Pith. sign in

REVIEW 4 major objections 6 minor 60 references

EmoPerso: Enhancing Personality Detection with Self-Supervised Emotion-Aware Modelling

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read EmoPerso's self-supervised emotion-aware modelling lifts personality detection past 80% average Macro-F1 on Kaggle and to 68.60% on Pandora, surpassing prior state-of-the-art models.

desk verdict Strong SOTA claim on MBTI detection, but the emotion pseudo-label recipe is underspecified; ask for the code and a random-label control before believing the 8-point ablation gain. read the letter →

arxiv 2509.02450 v1 pith:BHLHECCE submitted 2025-09-02 cs.CL cs.LG

Lingzhi Shen , Xiaohao Cai , Yunfei Long , Imran Razzak , Guanming Chen , Shoaib Jameel This is my paper
classification cs.CLcs.LG
keywords personalitydetectionemotion-awaremodellingself-supervisedlearningmulti-taskMBTIcross-attentionreasoningchainstextaugmentation
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

EmoPerso tries to establish that personality detection from social-media text gets substantially better when the model is forced to model emotion at the same time—even though no emotion labels are available. It derives emotion pseudo-labels from linguistic cues such as valence-bearing adjectives, intensifiers, and punctuation, then trains an emotion head jointly with the four binary personality-dimension heads, lets the personality representation re-attend to tokens under emotion-conditioned weights, and augments the data with LLM-generated paraphrases and completed masked spans. On the Kaggle and Pandora benchmarks the full system reports average Macro-F1 of 81.07% and 68.60%, respectively, the first reported pass above 80% on Kaggle and a 9.00-point lead over the strongest baseline. If the result holds, it suggests that inexpensive, automatically derived emotion signals can substitute for costly emotion annotation and that emotion–personality interaction is a genuinely load-bearing modelling choice, not a side decoration.

What carries the argument

The load-bearing mechanism is a pseudo-labelled emotion stream wired into a multi-task learner. The text is encoded by a frozen LLM; an attention-pooled shared representation feeds two lightweight heads, one predicting the four MBTI dimensions and one predicting emotion categories from pseudo-labels derived from lexical and stylistic cues such as valence adjectives, intensifiers, exclamation marks, and emotive punctuation. A multi-head cross-attention module uses the personality representation as a query over the token sequence, and an emotion-conditioned weighting layer re-weights the attended tokens before a cosine consistency loss aligns the emotion and personality spaces. Around this cor

What would settle it

Re-run EmoPerso with the emotion pseudo-labels shuffled within each training batch while keeping their marginal distribution fixed. If average Macro-F1 stays close to 81.07/68.60 instead of falling toward the w/o-emotions ablation level, then the specific emotion cues are not carrying the claimed signal. Separately, measure the agreement between the heuristic pseudo-labels and human emotion annotations on a held-out corpus; low agreement would indicate the emotion head is learning style cues, not emotions.

Watch

Extended reading notes

Core claim

The paper's central claim is that emotion is not a side signal for personality detection but a necessary auxiliary task: the model should predict emotional categories from a post at the same time as it predicts the author's personality, then use the inferred emotion representation to decide which tokens matter for personality. The emotion labels are never externally annotated; they are pseudo-labels built from affective heuristics present in the text, so the whole pipeline stays self-supervised with respect to emotion. On the Kaggle benchmark EmoPerso reports 81.07% average Macro-F1 across the four MBTI dimensions, exceeding the strongest prior system by 9.00 points and marking the first rep

Load-bearing premise

The result depends on the hand-crafted emotion pseudo-labels (valence adjectives, intensifiers, punctuation cues) being valid, stable stand-ins for real emotion categories; if those heuristics are arbitrary or dataset-specific, the emotion stream may be fitting surface patterns that will not generalise to new text.

