Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Explainable Depression Detection using Masked Hard Instance Mining

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

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

Pith's one-line read This paper argues that masking a depression detector's most-attended sentences during a second training phase yields attention that is more evenly spread and more clinically aligned, improving prediction and explainability on both Thai…

desk verdict A legitimate transfer of MHIM to text depression detection with the right instinct for external labels, but the explainability claim is confounded by the method's own entropy increase and lacks the controlled baselines it needs. read the letter →

arxiv 2505.24609 v1 pith:NJTHPKVM submitted 2025-05-30 cs.CL

Patawee Prakrankamanant , Shinji Watanabe , Ekapol Chuangsuwanich This is my paper
classification cs.CL
keywords depressiondetectionexplainableAIattentionmechanismmulti-instancelearningmaskedhardinstancemininglow-resourceNLPThaiclinicalinterviewsDAIC-WOZ
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

Depression screening models typically output a score without a trustworthy explanation. This paper argues that a model can be trained to give both. The proposed method, Masked Hard Instance Mining (MHIM), works in two phases: a first Dual Encoder learns normally, then a second is trained with the first model's highest-attended sentences masked out, forcing its attention to spread across more of the interview. On Thai (Thai-Maywe) and English (DAIC-WOZ) clinical interviews, the masked model improves prediction error, attention entropy, and agreement with clinician-labeled important sentences, while the same attention weights become better at locating sentences whose removal changes the prediction. If the claim holds, a cheap training-time regularizer can improve both accuracy and explainability in low-resource clinical NLP.

What carries the argument

The load-bearing mechanism is a training-time attention mask inside the Multiple Instance Learning aggregation layer, where each interview is a bag of sentence instances and attention weights serve as the pooling operation. In phase 1, a donor Dual Encoder is trained normally. In phase 2, a second Dual Encoder is trained while the donor's top attention weights are masked out of the receiver, so the receiver cannot rely on the same few sentences and must find alternative evidence. The paper tracks attention entropy as a direct measure of diversification, and uses Recall@k against clinician labels, plus an RMSE-after-dropping sensitivity analysis, as measures of explainability.

What would settle it

Evaluate the baseline and MHIM models on a held-out set of clinical interviews where expert annotators label every sentence's relevance to the target symptom, then compare Recall@k of each model's attention rankings; if the MHIM receiver does not beat the baseline donor at multiple k values, or if the RMSE-drop curves are reproduced by randomly chosen sentence removals, the central explainability claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that forcing attention diversification through MHIM makes the attention mechanism a better explanation device without sacrificing prediction quality. Starting from the Dual Encoder, which pools sentence representations with attention, the authors train a donor model, obtain its attention scores, and then train a second model that is blinded to the donor's top-attended sentence embeddings, plus a few random low-attention masks. At inference the mask is removed, yet the receiver's learned attention has higher entropy and better predictive performance on both datasets. On Thai HAM-D 1, where clinician-provided Importance Sentence Labels exist, the MHIM model's Recall@k beats the baseline at every cutoff and the qualitative top-3 sentences show more symptom-related content; on overall HAM-D and DAIC-WOZ PHQ-8, where no sentence labels exist, the authors use the proxy that removing the model's top-attended sentences raises RMSE more than random removal, which they interpret as evidence that the top attention identifies important evidence.

Load-bearing premise

The gain rests on the assumption that forcing attention to spread out during training makes attention weights at inference genuinely better at pointing to clinically relevant sentences, and that the evaluation metrics, namely Recall@k on one labeled item plus the RMSE-after-dropping proxy, actually measure that relevance.

