs on Fine\-Grained Emotion Taxonomies

Lawrence Obiuwevwi, Krzysztof J\. Rechowicz, Jessica M\. Johnson, Vikas Ashok, Sachin Shetty, & Sampath Jayarathna

###### Abstract

Emotion recognition in natural language is a foundational challenge in affective computing, with critical implications for human\-computer interaction, mental health support, and conversational AI\. This paper presents a rigorous, unified zero\-shot evaluation of three leading commercial large language models: Claude \(claude\-sonnet\-4\-6\), ChatGPT \(GPT\-5\.4\), and Gemini \(gemini\-2\.5\-flash\)\. The models were queried through their respective production APIs as of April 2026 on a fine\-grained 13\-class emotion classification task\. Using a stratified 1,000\-sentence sample from theboltuix/emotions\-datasetcomprising 131,306 sentences across 13 categories, a single uniform prompt with no exemplars was applied identically across all models\. Gemini achieves the highest accuracy \(39\.9%\) and macro\-F1 \(0\.363\), followed by GPT\-5\.4 \(38\.8%, F1 = 0\.291\) and Claude \(38\.0%, F1 = 0\.159\)\. All models excel onsarcasmanddesirewhile consistently failing onlove,confusion, andshame\. McNemar’s tests reveal no statistically significant pairwise differences \(p\>0\.10p\>0\.10\), suggesting convergence at a shared zero\-shot ceiling\. Claude’s markedly lower macro\-F1 exposes a class\-imbalance prediction bias\. These findings highlight the current limitations of frontier AI systems in zero\-shot fine\-grained emotion classification\.

## IIntroduction

Emotions shape human communication and underpin high\-stakes applications including mental health monitoring\[[3](https://arxiv.org/html/2607.00968#bib.bib24)\], empathetic dialogue systems\[[21](https://arxiv.org/html/2607.00968#bib.bib25)\], and human\-computer interaction\. The ability of an AI system to recognize affective states is therefore a prerequisite for deployment in any domain where human welfare is at stake\.

Despite rapid advances in large language models \(LLMs\), their affective intelligence remains poorly characterized at fine\-grained resolution\. Most studies probe coarse polarity or the Ekman six basic emotions\[[8](https://arxiv.org/html/2607.00968#bib.bib1)\]rather than richer taxonomies\[[22](https://arxiv.org/html/2607.00968#bib.bib17),[10](https://arxiv.org/html/2607.00968#bib.bib16)\], and direct cross\-provider comparisons under identical experimental conditions are absent from the literature\. This gap matters: practitioners selecting an API for emotion\-aware applications must rely on ad hoc or proprietary benchmarks that may not reflect real\-world linguistic diversity\.

This paper addresses the gap through four research questions: zero\-shot accuracy across 13 classes \(RQ1\); pairwise statistical differences using McNemar’s test \(RQ2\); per\-class strengths and failure modes \(RQ3\); and the effect of sentence length on accuracy \(RQ4\)\. We contribute\(i\)the first direct zero\-shot comparison of Claude \(claude\-sonnet\-4\-6\), ChatGPT \(GPT\-5\.4\), and Gemini \(gemini\-2\.5\-flash\) on a 13\-class task using production APIs as of April 2026;\(ii\)per\-emotion accuracy breakdowns revealing systematic failure modes;\(iii\)McNemar significance testing; and\(iv\)a sentence\-length moderation analysis\.

