Pith. sign in

REVIEW 4 major objections 6 minor 43 references

Exploring the Panorama of Anxiety Levels: A Multi-Scenario Study Based on Human-Centric Anxiety Level Detection and Personalized Guidance

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper reports a dialogue-based system that sorts anxiety into three levels with over 94% accuracy and supplies personalized explanations.

desk verdict The 94% accuracy claim is undermined by label circularity: ground truth is the GPT-4 prompt condition itself, so the paper measures prompt recovery, not anxiety detection. read the letter →

arxiv 2503.15527 v1 pith:S3XHUBA2 submitted 2025-02-04 cs.HC cs.AIcs.LG

Longdi Xian , Junhao Xu This is my paper · ORCID
classification cs.HCcs.AIcs.LG
keywords anxietydetectiontransformerencoderslargelanguagemodelsGPT-4simulateddialoguesnaturalprocessingmentalhealthpersonalizedguidancemulti-scenarioconversationdataset
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 single automated pipeline can read a two-person conversation, classify the speaker's anxiety into no, mild, or severe, and then explain and act on that classification. The authors generate a Chinese dialogue corpus with GPT-4 across education, work, daily life, and social scenarios; fine-tune four transformer encoders; and combine their predictions with a retrieval-based knowledge base that points to the anxiety-relevant sentences and produces personalized advice. The reported result is over 94% accuracy in categorical prediction. If this works on real human text, it would give mental-health support systems an interpretable alternative to bare category outputs.

What carries the argument

The machinery is a text-classification pipeline built from fine-tuned transformer encoders (BERT, RoBERTa, DistilBERT, ELECTRA) that map each dialogue through word and positional embeddings to a [CLS]-token representation, followed by a softmax layer over the three anxiety classes. A scenario-specific integration rule resolves disagreements among models, and five-fold cross-validation estimates accuracy. On the explanation side, a vector store of anxiety-related documents is retrieved through a conversational chain that receives both the conversation and the predicted label, enabling GPT-4 to quote the sentence that carried the anxiety and to generate advice.

What would settle it

Take the trained models and run them on a corpus of real human conversations whose anxiety levels are established by validated instruments such as clinical interviews or standardized self-report scales. If accuracy falls well below 94% or the quoted 'anxious sentences' do not match the sentences flagged by human raters, the framework's claim to human-centric anxiety detection would be refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that transformer encoder models trained on GPT-4-simulated multi-scenario conversations can predict anxiety levels with over 94% accuracy, and that a retrieval chain can turn each prediction into a concrete reason and a tailored suggestion. Different models are best in different contexts: RoBERTa is strongest in education, work, and daily life, while ELECTRA and BERT are combined for social scenarios. The paper frames this as a person-centered framework: the predicted label and the original dialogue are fed back together to the knowledge base, so the explanation is tied to the interlocutor's own sentences.

Load-bearing premise

The load-bearing assumption is that a GPT-4 prompt saying 'generate a dialogue with a little anxiety' produces text whose assigned label is a true measure of a person's anxiety; the 94% accuracy is computed against these self-generated labels, not against independently assessed human anxiety.

Editorial extensions

If this is right

  • A text-only anxiety screener could operate in ordinary settings where questionnaires are impractical, flagging mild or severe anxiety from how people talk.
  • The explanation step gives users a sentence-level reason for the prediction, which is the difference between a red flag and an actionable cue in counseling or self-help.
  • Scenario-specific model selection suggests that single global classifiers may be less accurate than routing to the best model for each context.
  • The generated Chinese multi-scenario dataset provides a reusable foundation for future work on simulated versus real emotional dialogue.