Editorial extensions

If this is right

  • On Thai-Maywe, adding MHIM to the Dual Encoder lowers RMSE on HAM-D 1 from 0.54 to 0.48 and on overall HAM-D from 3.54 to 3.15, while MAE improves on HAM-D 1.
  • Attention entropy rises with MHIM, from 2.406 to 2.496 on HAM-D 1 and from 2.158 to 2.819 on overall HAM-D, supporting the claim that attention spreads more evenly.
  • Against clinician-provided Importance Sentence Labels on HAM-D 1, MHIM raises Recall@k at every reported cutoff, for example from 0.52 to 0.62 at k=50% and from 0.85 to 0.90 at k=80%.
  • On the English DAIC-WOZ dataset, MHIM keeps prediction RMSE comparable or slightly better (0.733 to 0.717) while the attention-based dropping analysis shows improved sensitivity to important sentences.
  • The two-phase masking also improves the Single Encoder variant, so the benefit does not depend on the Dual Encoder architecture specifically.

Reading between the lines

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

  • A natural one-phase alternative would be to add an explicit attention-entropy regularizer during training; if entropy gain is the true cause, a single run could capture most of MHIM's benefit at lower cost.
  • The RMSE-drop proxy measures sensitivity, not clinical validity, so an independent human study of whether the model's chosen sentences are genuinely more useful to clinicians would strengthen or refute the explanation claim.
  • Because MHIM only requires attention weights from a donor model, it could transfer to other weakly labeled bag-classification tasks, such as pathology slide analysis or social-media screening, where attention overfits to a few instances.
  • The paper's observation that the LLM baseline underperforms on longer explanation lists hints that attention-based MIL may be better suited than prompting for exhaustive explanations.
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 / 4 minor

Summary. The paper proposes Masked Hard Instance Mining (MHIM), a two-phase training procedure for attention-based depression detection. In phase 1 a Dual Encoder is trained to predict depression severity; in phase 2 a second model is trained with attention masking derived from the phase-1 model: t top-attended and b bottom-attended sentence embeddings are masked, forcing the receiver to distribute attention more broadly. The method is evaluated on Thai-Maywe (HAM-D) and DAIC-WOZ (PHQ-8 item 2), comparing Dual Encoder, Single Encoder, and LLM prompting baselines on RMSE/MAE and on attention-based explainability metrics: Recall@k against human Importance Sentence Labels for HAM-D1, and a drop-top-attention sensitivity analysis for overall HAM-D and DAIC-WOZ. The authors claim that MHIM significantly improves both prediction accuracy and explainability.

Significance. If the claims held, MHIM would be a cheap and appealing regularizer for clinical NLP: training to spread attention would improve both accuracy and interpretability in low-resource settings, an important problem. The paper has useful strengths: it evaluates on two languages/corpora, reports attention entropy, makes one direct comparison with human Importance Sentence Labels (Thai HAM-D1), and includes baselines from prior work. However, the current evidence does not establish the central claims: prediction gains are within noise, and the explainability evidence is either confounded with entropy or circular. With additional controlled experiments and significance testing, the ideas could become a solid contribution.