## IIRelated Work

### II\-AEmotion Classification and Datasets

Computational emotion recognition has progressed from lexicon\-based resources such as the NRC Lexicon\[[17](https://arxiv.org/html/2607.00968#bib.bib11)\]and SemEval affective tasks\[[23](https://arxiv.org/html/2607.00968#bib.bib8),[16](https://arxiv.org/html/2607.00968#bib.bib9),[4](https://arxiv.org/html/2607.00968#bib.bib10)\]through deep learning classifiers to transformer fine\-tuning paradigms achieving state\-of\-the\-art results on benchmarks including GoEmotions\[[6](https://arxiv.org/html/2607.00968#bib.bib6)\]\(58K comments, 27 categories\)\. BERT\[[7](https://arxiv.org/html/2607.00968#bib.bib12),[20](https://arxiv.org/html/2607.00968#bib.bib2)\]and RoBERTa\[[12](https://arxiv.org/html/2607.00968#bib.bib13)\]established the transformer paradigm as dominant for supervised emotion classification\. Theboltuix/emotions\-datasetextends the Ekman taxonomy with social \(shame, guilt\)\[[24](https://arxiv.org/html/2607.00968#bib.bib5)\], cognitive \(confusion\), rhetorical \(sarcasm\)\[[9](https://arxiv.org/html/2607.00968#bib.bib21)\], and motivational \(desire\) categories, offering broader ecological validity\.

Emotion\-aware AI may also complement multimodal assistive systems that integrate cognitive, physiological, and attentional context in knowledge\-work environments\[[18](https://arxiv.org/html/2607.00968#bib.bib26)\]\. Related human\-computer interaction research has examined how eye\-tracking signals can characterize variations in cognitive effort during digital\-document reading\[[14](https://arxiv.org/html/2607.00968#bib.bib4),[13](https://arxiv.org/html/2607.00968#bib.bib3),[25](https://arxiv.org/html/2607.00968#bib.bib27)\]\.

### II\-BLLMs for Emotion and Sentiment

GPT\-3\[[2](https://arxiv.org/html/2607.00968#bib.bib14)\]demonstrated zero\-shot classification through in\-context learning, with instruction tuning\[[19](https://arxiv.org/html/2607.00968#bib.bib19),[26](https://arxiv.org/html/2607.00968#bib.bib20)\]substantially improving generalization\. SentimentGPT\[[10](https://arxiv.org/html/2607.00968#bib.bib16)\]found LLMs competitive on coarse polarity but weaker on low\-frequency emotion classes\. Yang et al\.\[[22](https://arxiv.org/html/2607.00968#bib.bib17)\]evaluated ChatGPT and GPT\-4 on sentiment tasks but excluded Claude, Gemini, and a fine\-grained 13\-class taxonomythe gap this study directly addresses\.

## IIIMethodology

### III\-ADataset

Theboltuix/emotions\-dataset\[[1](https://arxiv.org/html/2607.00968#bib.bib7)\]is a publicly available Hugging Face corpus of 131,306 English sentences annotated with one of 13 mutually exclusive emotion labels:happiness, sadness, fear, anger, love, disgust, surprise, neutral, confusion, desire, shame, guilt,andsarcasm\. A stratified random sample ofN=1,000N=1\{,\}000sentences was drawn with seed = 42, preserving class proportions to simulate realistic deployment conditions\. Table[I](https://arxiv.org/html/2607.00968#S3.T1)reports the distribution\.Happinessis the most frequent class \(20\.5%,n=205n=205\);desireis the rarest \(1\.9%,n=19n=19\)\. This 10:1 ratio has direct consequences for macro\-F1: models concentrating predictions on high\-frequency classes achieve deceptively high accuracy while failing on minority classes\.

TABLE I:Class Distribution in Stratified Sample \(N=1,000N=1\{,\}000\)
### III\-BModels and API Configuration

Three frontier LLMs were queried through their production APIs as of April 2, 2026: Claude \(claude\-sonnet\-4\-6\) through Anthropic’s Messages API with default sampling; ChatGPT \(GPT\-5\.4\) through OpenAI’s Chat Completions API withtemperature=0; and Gemini \(gemini\-2\.5\-flash\) through Google’s Generative AI API with default sampling\. All models receivedmax\_tokens=10, enforcing single\-word responses\. Responses not matching one of the 13 canonical labels after case normalization were treated as errors; all three models exhibited strong label adherence in practice\.