Reading between the lines

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

  • A decisive next experiment would replace GPT-4-generated labels with human annotation of real dialogues; if the model was actually learning prompt stereotypes rather than anxiety, its accuracy on human text would drop.
  • The same pipeline could be redirected to other emotional states or to real-time settings by swapping the scenario prompts and the retrieval corpus, though the paper does not test those extensions.
  • The study's reported accuracy is measured on the same generated distribution that produced the labels; interpreting it as clinical capability would require external validation on independently assessed human anxiety.
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. This paper proposes a three-part framework for anxiety-level detection in Chinese dialogues: (1) a GPT-4-generated synthetic dataset of two-person conversations across education, work, daily life, and social scenarios, with labels for no, mild, and severe anxiety; (2) fine-tuned transformer classifiers (BERT, RoBERTa, DistilBERT, ELECTRA) that predict these three anxiety levels, with a per-scenario model-integration scheme; and (3) a LangChain+GPT-4 knowledge base intended to provide explanations of the predicted level and personalized guidance. The authors report over 94% accuracy in categorical prediction and claim the advice is highly personalized.

Significance. The interpretability angle is timely: combining classification with explanations and tailored suggestions could be valuable for mental-health support tools. However, the scientific contribution depends entirely on the validity of the ground-truth labels, which are derived from the same GPT-4 prompt that generates the dialogue. The near-perfect accuracy figures are therefore best interpreted as measuring how well the model recovers the prompt condition, not as evidence of anxiety detection. The paper also provides no code, data, error bars, statistical tests, or independent human evaluation. The framework may have engineering interest, but the central empirical claim is not established by the evidence presented.

major comments (4)
  1. [Section 3.2.1, Table 1; Section 5.2] The ground-truth labels are defined by the generation prompt: the prompt instructs GPT-4 to generate a Chinese dialog 'that reflects a little anxiety' (or no/significant anxiety), and that instruction is stored as the label. The 94%+ accuracy reported in Section 5.2 is thus a measure of how well the classifier recovers the prompt-conditioning, not a validated measure of anxiety-level detection. This circularity is load-bearing because the abstract and conclusion claim the model 'can accurately recognize different levels of anxiety states'; without independent labels (e.g., human clinical ratings or validated scales), the central contribution is unsupported.
  2. [Section 4.2] The hyperparameter description is internally contradictory: the bullet list states 'Learning Rate: Set to 1e-5', and the same paragraph then gives 'learning_rate=5e-5'. No error bars, confidence intervals, or statistical tests are reported for any metric. The model-integration scheme in Section 5.2 is described as selecting the best model per scenario on the basis of the same validation folds used for the performance report, which introduces optimistic selection bias; the reported 'intersection' accuracy has no accompanying breakdown of how often the integration rule changed the final prediction.
  3. [Section 3.3 and Section 5.3] The knowledge base, the explanation generation, and the advice generation are all built from the same GPT-4-generated dialogues used to train the classifier, and the examples in Table 5 are four anecdotal cases with no human evaluation or user study. Therefore the claim that the advice is 'highly personalized and relevant' is not empirically supported; it is a description of the system's behavior on hand-picked inputs, not a measured outcome.
  4. [Section 5.2, Figure 6] The near-perfect accuracy values (e.g., 1.00 for RoBERTa in Daily and Social scenarios) are a red flag that the synthetic data are nearly separable by superficial lexical or stylistic cues that GPT-4 reliably produces when following the prompt. The paper provides no confusion matrices, error analysis, or evaluation on real-world or human-labeled conversations, so the generalization claims in the abstract and conclusion are untested.
minor comments (6)
  1. [Section 3.1] The phrase 'cue engineering' should be 'prompt engineering'; the same term is used repeatedly, and in Section 6 'harm connection GPT4.0' appears to be a typo for 'combining GPT4.0' or similar.
  2. [Equation (10), Section 4.1] The cross-entropy formula is written with a double sum over i and j but the indices and the true-label term y_i are not clearly defined; the equation should be corrected to the standard form over classes and samples.
  3. [Table 2] The example dialogue in the first row of Table 2 is inconsistent with the segmentation shown: the text says 'Oh, nothing much' but the segmentation lists 'Oh, ,, nothing, special, planned' and other words not present in the dialogue.
  4. [References] Several references are incomplete or inconsistently formatted (e.g., [18] and [21] are marked 'preprint' despite appearing in journals; [43] lacks author names), and some references are not cited in the body text.
  5. [Section 4.2] The text states that the response length is limited to 150 tokens and that 421 iterations are run per scenario, but it does not report how many generated dialogues failed validation, the average length of the final conversations, or any quality-filtering criteria.
  6. [Section 5.2] The paper reports per-scenario accuracies but does not report an aggregated overall accuracy with confidence intervals; the abstract's 'over 94% accuracy' is not directly tied to a specific table or metric definition.