major comments (4)
  1. [Section 5.1, Table 1] The claim that MHIM 'significantly improves' prediction is not supported by inferential statistics; the reported gains are mostly within one standard deviation (e.g., Thai HAM-D1 RMSE 0.54±0.28 vs 0.48±0.28, MAE 0.37±0.19 vs 0.33±0.19), and overall HAM-D MAE actually worsens from 2.47±0.76 to 2.52±0.60, while the DAIC-WOZ results in Section 5.3 (RMSE 0.733 to 0.717) are reported without error bars or significance tests.
  2. [Section 5.2, Table 2] Recall@k against human ISL labels is reported only for HAM-D1, with no precision, error bars, or significance tests, and no random/uniform or entropy-matched baseline; because the method's explicit mechanism is to increase attention entropy (Section 3, Table 1 shows entropy rising from 2.406 to 2.496 and 2.158 to 2.819), the large gains at k=50-90% are mechanically expected for any spread-out attention distribution, so the table cannot currently distinguish 'better explanations' from 'broader attention'.
  3. [Section 5.2, Figures 2-3] The 'drop top-attended sentences and measure RMSE rise' analysis measures model reliance, not clinical relevance, making it circular as an explainability claim for overall HAM-D and DAIC-WOZ; this is not an external validation, especially since the paper concedes in Section 5.2 that the model is 'prone to cases where a single high-ISL sentence may lead to a false positive.'
  4. [Section 3 and Section 5.2] Because the phase-2 training objective explicitly encourages more spread attention and the same attention weights are then used as the explanation, the explainability evaluation needs to control for entropy (e.g., match entropy between baseline and MHIM or report precision at fixed recall); without such a control, the self-referential loop between training objective and evaluation metric remains unresolved.
minor comments (4)
  1. [Sections 3 and 4.1] 'PHD-Q' appears to be a typo for 'PHQ-8' (twice), and 'HAM-D or PHD-Q score' should be corrected.
  2. [Section 3] 'The masks comes from two sets' should read 'The masks come from two sets'; additionally, Table 1 shows '0.542.52' with a missing separator between MAE and entropy for the MHIM row.
  3. [Section 2.4] The phrase 'adjacency to human attention' is vague; consider clarifying what property of attention is being claimed.
  4. [Section 4.3, Equation 2] The notation uses n for both the number of sentences and the denominator sum; if the index i runs over sentences, clarify that and define n accordingly.

Circularity Check

1 steps flagged · score 2.0 of 10

No formal circularity: ISL-based Recall@k is externally grounded, but the overall-HAM-D/DAIC sensitivity analysis shares the attention mechanism with the explanation, creating a mild self-referential loop.

  1. other [Section 5.2 (Model Explainability on the Thai-Maywe dataset) and Figures 2-3]
    "For the overall HAM-D, since there are no ISL labels. Instead, we performed a kind of sensitivity analysis on how the model prediction will change if different sentences were dropped from the input. If important sentences are dropped, it should have a high impact on the model's performance."

    In this proxy evaluation, the candidate 'important sentences' are selected by the model's own top attention scores, and the evidence is the change in the same model's prediction when those sentences are removed. The selection criterion (attention) and the evaluation criterion (attention-weighted impact on the model) are therefore the same machinery being explained. This makes the RMSE-shift test a within-model consistency check rather than an external test of HAM-D relevance, so the conclusion that MHIM is 'more effective at identifying important sentences' is partially self-referential. However, this step is not load-bearing for the whole paper because Table 2's Recall@k uses external ISL labels, and no formal equation reduces the claimed explainability gain to the training objective.

full rationale

The paper's derivation chain is mostly self-contained. The central explainability claim for HAM-D1 is measured by Recall@k (Eq. 2) against human ISL labels, which are external to the model and independent of the MHIM training objective. This is genuine evidence, not circular. The main soft spot is the sensitivity analysis used for overall HAM-D and DAIC-WOZ: it drops sentences chosen by the model's attention and measures the RMSE increase in the same model, so the test confirms that the model's attention is influential to the model's own output, not that the sentences are clinically relevant. This is a self-referential proxy, but it is not a formal derivation of the result from its inputs, and the paper itself concedes a false-positive limitation. The prediction-performance gains (Table 1, DAIC) are small and often within one standard deviation, which is a statistical-significance concern, not circularity. The citations to [8], [14], and related work are to external methods; the in-group dataset/encoder citations ([10], [31]) do not carry the load of the central claim. Overall, no step reduces by construction to its own input, so the paper receives a low score with a minor self-referential caveat.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. It relies on the standard MIL framing, the assumption that attention scores are meaningful explanations, and the ad hoc hypothesis that attention spread improves both accuracy and explainability. The only tuned parameters are the mask counts t and b, whose values are not disclosed.