### III\-CPrompt Design

A single\-turn prompt was used identically across all models, with no system message and no few\-shot examples\[[2](https://arxiv.org/html/2607.00968#bib.bib14),[27](https://arxiv.org/html/2607.00968#bib.bib18)\]:

“Classify the emotion expressed in this sentence\. Reply with ONLY one word from this list: happiness, sadness, fear, anger, love, disgust, surprise, neutral, confusion, desire, shame, guilt, sarcasm\. Sentence:\{sentence\}\. Emotion:”

The full label list constrains the output space; imperative instruction style elicits consistent outputs from instruction\-tuned models\[[19](https://arxiv.org/html/2607.00968#bib.bib19)\]\. No chain\-of\-thought preamble is included, ensuring performance reflects raw affective inference\[[27](https://arxiv.org/html/2607.00968#bib.bib18)\]\.

### III\-DEvaluation Metrics

Four metrics were computed per model: \(i\)Overall Accuracy; \(ii\)Macro\-F1\(F​1macF1\_\{\\text\{mac\}\}\), the primary metric for minority\-class evaluation; \(iii\)Weighted\-F1\(F​1wtF1\_\{\\text\{wt\}\}\); and \(iv\)Cohen’s Kappa\(κ\\kappa\)\[[5](https://arxiv.org/html/2607.00968#bib.bib23)\]\. Pairwise comparisons usedMcNemar’s test\[[15](https://arxiv.org/html/2607.00968#bib.bib22)\]with continuity correction\.

## IVResults

### IV\-AOverall Performance

Table[II](https://arxiv.org/html/2607.00968#S4.T2)summarizes aggregate performance\. Gemini leads on all four metrics \(accuracy 39\.9%, macro\-F1 0\.363,κ=0\.337\\kappa=0\.337\), followed by GPT\-5\.4 \(38\.8%, 0\.291, 0\.327\) and Claude \(38\.0%, 0\.159, 0\.320\)\. All three models substantially exceed the majority\-class baseline \(20\.5%\) and random baseline \(7\.7%\), yet fall far below human\-level performance \(\>\>70%\)\[[6](https://arxiv.org/html/2607.00968#bib.bib6)\]\. The narrow 1\.9 pp spread suggests convergence near a shared zero\-shot ceiling, confirmed by the McNemar results below\.

TABLE II:Overall Performance \(N=1,000N=1\{,\}000\)
### IV\-BPer\-Emotion Accuracy

Table[III](https://arxiv.org/html/2607.00968#S4.T3)and Fig\.[2](https://arxiv.org/html/2607.00968#S4.F2)present per\-class accuracy\.Sarcasm\(96\.2%–100%\) anddesire\(89\.5%–94\.7%\) stand in stark contrast to all other categories\.Anger\(53\.3%–61\.7%\),fear\(44\.1%–48\.5%\), andsadness\(43\.1%–55\.4%\) form a mid\-performance cluster\.Love\(14\.5%–27\.6%\),confusion\(22\.9%–27\.1%\), andshame\(17\.6%–29\.4%\) are consistently the hardest categories\. Notable model\-specific divergences include Gemini collapsing onneutral\(14\.8% vs\. Claude’s 31\.1%\), Claude leading onguilt\(50\.0%\) andneutral, and GPT\-5\.4 leading onsadnessandlove\.

TABLE III:Per\-Emotion Accuracy \(%\) by Model![Refer to caption](https://arxiv.org/html/2607.00968v1/confusion_matrices.png)Figure 1:Normalized confusion matrices for Claude, GPT\-5\.4, and Gemini\. Off\-diagonal mass concentrates on semantically proximal pairs:love↔\\leftrightarrowhappiness,fear↔\\leftrightarrowsadness, andconfusion↔\\leftrightarrowneutral\.![Refer to caption](https://arxiv.org/html/2607.00968v1/emotion-accuracy.png)Figure 2:Per\-emotion accuracy heatmap \(%\)\. Sarcasm scores highest \(96\.2–100%\); love and neutral are consistently weakest\.Fig\.[1](https://arxiv.org/html/2607.00968#S4.F1)presents normalized confusion matrices\. Cross\-model patterns includelovemisclassified ashappiness, confirming semantic overlap between positive affect categories;confusionmisclassified asneutralandsadness; andfearfrequently assigned tosadness, suggesting that shared negative valence dominates arousal\-level distinctions\.