Circularity Check

2 steps flagged · score 8.0 of 10

The 94% accuracy claim is circular: labels are GPT-4 prompt conditions, so the classifier is rewarded for recovering generation instructions, not detecting anxiety.

  1. self definitional [Section 3.2.1 Dataset Generating, step 2 (and Table 1)]
    "it constructs a prompt (shown in Table 1) by adding instructions to the text to generate a Chinese dialog that reflects a little anxiety. ... Each result (generated dialog) is added to the list of questions, and the corresponding anxiety level (anxiety level) is added to the list of answers."

    The ground-truth label is not an independent measurement of anxiety; it is the conditioning variable in the GPT-4 prompt. Because the same prompt specifies 'no anxiety, a little anxiety, super anxiety' and the returned level is stored as the label, the classification target is defined by the generation condition. A classifier that picks up GPT-4's prompt-induced phrasing is scoring recovery of the instruction, not detection of a psychological state. The task is therefore circular by construction: X (anxiety level) is defined by Y (prompt instruction), and Y generates the data used to train and evaluate the model.

  2. fitted input called prediction [Abstract; Section 5.2 Modeling Performance]
    "The study shows that the developed model achieves more than 94% accuracy in categorical prediction and that the advice provided is highly personalized."

    The 94% accuracy is measured on the same GPT-4-generated dataset whose labels were assigned as the prompt condition (Section 3.2.1). The reported 'prediction' therefore reduces to a reproduction check on the training-generation pipeline: high accuracy shows that transformer encoders can recognize the lexical/stylistic traces GPT-4 emits when told to write a 'little anxiety' or 'no anxiety' dialogue. No human ratings, clinical labels, or real-world conversations are used as an external benchmark, so the accuracy claim cannot support the conclusion that the model detects anxiety levels in people.

full rationale

The central empirical claim is the 94% accuracy in anxiety-level classification. That claim is measured against labels that are themselves the outputs of the prompt instructions used to generate the dialogues: each generated dialog is stored with the anxiety level named in the prompt as its ground truth. This makes the classification objective circular by construction. The knowledge-base explanation and advice components, built from the same GPT-4 generated documents and delivered by GPT-4, are not independently evaluated. There is no evidence of a self-citation chain or imported uniqueness theorem; the circularity is in the data-generation protocol, not in citation practice. Score 8 rather than 10 because the architecture itself (transformer fine-tuning on a text-label task) is a real technical procedure; what collapses is the interpretation of the accuracy as anxiety detection.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that GPT-4's synthetic conversations and self-assigned labels are a valid proxy for real anxiety. The only explicit free choices are ordinary training hyperparameters plus a data-dependent model selection rule. No new entities are introduced.