Editorial extensions

If this is right

  • If the reported numbers replicate, EmoPerso sets a new operating point for MBTI detection from social media text: 81.07% average Macro-F1 on Kaggle and 68.60% on Pandora, with the I/E and S/N dimensions crossing roughly 80% for the first time.
  • Emotion features can be obtained without emotion annotation: the pseudo-label heuristic makes the emotion stream trainable on the same personality-labelled data, which matters because high-quality personality labels are scarce and emotion labels are scarcer.
  • Every major component contributes: removing the emotion stream, the generative augmentation, the multi-task objective, the cross-attention, or the reasoning chains each lowers Macro-F1, and removing emotion features costs the most.
  • The framework is backbone-agnostic in the paper's experiments: swapping the frozen LLM for a different large model yields comparable average performance, suggesting that the gains come from the emotion-aware training structure rather than from one particular pretrained model.

Reading between the lines

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

  • An untested implication is that the same emotion-conditioned multi-task recipe could transfer to other trait-inference tasks with sparse labels, such as Big Five, stance, or deception detection, where pseudo-labelled psycholinguistic cues are easy to derive.
  • Because the pseudo-labels are built from surface cues, a reasonable worry is that part of the gain is stylistic rather than emotional: the emotion head may be capturing register differences that correlate with MBTI self-reports. A direct test would replace the heuristic labels with labels from a validated emotion lexicon or with human emotion annotations and compare.
  • The reported emotion-contribution heatmap suggests specific emotion categories matter differently for different MBTI dimensions; if that pattern is stable, it could be used to generate personality-specific emotion lexicons or to design targeted augmentation.
  • The KL-regularized LLM augmentation step invites further ablation: varying the number of paraphrases, the mask rate, and the style conditions could reveal whether the current gains saturate or whether more diversity translates into further Macro-F1 improvements.
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 / 6 minor

Summary. The paper proposes EmoPerso, a self-supervised framework for MBTI personality detection from social media text. The method combines LLM-based data augmentation (style-conditioned paraphrasing and contextual completion), multi-task learning with pseudo-labeled emotion prediction, a cross-attention personality–emotion interaction module, and a STaR-style reasoning-chain selection mechanism. Using a frozen DeepSeek-V3 backbone, the authors report average Macro-F1 of 81.07% on Kaggle and 68.60% on Pandora, surpassing prior state-of-the-art by roughly 9 and 5.5 points, respectively. An extensive ablation study (Table 3) attributes the largest contribution to the emotion branch, with an 8.04-point drop on Kaggle and 5.39-point drop on Pandora when emotion features are removed. A source-code URL is provided.

Significance. If the results are reproducible and the mechanisms are as described, EmoPerso would be a practically valuable advance in personality detection, establishing a new performance level on two widely used benchmarks. The paper is also methodologically ambitious, integrating several ideas (self-supervised emotion pseudo-labeling, MTL, cross-attention, STaR) in one framework, and it includes a thorough ablation study with t-SNE and attention visualizations. The main caveats are that the central independent variable—the emotion pseudo-label construction—is not specified in the manuscript, and the headline SOTA comparison is confounded by the much larger LLM backbone relative to the published baselines. Both issues are addressable, but they currently prevent full confidence in the causal claim that emotion-aware self-supervision drives the gains.