free parameters (2)
  • t (number of top-attention masks) = not reported
    Chosen on validation set; r=2t, total N=t+b. No values given.
  • b (number of bottom-attention masks) = not reported
    Also selected on validation set; not reported.
assumptions (4)
  • domain assumption Attention weights in the trained model reflect input relevance for depression prediction.
    Section 2.4 assumes attention provides explainability; this is the basis for Recall@k and dropping analyses.
  • ad hoc to paper Spreading attention across more sentences causally improves prediction and explainability.
    Section 3 hypothesis: low-entropy attention overfits; the method forces spread. No independent evidence that spread is causally beneficial.
  • domain assumption The donor model's attention masks transfer appropriately to the receiver in two-phase training.
    Section 3 Phase 2 assumes donor masks are a good guide for the receiver.
  • domain assumption MIL bag assumption: the interview label is determined by at least one positive instance.
    Section 2.2; inherited from [8].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explainable Depression Detection using Masked Hard Instance Mining." pith.science (2026). https://pith.science/paper/NJTHPKVM

@misc{pith2026250524609,
  author       = {Pith},
  title        = {Pith review of: Explainable Depression Detection using Masked Hard Instance Mining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NJTHPKVM}},
  note         = {Machine review of arXiv:2505.24609}
}
read the original abstract

This paper addresses the critical need for improved explainability in text-based depression detection. While offering predictive outcomes, current solutions often overlook the understanding of model predictions which can hinder trust in the system. We propose the use of Masked Hard Instance Mining (MHIM) to enhance the explainability in the depression detection task. MHIM strategically masks attention weights within the model, compelling it to distribute attention across a wider range of salient features. We evaluate MHIM on two datasets representing distinct languages: Thai (Thai-Maywe) and English (DAIC-WOZ). Our results demonstrate that MHIM significantly improves performance in terms of both prediction accuracy and explainability metrics.

Figures

Figures reproduced from arXiv: 2505.24609 by the authors.

Figure 1
Figure 1. The overview of our proposed pipeline. The training process consists of two phases: 1 and 2. The first phase follows the standard Dual Encoder training procedure, where each input sentence is encoded and aggregated to form an interview representation. This representation is then processed through an attention layer to predict the depression severity score (HAM-D or PHQ-8). In the second phase, the process remains si… view at source ↗
Figure 2
Figure 2. shows the effect of dropping the important sen￾tences according to the top attention scores. The figure com￾pares dropping sentences using the attention scores from the Dual Encoder baseline, the Dual Encoder with MHIM, and ran￾dom dropping. The plot demonstrates that our technique is more effective at identifying important sentences, as shown by the up￾ward RMSE shift compared to other methods [PITH_FULL_IMAGE:fig… view at source ↗
Figure 3
Figure 3. Prediction results on the DAIC-WOZ dataset when the top-k features are dropped using different dropping criterion. 6. Conclusion We proposed a 2-phase attention-based depression detec￾tion model training pipeline that leveraged the Masked Hard Instance Mining (MHIM) technique and the Multi-instance Learning (MIL) framework. The pipeline was specifically de￾signed to promote diversified attending pattern, thereby les… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages

  1. [1]

    Despite its widespread impact, many individuals re- main unaware of their susceptibility to depression, which of- ten leads to delayed diagnosis and treatment [2]

    Introduction Major Depressive Disorder (MDD) is a pervasive mental health condition affecting individuals worldwide [1], and the World Health Organization (WHO) estimates it as a leading cause of disability. Despite its widespread impact, many individuals re- main unaware of their susceptibility to depression, which of- ten leads to delayed diagnosis and ...

    work page

  2. [2]

    Explainable Depression Detection using Masked Hard Instance Mining

    Background 2.1. Dual Encoder Model The state-of-the-art (SOTA) model for text-based depression de- tection is the the Dual Encoder [8]. As shown in the top half of Figure 1, the dialogue between the interviewer and the subject is broken down into parts such as question and response pairs. These parts are then encoded using two different encoders, a Prefix...

    work page Pith review arXiv 2025

  3. [3]

    We find that this procedure often suffer from overfitting, especially on low-data regime