free parameters (5)
  • learning_rate = 1e-5 (also stated as 5e-5 in the same section)
    Reported in Section 4.2; contradictory, no tuning or sensitivity analysis.
  • batch_size = 8
    Section 4.2; chosen value without justification.
  • epochs = 2
    Section 4.2; chosen value without justification.
  • cross_validation_folds = 5
    Section 4.2; StratifiedKFold k=5 is a methodological choice, not derived.
  • per_scenario_model_selection = RoBERTa for education/work/daily; ELECTRA and Bert for social
    Section 5.2; models are chosen based on performance on the same folds, making the accuracy a fitted result.
assumptions (4)
  • domain assumption GPT-4-generated simulated conversations are representative of real anxiety-related human communication
    The dataset is exclusively generated by GPT-4 in Section 3.2.1; no real human conversations are collected.
  • ad hoc to paper The anxiety labels assigned by the generation prompt are valid ground truth
    Labels are set by the prompt ('reflects a little anxiety'), and accuracy in Section 5.2 is computed against these labels.
  • ad hoc to paper Accuracy on synthetic data transfers to real-world anxiety detection
    The abstract and conclusion generalize the result to mental health assistance without real-world validation.
  • ad hoc to paper The knowledge base built from the same GPT-4 documents provides valid explanations and personalized advice
    Sections 3.3 and 5.3 assume the RAG system's outputs are relevant and helpful without user or expert evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring the Panorama of Anxiety Levels: A Multi-Scenario Study Based on Human-Centric Anxiety Level Detection and Personalized Guidance." pith.science (2026). https://pith.science/paper/S3XHUBA2

@misc{pith2026250315527,
  author       = {Pith},
  title        = {Pith review of: Exploring the Panorama of Anxiety Levels: A Multi-Scenario Study Based on Human-Centric Anxiety Level Detection and Personalized Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S3XHUBA2}},
  note         = {Machine review of arXiv:2503.15527}
}
read the original abstract

More and more people are experiencing pressure from work, life, and education. These pressures often lead to an anxious state of mind, or even the early symptoms of suicidal ideation. With the advancement of artificial intelligence (AI) technology, large language models have become one of the most prominent technologies. They are often used for detecting psychological disorders. However, current studies primarily provide categorization results without offering interpretable explanations for these results. To address this gap, this study adopts a person-centered perspective and focuses on GPT-generated multi-scenario simulated conversations. These simulated conversations were selected as data samples for the study. Various transformer-based encoder models were utilized to develop a classification model capable of identifying different levels of anxiety. Additionally, a knowledge base focusing on anxiety was constructed using LangChain and GPT-4. When analyzing classification results, this knowledge base was able to provide explanations and reasons most relevant to the interlocutor's anxiety situation. The study demonstrates that the proposed model achieves over 94% accuracy in categorical prediction, and the advice provided is highly personalized and relevant.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 42 canonical work pages

  1. [1]

    Associations between maternal mental health, child dental anxiety, and oral health of 6- to 12-year-olds in nigeria

    Abiola Adetokunbo Adeniyi, Morenike Oluwatoyin Folayan, Nneka Maureen Chukwumah, Elizabeth Obhioneh Oziegbe, and Maha El Tantawi. Associations between maternal mental health, child dental anxiety, and oral health of 6- to 12-year-olds in nigeria. BRAZILIAN ORAL RESEARCH, 2023

    work page 2023

  2. [2]

    A. A. Azmee, M. Murikipudi, M. A. A. H. Khan, and Y. Pei. Sentence level analysis for detecting mental health causes using social media posts. In 2023 IEEE 47th Annual Computers, Software, and Applications Conference (COMPSAC), pages 1388–1393, 2023

    work page 2023

  3. [3]

    Chudzicka-Czupała, N

    A. Chudzicka-Czupała, N. Hapon, S.-K. Chiang, M. Żywiołek-Szeja, L. Karamushka, C. T. Lee, D. Grabowski, M. Paliga, J. D. Rosenblat, R. Ho, R. S. McIntyre, and Y.-L. Chen. Depression, anxiety and post-traumatic stress during the 2022 russo-ukrainian war, a comparison between populations in poland, ukraine, and taiwan. Scientific Reports, 13(1):3602, 2023

    work page 2022

  4. [4]

    Dominguez Camara

    V. Dominguez Camara. High frequency trading via transformer deep neural networks, 2023

    work page 2023

  5. [5]

    Sexual and gender minority adverse childhood experiences (sgm-aces), perceived social support, and adult mental health

    Armin A Dorri, Amy L Stone, Robert Salcido, Stephen T Russell, and Phillip W Schnarrs. Sexual and gender minority adverse childhood experiences (sgm-aces), perceived social support, and adult mental health. CHILD ABUSE & NEGLECT, 2023

    work page 2023

  6. [6]