major comments (4)
  1. [§3.2, Eq. (3); §4.1; Table 3] The emotion pseudo-label generator is the independent variable whose removal causes the largest ablation drop (w/o Emotions: −8.04 on Kaggle, −5.39 on Pandora), yet it is never specified. §3.2 says labels are 'selected from the inferred label set Ŷ_emo automatically constructed based on emotional cues,' and §4.1 only repeats 'affective heuristics (e.g., adjectives, intensifiers, and punctuation).' The number of emotion categories C_e, the label set, the exact cue lexicon, thresholds, and the derivation algorithm are all missing. Consequently, the emotion head's supervision signal cannot be inspected or reproduced from the text. Please provide the full construction procedure, and add a control that trains with randomly shuffled or label-free auxiliary targets (e.g., same multi-task setup with non-semantic pseudo-labels) to show that the gains are due to emotion semantics and not merely to
  2. [§4.2, Table 2] The claim of 'surpassing state-of-the-art' is confounded by the backbone. All listed baselines use much smaller architectures (e.g., RCNN, GCN, BERT-class encoders), while EmoPerso uses the large frozen DeepSeek-V3. The internal control 'Vanilla DeepSeek-V3' (66.76 on Kaggle) shows the backbone alone is strong, but the external comparison in Table 2 does not include any LLM-backbone baseline of comparable scale. Since the paper itself reports 'Replace DeepSeek-V3 with GPT-4o' in Table 3, such an LLM-based comparison is feasible. Please add a same-backbone or same-scale baseline (e.g., TAE or PsyCoT reimplemented on DeepSeek-V3, or at least a direct LLM-prompting baseline) to separate the method's contribution from the backbone's capacity. This does not invalidate the method but is necessary to support the 'first to pass 80%' wording.
  3. [§4.1; Tables 2 and 3] All results are said to be averaged over ten runs, but Tables 2 and 3 report only point estimates. The central empirical claims rest on differences between configurations (e.g., 8.04, 5.39, 4.97 points), and smaller differences such as w/o KL Divergence vs. EmoPerso (80.24 vs. 81.07 on Kaggle) could easily be within run-to-run variance. Please report standard deviations or confidence intervals, and where feasible paired significance tests (e.g., over the same ten seeds) for the main SOTA comparison and the key ablations. Without this, the relative ranking of close configurations is not statistically grounded.
  4. [§3.4, Eqs. (10)–(11)] The reasoning-chain selection uses the model's own predicted probabilities and its own predicted emotion features (IG and MI are computed from the model's outputs), creating a self-referential feedback loop. The STaR component is credited with a 5.65-point gain on Kaggle (75.42 vs. 81.07 in Table 3), but it is not clear that the selection criterion is selecting genuinely informative chains rather than chains that the model already finds likely. Please add a control that selects random chains, or chains scored by a held-out reward function, to demonstrate that the IG/MI selection mechanism itself—not merely the presence of extra generated text—is responsible for the improvement.
minor comments (6)
  1. [§3.2 / §3.3] Notation clash: W_emo is used both for the emotion classification head in §3.2 and for the emotion-conditioned attention projection in §3.3. Please use distinct symbols.
  2. [§3.2 / §4.1 / Eq. (7), Eq. (12)] The loss decomposition is inconsistent. §3.2 defines λ_pers and λ_emo inside L_MTL, while Algorithm 1 and Eq. (7)/(12) use λ_MTL, λ_cross, λ_star. §4.1 says the MTL ratio is 0.7:0.3 but does not say whether this corresponds to λ_pers:λ_emo or to λ_MTL as a whole. Please unify notation and clarify the hyperparameter dependencies.
  3. [§4.1] 'The input sequence is standardized to 2,048 tokens (median) with a hidden size of 4,096'—it is unclear whether this hidden size refers to DeepSeek-V3's hidden dimension or a task-specific layer. Please specify the backbone configuration and the pooling/truncation strategy for shorter posts.
  4. [§4.2] 'This marks the first time that a personality detection model has surpassed 80% on this dataset' is a strong claim that depends on the comparison set in Table 2. Please either scope the claim to the methods considered or survey prior reported numbers carefully, since LLM-based methods may not all be included.
  5. [Figure 4] The emotion-importance heatmap is described in prose but the exact scoring formula is not given. Please state the mathematical definition of the importance scores used in the figure, or point to a formula in the text.
  6. [§6] Minor wording: 'personality detection and classification based on emotions and spoken language' seems to refer to text-only input; presumably 'spoken' should be 'written' or 'natural language.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the emotion pseudo-labels are a hand-crafted input transform, not a renaming of the personality labels, and the final supervision remains the ground-truth MBTI labels.