### IV\-CStatistical Significance

Table[IV](https://arxiv.org/html/2607.00968#S4.T4)presents McNemar’s testpp\-values\. No comparison reaches significance atα=0\.05\\alpha=0\.05, supporting convergence: accuracy differences are consistent with random variation, and no model can be declared definitively superior under this protocol\.

TABLE IV:McNemar’s Testpp\-Values \(α=0\.05\\alpha=0\.05\)
### IV\-DSentence Length Analysis

Performance peaks in the medium range \(5–15 words\): Claude = 41\.3%, GPT\-5\.4 = 42\.0%, and Gemini = 43\.6% \(n=567n=567\)\. Long sentences \(\>\>15 words\) produce the lowest accuracy \(33\.2%–34\.3%\), a∼\\sim9 pp drop consistent across all models \(Fig\.[3](https://arxiv.org/html/2607.00968#S4.F3)\), suggesting a structural property of zero\-shot affective inference\.

![Refer to caption](https://arxiv.org/html/2607.00968v1/length_vs_accuracy.png)Figure 3:Accuracy by sentence length\. Medium sentences \(5–15 words\) yield peak performance; long sentences \(\>\>15 words\) degrade accuracy by∼\\sim9 pp across all models\.

## VDiscussion

Zero\-shot ceiling:The 1\.9 pp accuracy spread and non\-significant McNemar results indicate all three models have converged near a shared zero\-shot ceiling\. For practitioners, provider choice is unlikely to affect raw accuracy\. However, the macro\-F1 gap \(0\.159 vs\. 0\.363, Claude vs\. Gemini\) is practically significant: in applications requiring minority\-class sensitivity \(e\.g\., shame, guilt\), Gemini is the preferred choice\.

Sarcasm/desire paradox and hard categories:Near\-perfect sarcasm \(96\.2%–100%\) and desire \(89\.5%–94\.7%\) accuracies likely reflect small support sizes \(n=26n=26,n=19n=19\) and atypical lexical prototypicality\[[11](https://arxiv.org/html/2607.00968#bib.bib15)\]rather than robust affective understanding\. Conversely,loveis the worst\-performing class for Claude and Gemini \(14\.5%\), driven by semantic overlap withhappiness\. The confusion matrix confirms that mostlovemisclassifications land onhappiness, consistent with GoEmotions\[[6](https://arxiv.org/html/2607.00968#bib.bib6)\], where positive social emotions formed the hardest discrimination cluster\.

Provider\-specific biases:Claude’s macro\-F1 of 0\.159less than half of Gemini’s 0\.363despite comparable accuracy \(38\.0%\) is the signature of majority\-class prediction bias\[[26](https://arxiv.org/html/2607.00968#bib.bib20)\]\. Claude over\-predictshappinesswhile underperforming on minority classes, likely amplified by alignment processes during pretraining\[[19](https://arxiv.org/html/2607.00968#bib.bib19)\]\. This is a critical limitation for mental health applications where rare emotions carry high clinical significance\. Gemini, by contrast, collapses onneutral\(14\.8% vs\. Claude’s 31\.1%\), systematically reassigning neutral sentences to emotional categories\. This pattern may reflect RLHF objectives rewarding emotionally engaging outputs\[[19](https://arxiv.org/html/2607.00968#bib.bib19)\], making Gemini unreliable for applications requiring neutral\-state detection\.