    Methods In the Dual Encoder, the input sentences are fed directly into the model for depression score prediction. We find that this procedure often suffer from overfitting, especially on low-data regime. Our investigation indicates that attention mechanism in the Dual Encoder tend to concentrate to only a small portion of the input features, exhibiting lo...

    work page

  4. [4]

    Datasets To benchmark the performance of our method, we have cho- sen two datasets for benchmarking: the DAIC-WOZ [11] and the Thai-Maywe dataset [10]

    Experiment Setup 4.1. Datasets To benchmark the performance of our method, we have cho- sen two datasets for benchmarking: the DAIC-WOZ [11] and the Thai-Maywe dataset [10]. The DAIC-WOZ dataset is a set of 189 clinical interviews conducted in English, utilizing the PHQ-8 depression screening questionnaire [25], which in- cludes eight items. For this stud...

    work page

  5. [5]

    Model Performance on the Thai-Maywe dataset Table 1 shows the prediction performance on the Thai-Maywe dataset of our method compared to other baselines

    Result & Discussion 5.1. Model Performance on the Thai-Maywe dataset Table 1 shows the prediction performance on the Thai-Maywe dataset of our method compared to other baselines. Both the Dual Encoder and Single Encoder models achieve improved performance by using the MHIM method, as measured by RMSE and MAE. Moreover, the MHIM properly encourages the att...

    work page

  6. [6]

    The pipeline was specifically de- signed to promote diversified attending pattern, thereby less prone to overfitting in low-data settings

    Conclusion We proposed a 2-phase attention-based depression detec- tion model training pipeline that leveraged the Masked Hard Instance Mining (MHIM) technique and the Multi-instance Learning (MIL) framework. The pipeline was specifically de- signed to promote diversified attending pattern, thereby less prone to overfitting in low-data settings. We found ...

    work page

  7. [7]

    Acknowledgements This research project is supported by the Second Century Fund (C2F), Chulalongkorn University

    work page

  8. [8]

    Depression and other common mental disorders: global health estimates,

    W. H. Organizationet al., “Depression and other common mental disorders: global health estimates,”World Health Organization, 2017

    work page 2017

Show all 39 references
  1. [9]

    Projections of global mortality and burden of disease from 2002 to 2030,

    C. D. Mathers and D. Loncar, “Projections of global mortality and burden of disease from 2002 to 2030,”PLoS medicine, vol. 3, no. 11, p. e442, 2006

    2002

  2. [10]

    A text classification framework for simple and effective early depression detection over social media streams,

    S. G. Burdisso, M. Errecalde, and M. Montes-y G ´omez, “A text classification framework for simple and effective early depression detection over social media streams,”Expert Systems with Appli- cations, vol. 133, pp. 182–197, 2019

    2019

  3. [11]

    Real-time acoustic based depression detection using machine learning techniques,

    B. Yalamanchili, N. S. Kota, M. S. Abbaraju, V . S. S. Nadella, and S. V . Alluri, “Real-time acoustic based depression detection using machine learning techniques,” in2020 international confer- ence on emerging trends in information technology and engineer- ing (ic-ETITE). IE...

    2020

  4. [12]

    Deep learning for depression detection of twitter users,

    A. H. Orabi, P. Buddhitha, M. H. Orabi, and D. Inkpen, “Deep learning for depression detection of twitter users,” inProceedings of the fifth workshop on computational linguistics and clinical psychology: from keyboard to clinic, 2018, pp. 88–97

    2018

  5. [13]

    Detecting linguistic traces of depression in topic-restricted text: Attending to self-stigmatized depression with nlp,

    J. Wolohan, M. Hiraga, A. Mukherjee, Z. A. Sayyed, and M. Mil- lard, “Detecting linguistic traces of depression in topic-restricted text: Attending to self-stigmatized depression with nlp,” inPro- ceedings of the first international workshop on language cognition and computati...