full rationale

The paper's derivation chain is not circular. The emotion pseudo-labels are described as derived from stylistic and affective cues in the input (e.g., adjectives, intensifiers, punctuation), not from the target personality labels (Section 3.2, Section 4.1). The personality head is trained with binary cross-entropy against ground-truth MBTI labels (Eq. 2), and the emotion head is an auxiliary multi-label task trained on pseudo-labels (Eq. 3). Even though L_cross aligns the personality representation with the emotion embedding, the final personality prediction is still supervised by the external MBTI labels, so the improvement is not forced by construction. The STaR reasoning-chain selection uses the model's own predicted probabilities and emotion features (Eqs. 9-11), but this is a standard self-training bootstrap; the final objective (Eq. 12) still contains the supervised personality loss, so the chain selection does not define the target. The paper's self-citations (e.g., [32], [33]) appear only in related-work context and are not used as the load-bearing justification for any component. The underspecification of the exact pseudo-label construction is a reproducibility and validity concern (a random-label control would strengthen the causal claim), but it is not a circularity: the pseudo-labels are not derived from the personality labels, and no equation in the paper reduces the predicted personality output to the emotion pseudo-labels or to a fitted parameter. Therefore, the central claim is not circular by definition or by self-citation chain.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The framework relies on a small set of hand-chosen loss weights, an unspecified pseudo-label emotion set, and several domain assumptions about LLM augmentation and MBTI validity. The most fragile item is the emotion pseudo-label construction, which is central to the claimed gains.

free parameters (8)
  • λ_pers:λ_emo ratio = 0.7:0.3
    Loss weight ratio between personality and emotion tasks in MTL; chosen by hand (Section 4.1).
  • λ_KL = 0.1
    Weight of KL divergence regularization in augmentation loss (Section 4.1).
  • λ_style = not reported
    Weight for style classification loss in L_gen; value not specified (Section 3.1).
  • λ_cross = not reported
    Weight for cross-attention consistency loss in Eq. (7); value not specified (Section 3.3).
  • λ_star = not reported
    Weight for reasoning chain entropy loss in Eq. (12); value not specified (Section 3.4).
  • λ_IG and λ_MI = not reported
    Weights for information gain and mutual information in chain selection (Eq. 10); values not specified (Section 3.4).
  • k (number of paraphrases) = 3
    Number of style-conditioned paraphrases generated per input (Section 4.1).
  • mask ratio = 10%
    Fraction of tokens masked for contextual completion (Section 4.1).
assumptions (5)
  • domain assumption MBTI self-reported personality labels are valid ground truth for personality detection
    Used throughout; both datasets are labeled by users' self-reported MBTI types (Section 4.1).
  • domain assumption LLM-generated paraphrases preserve the personality-relevant content of the original post
    The augmentation strategy in Section 3.1 relies on semantic fidelity of paraphrases to maintain personality signals.
  • domain assumption Affective heuristics (valence-bearing adjectives, intensifiers, punctuation) are valid proxies for basic emotions
    Used to construct pseudo-labels for the emotion head in Section 3.2.
  • ad hoc to paper The inferred emotion label set Y_emo is fixed and appropriate
    The label set is never specified; the method assumes such a set exists and is suitable for the auxiliary task (Section 3.2).
  • domain assumption DeepSeek-V3 provides sufficiently rich frozen representations for personality and emotion
    The backbone is frozen and all downstream components depend on its internal features (Section 4.1).
invented entities (1)
  • pseudo-labeled emotion category set (Ŷemo)
    purpose: Provides self-supervised emotion supervision for the MTL head and emotion-conditioned modulation
    The set of emotion categories and the heuristics used to build it are never described, so there is no way to test its validity outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EmoPerso: Enhancing Personality Detection with Self-Supervised Emotion-Aware Modelling." pith.science (2026). https://pith.science/paper/BHLHECCE

@misc{pith2026250902450,
  author       = {Pith},
  title        = {Pith review of: EmoPerso: Enhancing Personality Detection with Self-Supervised Emotion-Aware Modelling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BHLHECCE}},
  note         = {Machine review of arXiv:2509.02450}
}
read the original abstract

Personality detection from text is commonly performed by analysing users' social media posts. However, existing methods heavily rely on large-scale annotated datasets, making it challenging to obtain high-quality personality labels. Moreover, most studies treat emotion and personality as independent variables, overlooking their interactions. In this paper, we propose a novel self-supervised framework, EmoPerso, which improves personality detection through emotion-aware modelling. EmoPerso first leverages generative mechanisms for synthetic data augmentation and rich representation learning. It then extracts pseudo-labeled emotion features and jointly optimizes them with personality prediction via multi-task learning. A cross-attention module is employed to capture fine-grained interactions between personality traits and the inferred emotional representations. To further refine relational reasoning, EmoPerso adopts a self-taught strategy to enhance the model's reasoning capabilities iteratively. Extensive experiments on two benchmark datasets demonstrate that EmoPerso surpasses state-of-the-art models. The source code is available at https://github.com/slz0925/EmoPerso.