    H. Fang, G. Xu, Y. Long, and W. Tang. An effective electra-based pipeline for sentiment analysis of tourist attraction reviews. Applied Sciences, 12(21):10881, 2022

    work page 2022

  7. [7]

    Giacchetti, G

    N. Giacchetti, G. M. Lattanzi, F. Aceti, N. Vanacore, and R. Williams. States of mind with respect to attachment: A comparative study between women who killed their children and mothers diagnosed with post-partum depression. Nordic Journal of Psychiatry, 77(1):3–13, 2023

    work page 2023

  8. [8]

    Godinic, B

    D. Godinic, B. Obrenovic, and A. Khudaykulov. Effects of economic uncertainty on mental health in the covid-19 pandemic context: Social identity disturbance, job uncertainty and psychological well-being model. International Journal of Innovation and Economic Development, 6(1):61–74, 2020

    work page 2020

Show all 43 references
  1. [9]

    Hull, Jacob Levine, Niels Bantilan, Angel N Desai, and Maimuna S Majumder

    Thomas D. Hull, Jacob Levine, Niels Bantilan, Angel N Desai, and Maimuna S Majumder. Complex psychological responses to the covid-19 pandemic: Digital phenotyping evidence from a large telemental health platform. JMIR FORMATIVE RESEARCH, 2021

    2021

  2. [10]

    Majumder

    Thomas Hull, Jacob Levine, Niels Bantilan, Angel Desai, and Maimuna S. Majumder. Digital phenotyping of complex psychological responses to the covid-19 pandemic. BMC Public Health, 21(1):314, 2021

    2021

  3. [11]

    Mental disorders detection with immediate message using roberta

    Chenghao Hu and Xiaobing Zhou. Mental disorders detection with immediate message using roberta. In IberLEF (Working Notes). CEUR Workshop Proceedings, 2023

    2023

  4. [12]

    Neuroscience

    Intelligence and C. Neuroscience. Retracted: A novel text mining approach for mental health prediction using bi-lstm and bert model. Computational Intelligence and Neuroscience, 2023:1–1, 2023

    2023

  5. [13]

    Kotera and D

    Y. Kotera and D. Sheffield. Nlp for japanese workers’ mental well-being: Pilot study. MENTAL HEALTH REVIEW JOURNAL, 2019

    2019

  6. [14]

    Kumar, R

    V. Kumar, R. Chaki, S. T. Talluri, G. Ramakrishnan, Y. F. Li, and G. Haffari. Question generation from paragraphs: A tale of two hierarchical models. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 2522–2529. IEEE, 2019

    2019

  7. [15]

    C. Lau, X. Zhu, and W. Y. Chan. Automatic depression severity assessment with deep learning using parameter-efficient tuning. Frontiers in Psychiatry, 14:1160291, 2023

    2023

  8. [16]

    Linden, L

    A. Linden, L. Best, F. Elise, D. Roberts, A. Branagan, Y. B. E. Tay, L. Crane, J. Cusack, B. Davidson, I. Davidson, C. Hearst, W. Mandy, D. Rai, E. Smith, and K. Gurusamy. Benefits and harms of interventions to improve anxiety, depression, and other mental health outcomes for ...

    2023

  9. [17]

    Associations of cardiorespiratory fitness, screen time and mental health among chinese school children

    Rui Li, Min-Zhe Zhang, Bo-Wen Tang, Jie Zhang, Geng-Cong Qu, Hui Li, and Yuan- Mei Shi. Associations of cardiorespiratory fitness, screen time and mental health among chinese school children. MEDICINE, 2022

    2022

  10. [18]

    Low, Laurie Rumker, Tanya Talkar, John Torous, Guillermo Cecchi, and Satrajit S

    Daniel M. Low, Laurie Rumker, Tanya Talkar, John Torous, Guillermo Cecchi, and Satrajit S. Ghosh. Natural language processing reveals vulnerable mental health support groups and heightened health anxiety on reddit during covid-19: Observational study (preprint). 2020