    2018

  6. [14]

    Xa-bilstm: a deep learning approach for depression detection in imbalanced data,

    Q. Cong, Z. Feng, F. Li, Y . Xiang, G. Rao, and C. Tao, “Xa-bilstm: a deep learning approach for depression detection in imbalanced data,” in2018 IEEE international conference on bioinformatics and biomedicine (BIBM). IEEE, 2018, pp. 1624–1627

    2018

  7. [15]

    Automatic depression severity assessment with deep learning using parameter-efficient tuning,

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

    2023

  8. [16]

    Explainable depression symp- tom detection in social media,

    E. Bao, A. P ´erez, and J. Parapar, “Explainable depression symp- tom detection in social media,”Health Information Science and Systems, vol. 12, no. 1, p. 47, 2024

    2024

  9. [17]

    An application for mental health monitoring using facial, voice, and questionnaire information,

    S. Boonvitchaikul, N. Cheetanom, T. Sompong, J. Sununt- nasuk, S. Thammarerkrit, P. Pongpanatapipat, P. Aeuepalisa, A. Kuasakunrungroj, C. Viriyasuthee, P. Prakrankamanantet al., “An application for mental health monitoring using facial, voice, and questionnaire information,” ...

    2023

  10. [18]

    The Distress Analysis Interview Corpus of human and computer interviews

    J. Gratch, R. Artstein, G. M. Lucas, G. Stratou, S. Scherer, A. Nazarian, R. Wood, J. Boberg, D. DeVault, S. Marsellaet al., “The Distress Analysis Interview Corpus of human and computer interviews.” inLREC. Reykjavik, 2014, pp. 3123–3128

    2014

  11. [19]

    Multiple instance learning: algorithms and applica- tions,

    B. Babenko, “Multiple instance learning: algorithms and applica- tions,”View Article PubMed/NCBI Google Scholar, vol. 19, 2008

    2008

  12. [20]

    Attention-based deep mul- tiple instance learning,

    M. Ilse, J. Tomczak, and M. Welling, “Attention-based deep mul- tiple instance learning,” inInternational conference on machine learning. PMLR, 2018, pp. 2127–2136

    2018

  13. [21]

    Multiple instance learning framework with masked hard instance mining for whole slide image classification,

    W. Tang, S. Huang, X. Zhang, F. Zhou, Y . Zhang, and B. Liu, “Multiple instance learning framework with masked hard instance mining for whole slide image classification,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4078–4087

    2023

  14. [22]

    P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks,

    X. Liu, K. Ji, Y . Fu, W. Tam, Z. Du, Z. Yang, and J. Tang, “P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), S. Muresan, P. Na...

    2022

  15. [23]

    Universal sentence encoder for English,

    D. Cer, Y . Yang, S.-y. Kong, N. Hua, N. Limtiaco, R. St. John, N. Constant, M. Guajardo-Cespedes, S. Yuan, C. Tar, B. Strope, and R. Kurzweil, “Universal sentence encoder for English,” inProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: S...

    2018

  16. [24]

    Multi instance neural networks,

    J. Ramon, “Multi instance neural networks,” inML-2000 Work- shop Attribute-Value and Relational Learning, 2000

    2000

  17. [25]

    A brief introduction to weakly supervised learning,

    Z.-H. Zhou, “A brief introduction to weakly supervised learning,” National science review, vol. 5, no. 1, pp. 44–53, 2018

    2018

  18. [26]

    Trans- mil: Transformer based correlated multiple instance learning for whole slide image classification,

    Z. Shao, H. Bian, Y . Chen, Y . Wang, J. Zhang, X. Jiet al., “Trans- mil: Transformer based correlated multiple instance learning for whole slide image classification,”Advances in neural information processing systems, vol. 34, pp. 2136–2147, 2021

    2021

  19. [27]

    Neural machine transla- tion by jointly learning to align and translate,

    D. Bahdanau, K. Cho, and Y . Bengio, “Neural machine transla- tion by jointly learning to align and translate,” in3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Y . Ben- gio and Y . LeCun, E...