Figures

Figures reproduced from arXiv: 2509.02450 by the authors.

Figure 1
Figure 1. An illustration of the cognitive distinction and in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of EmoPerso. The framework leverages LLMs for self-supervised emotion feature extraction. It [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Code-style illustration of style-conditioned para [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Emotion contribution to prediction on the Kaggle [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Comparison of the best training epoch (scaled) and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 47 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Jaan Aru, Aqeel Labash, Oriol Corcoll, and Raul Vicente. 2023. Mind the gap: Challenges of deep learning approaches to theory of mind. Artificial Intelligence Review 56, 9 (2023), 9141–9156

    work page 2023

  3. [3]

    Goran Bubaš. 2024. The use of GPT-4o and Other Large Language Models for the Improvement and Design of Self-Assessment Scales for Measurement of Interpersonal Communication Skills. arXiv preprint arXiv:2409.14050 (2024)

    work page Pith review arXiv 2024

  4. [4]

    Sabur Butt, Grigori Sidorov, and Alexander Gelbukh. 2025. Interpretation of Myers–Briggs Type Indicator personality profiles based on ambivert continuum scale. Expert Systems with Applications 264 (2025), 125689

    work page 2025

  5. [5]

    Giovanni Luca Cascio Rizzo, Jonah Berger, Matteo De Angelis, and Rumen Pozhar- liev. 2023. How sensory language shapes influencer’s impact.Journal of Consumer Research 50, 4 (2023), 810–825

    work page 2023

  6. [6]

    Quan Cheng and Wenwan Shi. 2025. Hierarchical multi-label text classification of tourism resources using a label-aware dual graph attention network. Information Processing & Management 62, 1 (2025), 103952

    work page 2025

  7. [7]

    Dan Cogan, Zu-En Su, Oded Kenneth, and David Gershoni. 2023. Deterministic generation of indistinguishable photons in a cluster state. Nature Photonics 17, 4 (2023), 324–329

    work page 2023

  8. [8]

    Jiequan Cui, Beier Zhu, Qingshan Xu, Zhuotao Tian, Xiaojuan Qi, Bei Yu, Han- wang Zhang, and Richang Hong. 2025. Generalized Kullback-Leibler Divergence Loss. arXiv preprint arXiv:2503.08038 (2025)

    arXiv 2025

Show all 60 references
  1. [9]

    P. Ekman. 1992. Are there basic emotions? Psychological Review 99, 3 (1992), 550–553

    1992

  2. [10]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  3. [11]

    Jorge Luis Guerra, Carlos Catania, and Eduardo Veas. 2022. Datasets are not enough: Challenges in labeling network traffic. Computers & Security 120 (2022), 102810

    2022

  4. [12]

    Linmei Hu, Hongyu He, Duokang Wang, Ziwang Zhao, Yingxia Shao, and Liqiang Nie. 2024. LLM vs Small Model? Large Language Model Based Text Augmentation Enhanced Personality Detection Model. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 18234–18242

    2024

  5. [13]

    Shadi Jaradat, Richi Nayak, Alexander Paz, Huthaifa I Ashqar, and Mohammad El- henawy. 2024. Multitask learning for crash analysis: A fine-tuned llm framework using twitter data. Smart Cities 7, 5 (2024), 2422–2465

    2024

  6. [14]

    Jamin Rahman Jim, Md Apon Riaz Talukder, Partha Malakar, Md Mohsin Kabir, Kamruddin Nur, and Mohammed Firoz Mridha. 2024. Recent advancements and challenges of NLP-based sentiment analysis: A state-of-the-art review. Natural Language Processing Journal (2024), 100059

    2024

  7. [15]

    Francois Leonardi, Patrick Feldman, Matthew Almeida, William Moretti, and Charles Iverson. 2024. Contextual feature drift in large language models: An examination of adaptive retention across sequential inputs. (2024)

    2024

  8. [16]

    Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hong- sheng Li, and Yu Qiao. 2023. Uniformer: Unifying convolution and self-attention for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelli- gence 45, 10 (2023), 12581–12600

    2023

  9. [17]

    Yuming Li, Johnny Chan, Gabrielle Peko, and David Sundaram. 2024. An explana- tion framework and method for AI-based text emotion analysis and visualisation. Decision Support Systems 178 (2024), 114121

    2024

  10. [18]

    Yang Li, Amirmohammad Kazemeini, Yash Mehta, and Erik Cambria. 2022. Mul- titask learning for emotion and personality traits detection. Neurocomputing 493 (2022), 340–350

    2022

  11. [19]

    Zheng Li, Dawei Zhu, Qilong Ma, Weimin Xiong, and Sujian Li. 2024. EERPD: Leveraging Emotion and Emotion Regulation for Improving Personality Detec- tion. arXiv preprint arXiv:2406.16079 (2024)

    2024 arXiv

  12. [20]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  13. [21]

    Wenjie Liu, Zhijie Ren, and Liang Chen. 2025. Knowledge reasoning based on graph neural networks with multi-layer top-p message passing and sparse negative sampling. Knowledge-Based Systems (2025), 113063

    2025

  14. [22]

    Wenjuan Liu, Zhengyan Sun, Subo Wei, Shunxiang Zhang, Guangli Zhu, and Lei Chen. 2024. PS-GCN: Psycholinguistic graph and sentiment semantic fused graph convolutional networks for personality detection. Connection Science 36, 1 (2024), 2295820

    2024

  15. [23]

    Qing Luo, Wei Zeng, Manni Chen, Gang Peng, Xiaofeng Yuan, and Qiang Yin

  16. [24]

    Veronica Lynn, Niranjan Balasubramanian, and H Andrew Schwartz. 2020. Hi- erarchical modeling for user personality prediction: The role of message-level attention. In Proceedings of the 58th annual meeting of the association for compu- tational linguistics. 5306–5316

    2020

  17. [25]

    Walter Mischel and Yuichi Shoda. 1995. A cognitive-affective system theory of personality: reconceptualizing situations, dispositions, dynamics, and invariance in personality structure. Psychological review 102, 2 (1995), 246

    1995

  18. [26]

    Max Murphy. 2024. Artificial Intelligence and Personality: Large Lan- guage Models’ Ability to Predict Personality Type. Emerging Media (2024), 27523543241257291

    2024

  19. [27]

    Benjamin Nelson, Ari Winbush, Steven Siddals, John Torous, Nick Allen, and Matthew Flathers. 2025. Evaluating the Performance of Large Language Models in Identifying Human Facial Emotions: GPT 4o, Gemini 2.0 Experimental, and Claude 3.5 Sonnet. (2025)

    2025

  20. [28]

    Ruthie Pliskin, Anat Ruhrman, and Eran Halperin. 2020. Proposing a multi- dimensional, context-sensitive approach to the study of ideological (a) symmetry in emotion. Current Opinion in Behavioral Sciences 34 (2020), 75–80

    2020

  21. [29]

    J Prasanthi and G Anuradha. 2021. SURVEY ON PERSONALITY DETECTION USING DEEP LEARNING TECHNIQUES. In 2021 6th International Conference on Communication and Electronics Systems (ICCES) . IEEE, 1–8

    2021

  22. [30]

    Mustafa Safdari, Greg Serapio-García, Clément Crepy, Stephen Fitz, Peter Romero, Luning Sun, Marwa Abdulhai, Aleksandra Faust, and Maja Matarić. 2023. Per- sonality traits in large language models. arXiv preprint arXiv:2307.00184 (2023)

    2023 arXiv

  23. [31]

    Aditya G Shanmukha, RS Shamyuktha, S Karan, Deepa Gupta, and Suja Palaniswamy. 2024. Advancing Personality Detection through Word Embed- ments and Deep Learning: An Examination Using the MBTI Dataset. In2024 IEEE Recent Advances in Intelligent Computational Systems (RAICS) . IEEE, 1–6

    2024

  24. [32]