    2020

  11. [19]

    L. Lu, H. Shen, L. Tan, Q. Huang, Q. Chen, M. Liang, L. He, and Y. Zhou. Prevalence and factors associated with anxiety and depression among community-dwelling older adults in hunan, china: A cross-sectional study. BMC Psychiatry, 23(1):107, 2023

    2023

  12. [20]

    Maalouf, S

    E. Maalouf, S. Hallit, P. Salameh, and H. Hosseini. Depression, anxiety, insomnia, stress, and the way of coping emotions as risk factors for ischemic stroke and their influence on stroke severity: A case–control study in lebanon. Frontiers in Psychiatry, 14:1097873, 2023

    2023

  13. [21]

    Kunik, Aanand Naik, Hua Xu, and Javad Razjouyan

    Arash Maghsoudi, Sara Nowakowski, Ritwick Agrawal, Amir Sharafkhaneh, Sadaf Aram, Mark E. Kunik, Aanand Naik, Hua Xu, and Javad Razjouyan. Using natural language processing to assess the psychological effect of covid-19 pandemic on insomnia via tweets: A pre-post retrospective...

    2021

  14. [22]

    Leveraging twitter using artificial intelligence to explore mental health insights in the uk during the covid-19 pandemic (preprint)

    Christopher Marshall, Kate Lanyi, Rhiannon Green, Georgie Wilkins, Fiona Pearson, and Dawn Craig. Leveraging twitter using artificial intelligence to explore mental health insights in the uk during the covid-19 pandemic (preprint). 2021

    2021

  15. [23]

    Using natural language processing to explore mental health insights from uk tweets during the covid-19 pandemic: Infodemiology study

    Christopher Marshall, Kate Lanyi, Rhiannon Green, Georgina C Wilkins, Fiona Pearson, and Dawn Craig. Using natural language processing to explore mental health insights from uk tweets during the covid-19 pandemic: Infodemiology study. JMIR INFODEMIOLOGY, 2022

    2022

  16. [24]

    J. Oh, M. Kim, H. Park, and H. Oh. Are you depressed? analyze user utterances to detect depressive emotions using distilbert. Applied Sciences, 13(10):6223, 2023

    2023

  17. [25]

    Ozçift, K

    A. Ozçift, K. Akarsu, F. Yumuk, and C. Söylemez. Advancing natural language processing (nlp) applications of morphologically rich languages with bidirectional encoder representations from transformers (bert): an empirical case study for turkish. Automatika: časopis za autom...

    2021

  18. [26]

    Psychosocial stressors and maternal mental health in the u.s

    Angela M Parcesepe, Sarah G Kulkarni, Christian Grov, Rebecca Zimba, William You, Drew A Westmoreland, Amanda Berry, Shivani Kochhar, Madhura S Rane, Chloe Mirzayi, Andrew R Maroko, and Denis Nash. Psychosocial stressors and maternal mental health in the u.s. during the first ...

    2023

  19. [27]

    K. K. Patel, A. Pal, K. Saurav, and P. Jain. Mental health detection using transformer bert. In Handbook of Research on Lifestyle Sustainability and Management Solutions Using AI, Big Data Analytics, and Visualization, pages 91–108. IGI Global, 2022

    2022

  20. [28]

    Heidi F Pedersen, Tobias K Stripp, Niels C Hvidt, Tor-Arne Isene, Peter la Cour, Gry Stålsett, and Lars J Danbolt. Mental health of scandinavians during the first lockdown in the covid-19 pandemic: Psychosocial resources and coping strategies as protective or risk factors for...