Implications and limitations:Accuracy in the 38%–40% range falls well short of deployment thresholds for emotionally sensitive contexts\[[3](https://arxiv.org/html/2607.00968#bib.bib24),[21](https://arxiv.org/html/2607.00968#bib.bib25)\]\. Key limitations include the zero\-shot\-only evaluation \(few\-shot prompting may substantially improve all models\), reliance on a single English\-language dataset, and high\-variance estimates for rare classes \(n<30n<30\)\. The results reflect API behavior as of April 2, 2026; continuous model updates may alter findings\.

## VIConclusion

We presented the first direct zero\-shot evaluation of Claude, GPT\-5\.4, and Gemini on 13\-class emotion classification using production APIs\. Gemini led on all metrics \(accuracy 39\.9%, macro\-F1 0\.363\), although no pairwise difference was statistically significant, suggesting convergence near a shared zero\-shot ceiling\. Our results reveal that frontier LLMs remain substantially below human\-level affective awareness in zero\-shot fine\-grained settings\. Future work should explore few\-shot and chain\-of\-thought prompting\[[27](https://arxiv.org/html/2607.00968#bib.bib18)\]for minority\-class improvement, fine\-tuned baselines to quantify the zero\-shot gap, and multilingual evaluation to probe cross\-lingual affective generalization\.

## Acknowledgement

This work is supported in part by NSF 245523\. Any opinions, findings, and conclusions or recommendations expressed in this material are of the author\(s\) and do not necessarily reflect those of the sponsors\.