    Lingzhi Shen, Yunfei Long, Xiaohao Cai, Guanming Chen, Imran Razzak, and Shoaib Jameel. 2025. Less but Better: Parameter-Efficient Fine-Tuning of Large Language Models for Personality Detection. arXiv preprint arXiv:2504.05411 (2025)

    2025 arXiv

  25. [33]

    Lingzhi Shen, Yunfei Long, Xiaohao Cai, Guanming Chen, Yuhan Wang, Imran Razzak, and Shoaib Jameel. 2025. Ll4g: Self-supervised dynamic optimization for graph-based personality detection. arXiv preprint arXiv:2504.02146 (2025)

    2025 arXiv

  26. [34]

    Lingzhi Shen, Yunfei Long, Xiaohao Cai, Imran Razzak, Guanming Chen, Kang Liu, and Shoaib Jameel. 2025. Gamed: Knowledge adaptive multi-experts decou- pling for multimodal fake news detection. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data...

    2025

  27. [35]

    Zhouhao Sun, Xiao Ding, Li Du, Yunpeng Xu, Yixuan Ma, Yang Zhao, Bing Qin, and Ting Liu. 2025. Information Gain-Guided Causal Intervention for Autonomous Debiasing Large Language Models. arXiv preprint arXiv:2504.12898 (2025)

    2025 arXiv

  28. [36]

    Hossein Dabiriyan Tehrani, Sara Yamini, and Alexander T Vazsonyi. 2024. Par- enting styles and Big Five personality traits among adolescents: A meta-analysis. Personality and Individual Differences 216 (2024), 112421

    2024

  29. [37]

    Teng Teng, Huifang Li, Yulin Fang, and Lingzhi Shen. 2022. Understanding the differential effectiveness of marketer versus user-generated advertisements in closed social networking sites: An empirical study of WeChat. Internet Research 32, 6 (2022), 1910–1929

    2022

  30. [38]

    Iris Van Rooij, Olivia Guest, Federico Adolfi, Ronald de Haan, Antonina Kolokolova, and Patricia Rich. 2024. Reclaiming AI as a theoretical tool for cognitive science. Computational Brain & Behavior 7, 4 (2024), 616–636

    2024

  31. [39]

    Di Wang, Ronghao Yang, Hanhu Liu, Haiqing He, Junxiang Tan, Shaoda Li, Yichun Qiao, Kangqi Tang, and Xiao Wang. 2022. HFENet: hierarchical feature extraction network for accurate landcover classification. Remote Sensing 14, 17 (2022), 4244

    2022

  32. [40]

    Hongyu Wang, Dandan Zhang, Jun Feng, Lucia Cascone, Michele Nappi, and Shaohua Wan. 2024. A multi-objective segmentation method for chest X-rays based on collaborative learning from multiple partially annotated datasets. Infor- mation Fusion 102 (2024), 102016

    2024

  33. [41]

    Y. Wang, D. Li, K. Funakoshi, and M. Okumura. 2023. Emp: Emotion-guided multi-modal fusion and contrastive learning for personality traits recognition. In Proceedings of the 2023 ACM International Conference on Multimedia Retrieval . 243–252. EmoPerso: Enhancing Personality De...

    2023

  34. [42]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  35. [43]

    P William, N Yogeesh, Vishal M Tidake, Snehal Sumit Gondkar, K Vengatesan, et al. 2023. Framework for implementation of personality inventory model on natural language processing with personality traits analysis. In2023 International Conference on Intelligent Data Communicatio...

    2023

  36. [44]

    Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Lidia Sam Chao, and Derek Fai Wong. 2025. A survey on LLM-generated text detection: Necessity, methods, and future directions. Computational Linguistics (2025), 1–66

    2025

  37. [45]

    Xuansheng Wu, Jiayi Yuan, Wenlin Yao, Xiaoming Zhai, and Ninghao Liu. 2025. Interpreting and steering llms with mutual information-based explanations on sparse autoencoders. arXiv preprint arXiv:2502.15576 (2025)

    2025 arXiv

  38. [46]

    Di Xue, Lifa Wu, Zheng Hong, Shize Guo, Liang Gao, Zhiyong Wu, Xiaofeng Zhong, and Jianshan Sun. 2018. Deep learning-based personality recognition from text posts of online social networks. Applied Intelligence 48, 11 (2018), 4232–4246