    2015

  20. [28]

    What does BERT look at? an analysis of BERT‘s attention,

    K. Clark, U. Khandelwal, O. Levy, and C. D. Manning, “What does BERT look at? an analysis of BERT‘s attention,” inProceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, T. Linzen, G. Chrupała, Y . Belinkov, and D. Hupkes, Eds. Flo...

    2019

  21. [29]

    Mpnet: Masked and permuted pre-training for language understanding,

    K. Song, X. Tan, T. Qin, J. Lu, and T.-Y . Liu, “Mpnet: Masked and permuted pre-training for language understanding,”Advances in neural information processing systems, vol. 33, pp. 16 857– 16 867, 2020

    2020

  22. [30]

    Human attention maps for text classification: Do humans and neural networks focus on the same words?

    C. Sen, T. Hartvigsen, B. Yin, X. Kong, and E. Rundensteiner, “Human attention maps for text classification: Do humans and neural networks focus on the same words?” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai...

    2020

  23. [31]

    On exploring attention-based explanation for transformer models in text clas- sification,

    S. Liu, F. Le, S. Chakraborty, and T. Abdelzaher, “On exploring attention-based explanation for transformer models in text clas- sification,” in2021 IEEE International Conference on Big Data (Big Data), 2021, pp. 1193–1203

    2021

  24. [32]

    Towards automatic text- based estimation of depression through symptom prediction,

    K. Milintsevich, K. Sirts, and G. Dias, “Towards automatic text- based estimation of depression through symptom prediction,” Brain Informatics, vol. 10, no. 1, p. 4, 2023

    2023

  25. [33]

    The phq-8 as a measure of current depression in the general population,

    K. Kroenke, T. W. Strine, R. L. Spitzer, J. B. Williams, J. T. Berry, and A. H. Mokdad, “The phq-8 as a measure of current depression in the general population,”Journal of affective disorders, vol. 114, no. 1-3, pp. 163–173, 2009

    2009

  26. [34]

    The hamilton rating scale for depression,

    M. Hamilton, “The hamilton rating scale for depression,” inAs- sessment of depression. Springer, 1986, pp. 143–152

    1986

  27. [35]

    The dsm-5: Classifi- cation and criteria changes,

    D. A. Regier, E. A. Kuhl, and D. J. Kupfer, “The dsm-5: Classifi- cation and criteria changes,”World psychiatry, vol. 12, no. 2, pp. 92–98, 2013

    2013

  28. [36]

    Roberta: A robustly optimized bert pre- training approach (2019),

    L. Yinhan, O. Myle, G. Naman, D. Jingfei, J. Mandar, C. Danqi, L. Omer, and L. Mike, “Roberta: A robustly optimized bert pre- training approach (2019),”arXiv preprint arXiv:1907.11692, pp. 1–13, 2019

    2019 arXiv

  29. [38]

    Wangchanberta: Pretraining transformer-based thai lan- guage models,

    L. Lowphansirikul, C. Polpanumas, N. Jantrakulchai, and S. Nu- tanong, “Wangchanberta: Pretraining transformer-based thai lan- guage models,”arXiv preprint arXiv:2101.09635, 2021

    2021 arXiv

  30. [39]

    ConGen: Unsupervised control and generalization distillation for sentence representation,

    P. Limkonchotiwat, W. Ponwitayarat, L. Lowphansirikul, C. Udomcharoenchaikit, E. Chuangsuwanich, and S. Nutanong, “ConGen: Unsupervised control and generalization distillation for sentence representation,” inFindings of the Association for Computational Linguistics: EMNLP 2022...

    2022

  31. [286]

    Available: https://aclanthology.org/W19-4828/

    [Online]. Available: https://aclanthology.org/W19-4828/

Pith tools

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