## References

- \[1\]\(2023\)Emotions\-dataset\.Note:HuggingFace DatasetsExternal Links:[Link](https://huggingface.co/datasets/boltuix/emotions-dataset)Cited by:[§III\-A](https://arxiv.org/html/2607.00968#S3.SS1.p1.3)\.
- \[2\]T\. B\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell,et al\.\(2020\)Language models are few\-shot learners\.Advances in Neural Information Processing Systems \(NeurIPS\)33,pp\. 1877–1901\.External Links:[Document](https://dx.doi.org/10.5555/3495724.3495883)Cited by:[§II\-B](https://arxiv.org/html/2607.00968#S2.SS2.p1.1),[§III\-C](https://arxiv.org/html/2607.00968#S3.SS3.p1.1)\.
- \[3\]S\. Chancellor and M\. De Choudhury\(2020\)Methods in predictive techniques for mental health status on social media: a critical review\.npj Digital Medicine3\(43\)\.External Links:[Document](https://dx.doi.org/10.1038/s41746-020-0233-7)Cited by:[§I](https://arxiv.org/html/2607.00968#S1.p1.1),[§V](https://arxiv.org/html/2607.00968#S5.p4.1)\.
- \[4\]A\. Chatterjee, K\. N\. Narahari, M\. Joshi, and P\. Agrawal\(2019\)SemEval\-2019 task 3: EmoContext contextual emotion detection in text\.InProceedings of the 13th International Workshop on Semantic Evaluation \(SemEval\-2019\),pp\. 39–48\.External Links:[Document](https://dx.doi.org/10.18653/v1/S19-2005)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[5\]J\. Cohen\(1960\)A coefficient of agreement for nominal scales\.Educational and Psychological Measurement20\(1\),pp\. 37–46\.External Links:[Document](https://dx.doi.org/10.1177/001316446002000104)Cited by:[§III\-D](https://arxiv.org/html/2607.00968#S3.SS4.p1.3)\.
- \[6\]D\. Demszky, D\. Movshovitz\-Attias, J\. Ko, A\. Cowen, G\. Nemade, and S\. Ravi\(2020\)GoEmotions: a dataset of fine\-grained emotions\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics \(ACL\),pp\. 4040–4054\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.372)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1),[§IV\-A](https://arxiv.org/html/2607.00968#S4.SS1.p1.2),[§V](https://arxiv.org/html/2607.00968#S5.p2.2)\.
- \[7\]J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova\(2019\)BERT: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics \(NAACL\),pp\. 4171–4186\.External Links:[Document](https://dx.doi.org/10.18653/v1/N19-1423)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[8\]P\. Ekman\(1992\)An argument for basic emotions\.Cognition & Emotion6\(3–4\),pp\. 169–200\.External Links:[Document](https://dx.doi.org/10.1080/02699939208411068)Cited by:[§I](https://arxiv.org/html/2607.00968#S1.p2.1)\.
- \[9\]A\. Joshi, P\. Bhattacharyya, and M\. J\. Carman\(2017\)Automatic sarcasm detection: a survey\.ACM Computing Surveys50\(5\),pp\. 73:1–73:22\.External Links:[Document](https://dx.doi.org/10.1145/3124420)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[10\]K\. Kheiri and H\. Karimi\(2023\)SentimentGPT: exploiting GPT for advanced sentiment analysis and its departure from current machine learning\.arXiv preprint arXiv:2307\.10234\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2307.10234)Cited by:[§I](https://arxiv.org/html/2607.00968#S1.p2.1),[§II\-B](https://arxiv.org/html/2607.00968#S2.SS2.p1.1)\.
- \[11\]H\. Larochelle, D\. Erhan, and Y\. Bengio\(2008\)Zero\-data learning of new tasks\.InProceedings of the 23rd AAAI Conference on Artificial Intelligence \(AAAI\),pp\. 646–651\.External Links:[Link](https://aaai.org/papers/00646-aaai08-103-zero-data-learning-of-new-tasks/)Cited by:[§V](https://arxiv.org/html/2607.00968#S5.p2.2)\.
- \[12\]Y\. Liu, M\. Ott, N\. Goyal, J\. Du, M\. Joshi, D\. Chen, O\. Levy, M\. Lewis, L\. Zettlemoyer, and V\. Stoyanov\(2019\)RoBERTa: a robustly optimized BERT pretraining approach\.arXiv preprint arXiv:1907\.11692\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.1907.11692)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[13\]B\. Mahanama, G\. Jayawardena, Y\. Abeysinghe, V\. Ashok, and S\. Jayarathna\(2022\)Multidisciplinary reading patterns of digital documents\.In2022 Symposium on Eye Tracking Research and Applications,pp\. 1–2\.Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p2.1)\.
- \[14\]B\. Mahanama, G\. Jayawardena, and S\. Jayarathna\(2021\)Analyzing unconstrained reading patterns of digital documents using eye tracking\.In2021 ACM/IEEE Joint Conference on Digital Libraries \(JCDL\),pp\. 282–283\.Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p2.1)\.
- \[15\]Q\. McNemar\(1947\)Note on the sampling error of the difference between correlated proportions or percentages\.Psychometrika12\(2\),pp\. 153–157\.External Links:[Document](https://dx.doi.org/10.1007/BF02295996)Cited by:[§III\-D](https://arxiv.org/html/2607.00968#S3.SS4.p1.3)\.