    2023

  21. [29]

    Mental health is related to metabolic syndrome: The hispanic community health study/study of latinos

    Shannon L Richard, Brenna N Renn, Jinyoung Kim, Dieu-My T Tran, and Du Feng. Mental health is related to metabolic syndrome: The hispanic community health study/study of latinos. PSYCHONEUROENDOCRINOLOGY, 2023

    2023

  22. [30]

    Prevalence of mental disorders and associated factors in korean adults: National mental health survey of korea 2021

    Soo Jung Rim, Bong-Jin Hahm, Su Jeong Seong, Jee Eun Park, Sung Man Chang, Byung-Soo Kim, Hyonggin An, Hong Jin Jeon, and Subin Park. Prevalence of mental disorders and associated factors in korean adults: National mental health survey of korea 2021. PSYCHIATRY INVESTIGATION, 2023

    2021

  23. [31]

    Rothman and A

    D. Rothman and A. Gulli. Transformers for Natural Language Processing: Build, train, and fine-tune deep neural network architectures for NLP with Python, PyTorch, TensorFlow, BERT, and GPT-3. Packt Publishing Ltd, 2022

    2022

  24. [32]

    D. Rothman. Transformers for Natural Language Processing: Build innovative deep neural network architectures for NLP with Python, PyTorch, TensorFlow, BERT, RoBERTa, and more. Packt Publishing Ltd, 2021

    2021

  25. [33]

    M. R. Sarkar, S. G. Anavatti, T. Dam, M. Pratama, and B. Al Kindhi. Enhancing wind power forecast precision via multi-head attention transformer: An investigation on single-step and multi-step forecasting. In 2023 International Joint Conference on Neural Networks (IJCNN), page...

    2023

  26. [34]

    Simon, T

    J. Simon, T. M. Helter, R. G. White, C. Van Der Boor, and A. Łaszewska. Impacts of the covid-19 lockdown and relevant vulnerabilities on capability well-being, mental health and social support: An austrian survey study. BMC Public Health, 21(1):314, 2021

    2021

  27. [35]

    Su and K

    P. Su and K. Vijay-Shanker. Investigation of bert model on biomedical relation extraction based on revised fine-tuning mechanism. 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 2522–2529, 2020

    2020

  28. [36]

    A survey of mental health chatbots using nlp

    Abha Tewari, Amit Chhabria, Ajay Singh Khalsa, Sanket Chaudhary, and Harshita Kanal. A survey of mental health chatbots using nlp. SOCIAL SCIENCE RESEARCH NETWORK, 2021

    2021

  29. [37]

    Patterns and predictors of depressive and anxiety symptoms within a population-based sample of adults diagnosed with covid-19 in michigan

    Andrea R Titus, Briana Mezuk, Jana L Hirschtick, Patricia McKane, Michael R Elliott, and Nancy L Fleischer. Patterns and predictors of depressive and anxiety symptoms within a population-based sample of adults diagnosed with covid-19 in michigan. SOCIAL PSYCHIATRY AND PSYCHIAT...

    2023

  30. [38]

    D. Vucetic. Efficient fine-tuning of BERT-like models. PhD thesis, McGill University (Canada), 2022

    2022

  31. [39]

    B. Wang, L. Shang, C. Lioma, X. Jiang, H. Yang, Q. Liu, and J. G. Simonsen. On position embeddings in bert. In International Conference on Learning Representations, 2020

    2020

  32. [40]

    B. Wang, D. Zhao, C. Lioma, Q. Li, P. Zhang, and J. G. Simonsen. Encoding word order in complex embeddings. arXiv preprint arXiv:1912.12333, 2019

    1912 arXiv

  33. [41]

    H. Wu, Y. Zhang, L. Liang, X. Mei, D. Han, B. Han, and K. C. Li. Multi-head attention- based model for reconstructing continuous missing time series data. The Journal of Supercomputing, pages 1–28, 2023

    2023

  34. [42]

    Zanabazar, A

    A. Zanabazar, A. Deleg, and M. Ravdan. A study of factors causing math anxiety among undergraduate students. International Journal of Innovative Research and Scientific Studies, 6(3):578–585, 2023

    2023

  35. [43]

    The capability evaluation of chatgpt, gpt-4 ,tongyi qianwen on security domain with dikwp analysis agi-aigc-gpt test and evaluation dikwp (global) lab. 2023

    2023

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.