    2018

  39. [47]

    Feifan Yang, Xiaojun Quan, Yunyi Yang, and Jianxing Yu. 2021. Multi-document transformer for personality detection. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 14221–14229

    2021

  40. [48]

    Feifan Yang, Tao Yang, Xiaojun Quan, and Qinliang Su. 2021. Learning to an- swer psychological questionnaire for personality detection. In Findings of the Association for Computational Linguistics: EMNLP 2021 . 1131–1142

    2021

  41. [49]

    Tao Yang, Jinghao Deng, Xiaojun Quan, and Qifan Wang. 2023. Orders are unwanted: dynamic deep graph convolutional network for personality detection. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 13896– 13904

    2023

  42. [50]

    Tao Yang, Tianyuan Shi, Fanqi Wan, Xiaojun Quan, Qifan Wang, Bingzhe Wu, and Jiaxiang Wu. 2023. PsyCoT: psychological questionnaire as powerful chain- of-thought for personality detection. arXiv preprint arXiv:2310.20256 (2023)

    2023 arXiv

  43. [51]

    Tao Yang, Feifan Yang, Haolan Ouyang, and Xiaojun Quan. 2021. Psycholinguistic tripartite graph network for personality detection.arXiv preprint arXiv:2106.04963 (2021)

    2021 arXiv

  44. [52]

    Hee Jun Yoon, Brent W Roberts, Madison N Sewell, Christopher M Napolitano, Christopher J Soto, Dana Murano, and Alex Casillas. 2024. Examining SEB skills’ incremental validity over personality traits in predicting academic achievement. Plos one 19, 1 (2024), e0296484

    2024

  45. [53]

    Eric Zelikman, YH Wu, Jesse Mu, and Noah D Goodman. 2024. STaR: Self-taught reasoner bootstrapping reasoning with reasoning. In Proc. the 36th International Conference on Neural Information Processing Systems , Vol. 1126

    2024

  46. [54]

    Haohao Zhu, Xiaokun Zhang, Junyu Lu, Youlin Wu, Zewen Bai, Changrong Min, Liang Yang, Bo Xu, Dongyu Zhang, and Hongfei Lin. 2024. Enhancing Textual Personality Detection toward Social Media: Integrating Long-term and Short-term Perspectives. arXiv preprint arXiv:2404.15067 (2024)

    2024

  47. [55]

    Haohao Zhu, Xiaokun Zhang, Junyu Lu, Liang Yang, and Hongfei Lin. 2024. Inte- grating multi-view analysis: Multi-view mixture-of-expert for textual personality detection. In CCF International Conference on Natural Language Processing and Chinese Computing. Springer, 359–371

    2024

  48. [56]

    Jianfeng Zhu, Ruoming Jin, and Karin G Coifman. 2025. Investigating Large Language Models in Inferring Personality Traits from User Conversations. arXiv preprint arXiv:2501.07532 (2025)

    2025 arXiv

  49. [57]

    Yangfu Zhu, Linmei Hu, Xinkai Ge, Wanrong Peng, and Bin Wu. 2022. Contrastive Graph Transformer Network for Personality Detection.. In IJCAI. 4559–4565

    2022

  50. [58]

    Yangfu Zhu, Linmei Hu, Nianwen Ning, Wei Zhang, and Bin Wu. 2022. A lexi- cal psycholinguistic knowledge-guided graph neural network for interpretable personality detection. Knowledge-Based Systems 249 (2022), 108952

    2022

  51. [59]

    Nikola Zubić, Federico Soldá, Aurelio Sulser, and Davide Scaramuzza. 2024. Limits of Deep Learning: Sequence Modeling through the Lens of Complexity Theory. arXiv preprint arXiv:2405.16674 (2024)

    2024 arXiv

  52. [2023]

    In 2023 IEEE 6th International Conference on Electronic Information and Communication Technology (ICEICT)

    Self-Attention and Transformers: Driving the Evolution of Large Language Models. In 2023 IEEE 6th International Conference on Electronic Information and Communication Technology (ICEICT). IEEE, 401–405

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.