- \[16\]S\. M\. Mohammad, F\. Bravo\-Marquez, M\. Salameh, and S\. Kiritchenko\(2018\)SemEval\-2018 task 1: affect in tweets\.InProceedings of the 12th International Workshop on Semantic Evaluation \(SemEval\-2018\),pp\. 1–17\.External Links:[Document](https://dx.doi.org/10.18653/v1/S18-1001)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[17\]S\. M\. Mohammad and P\. D\. Turney\(2013\)Crowdsourcing a word–emotion association lexicon\.Computational Intelligence29\(3\),pp\. 436–465\.External Links:[Document](https://dx.doi.org/10.1111/j.1467-8640.2012.00460.x)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[18\]L\. Obiuwevwi, K\. J\. Rechowicz, V\. Ashok, S\. Shetty, and S\. Jayarathna\(2026\)Cognitive prosthetic: an AI\-enabled multimodal system for episodic recall in knowledge work\.InProceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems,CHI EA ’26,pp\. 1–7\.External Links:[Document](https://dx.doi.org/10.1145/3772363.3798940)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p2.1)\.
- \[19\]L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. L\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray,et al\.\(2022\)Training language models to follow instructions with human feedback\.Advances in Neural Information Processing Systems \(NeurIPS\)35,pp\. 27730–27744\.External Links:[Document](https://dx.doi.org/10.5555/3600270.3602281)Cited by:[§II\-B](https://arxiv.org/html/2607.00968#S2.SS2.p1.1),[§III\-C](https://arxiv.org/html/2607.00968#S3.SS3.p3.1),[§V](https://arxiv.org/html/2607.00968#S5.p3.1)\.
- \[20\]K\. Pineda, A\. M\. Perrotti, F\. Poursardar, D\. Graber, and S\. Jayarathna\(2023\)Using bert to understand tiktok users’ adhd discussion\.In2023 IEEE 24th International Conference on Information Reuse and Integration for Data Science \(IRI\),pp\. 213–214\.Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[21\]H\. Rashkin, E\. M\. Smith, M\. Li, and Y\. Boureau\(2019\)Towards empathetic open\-domain conversation models: a new benchmark and dataset\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics \(ACL\),pp\. 5370–5381\.External Links:[Document](https://dx.doi.org/10.18653/v1/P19-1534)Cited by:[§I](https://arxiv.org/html/2607.00968#S1.p1.1),[§V](https://arxiv.org/html/2607.00968#S5.p4.1)\.
- \[22\]S\. Sabour, S\. Liu, Z\. Zhang, J\. Liu, J\. Zhou, A\. Sunaryo, T\. Lee, R\. Mihalcea, and M\. Huang\(2024\)EmoBench: evaluating the emotional intelligence of large language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 5986–6004\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.326)Cited by:[§I](https://arxiv.org/html/2607.00968#S1.p2.1),[§II\-B](https://arxiv.org/html/2607.00968#S2.SS2.p1.1)\.
- \[23\]C\. Strapparava and R\. Mihalcea\(2007\)SemEval\-2007 task 14: affective text\.InProceedings of the 4th International Workshop on Semantic Evaluations \(SemEval\-2007\),pp\. 70–74\.External Links:[Link](https://aclanthology.org/S07-1013)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[24\]J\. P\. Tangney and R\. L\. Dearing\(2002\)Shame and guilt\.Guilford Press,New York, NY\.External Links:ISBN 9781593859480Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p1.1)\.
- \[25\]K\. Thennakoon, Y\. Abeysinghe, P\. Thenahandi, L\. Obiuwevwi, V\. Ashok, and S\. Jayarathna\(2025\)How devices shape mental effort in digital document reading: an eye\-tracking study\.In2025 ACM/IEEE Joint Conference on Digital Libraries \(JCDL\),pp\. 237–240\.External Links:[Document](https://dx.doi.org/10.1109/JCDL67857.2025.00035)Cited by:[§II\-A](https://arxiv.org/html/2607.00968#S2.SS1.p2.1)\.
- \[26\]J\. Wei, M\. Bosma, V\. Y\. Zhao, K\. Guu, A\. W\. Yu, B\. Lester, N\. Du, A\. M\. Dai, and Q\. V\. Le\(2022\)Finetuned language models are zero\-shot learners\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Document](https://dx.doi.org/10.48550/arXiv.2109.01652)Cited by:[§II\-B](https://arxiv.org/html/2607.00968#S2.SS2.p1.1),[§V](https://arxiv.org/html/2607.00968#S5.p3.1)\.
- \[27\]J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. Chi, Q\. Le, and D\. Zhou\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.Advances in Neural Information Processing Systems \(NeurIPS\)35,pp\. 24824–24837\.External Links:[Document](https://dx.doi.org/10.5555/3600270.3602070)Cited by:[§III\-C](https://arxiv.org/html/2607.00968#S3.SS3.p1.1),[§III\-C](https://arxiv.org/html/2607.00968#S3.SS3.p3.1),[§VI](https://arxiv.org/html/2607.00968#S6.p1.1)\.