Pith. sign in

REVIEW 3 major objections 84 references

After screening, depressed users express risk in different layouts; one flat detector averages them away, so weak-prior routing to dense experts improves detection while staying deployable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 19:52 UTC pith:6CUWGJUJ

load-bearing objection Solid empirical methods paper: dense MoE under LUPI is the real win; the three soft layouts are secondary scaffolding, not the load-bearing mechanism. the 3 major comments →

arxiv 2607.04350 v1 pith:6CUWGJUJ submitted 2026-07-05 cs.CL

WPG-MoE: Weak-Prior-Guided Dense Mixture-of-Experts for User-Level Social Media Depression Detection

Xian Li , Yuanhe Tian , Yang Yang , Guoqing Wang , Yan Song This is my paper
classification cs.CL
keywords depression detectionsocial mediamixture of expertsweak priorslearning using privileged informationPHQ-9 screeninguser-level classificationevidence heterogeneity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Social-media depression detectors usually improve the screening step—picking risk posts, grounding symptoms, building clinical features—then hand everything to a single classifier. This paper argues that the real failure is what happens next: users still express risk differently (explicit self-disclosure, recurring episode-like blocks, or a few sparse high-intensity posts), and a monolithic model averages those patterns into one representation and one decision boundary. That averaging dilutes localized cues and especially hurts non-self-disclosing users. WPG-MoE answers with a dense mixture-of-experts on a shared language-model backbone. During training, offline LLM-extracted structure supplies soft user-level weak priors that gently route each user toward experts matched to those evidence layouts; at inference the system keeps only PHQ-9 template screening and the shared backbone, casting the gap as learning using privileged information. On Chinese and English user-level datasets the method beats strong screening-based baselines and shows readable gate behavior, while ablations and case studies tie the gains to preserving sparse and episode-supported cues that flat detectors blur.

Core claim

Post-screening evidence heterogeneity is a distinct failure mode for user-level social media depression detection: after risk posts are selected, depressed users still present overlapping but different evidence layouts, and a single detector that averages them dilutes localized signals. Softly routing users to dense experts with training-only weak priors induced from those layouts, under a LUPI train–deploy split that leaves only PHQ-9 screening at inference, recovers performance that flat detectors lose.

What carries the argument

WPG-MoE: dual-path evidence construction (privileged Path-A LLM scores for training; deployable Path-B PHQ-9 template similarity at inference) plus a dense five-expert mixture whose gates are softly guided by a three-component weak prior (self-disclosure, episode-supported, sparse high-risk) under LUPI, so privileged structure shapes routing without being required at deployment.

Load-bearing premise

The three soft evidence layouts pulled from the offline LLM scorer are stable, meaningful routing tendencies rather than artifacts of the scorer or residual label noise in the training data.

What would settle it

On a held-out set with human-adjudicated evidence layouts, replace weak-prior-guided dense experts by a single shared classifier (or by unguided MoE) while keeping the same backbone and PHQ-9 screening; if the AUPRC/F1 gap on sparse and episode-supported slices disappears or reverses, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper argues that after risk-post screening, depressed users still express risk through heterogeneous evidence layouts (self-disclosure, episode-supported, sparse high-risk, mixed), so a single detector averages away localized cues. It proposes WPG-MoE: a dense five-expert MoE on a shared LLM backbone, softly routed by user-level weak priors π derived from training-only LLM-structured Path-A fields, cast as LUPI so inference uses only PHQ-9 template screening (Path B), the shared backbone, history segments, and lightweight statistics. Contributions are (i) diagnosing post-screening heterogeneity, (ii) weak-prior-guided dense routing without hard subtypes, and (iii) a deployable LUPI split. Support includes three datasets under a unified holdout, seven baselines, matched-backbone replacements, ablations, human layout diagnostics (κ≈0.71), Path-A audits, controlled mixing, gate analyses, seed-wise spreads, and case studies.

Significance. If the result holds, this is a useful systems contribution for clinical NLP: it reframes user-level depression detection as a post-screening specialization problem and shows a practical LUPI recipe that keeps costly LLM annotation out of deployment. Strengths that should be credited include multi-dataset controlled comparison (Table 4), matched-backbone isolation of architecture vs encoder (Table 5 / A.9), human evidence-layout diagnostics with substantial agreement (Table 3; A.8), Path-A field audits (A.1/A.3), full ablations and seed-wise reliability (Tables 6, 23–24), and explicit train–deploy alignment mechanisms (Table 1). The work is significant for mental-health screening pipelines even if the three named layouts are better treated as soft inductive biases than as validated clinical subtypes.

major comments (3)
  1. Table 6 (§4.4) and the full matrix in Table 23 show a consistent component ordering that undercuts the paper’s distinctive mechanism claim. On SWDD, removing dense MoE drops AUPRC 0.830→0.693 and removing Path A drops to 0.758, while w/o Weak Priors (0.800) and w/o Route Loss (0.818) remain close to full model. The same pattern holds on Twitter and eRisk25. The Abstract, §1 contributions, and §2.3 present clinically grounded weak priors for self-disclosure / episode / sparse layouts as what makes routing work; the ablations instead isolate multi-view MoE capacity plus privileged Path-A supervision as the load-bearing pieces, with L_route and π as secondary shapers. The central claim remains defensible under a broader LUPI+dense-MoE reading, but the manuscript should restate the mechanism to match what Table 6 isolates, or add a control that keeps Path-A features while randomizing/shuffli
  2. §2.3 Eq. (5) and Appendix A.3: the three soft layouts are induced from the same Path-A fields used to build evidence blocks and silver evidence labels. Human audits show Path-A fields and user-layout assignment are recoverable (κ≈0.65–0.71; assignment Acc. 0.772), which rules out pure noise, but does not establish that these three tendencies are the causal routing structure rather than convenient summaries of privileged score mass. Controlled mixing (Fig. 4) and gate plots (Figs. 5–6) are compatible with overlapping multi-view capacity; they do not falsify a “richer training supervision” alternative. A load-bearing revision is either (i) an experiment that freezes Path-A candidate quality while ablating layout-specific prior structure, or (ii) a clearer claim that π is a soft training regularizer, not a validated clinical typology.
  3. §3.1 / Appendix A.7: SWDD results use a manually corrected self_reported split (399 raw 0→1, 16 raw 1→0). This is responsible, but main Table 4 and transfer cells that train on SWDD are not accompanied by a sensitivity check on the raw vs corrected labels for WPG-MoE and the strongest baselines. Because Path-A priors and self-disclosure routing lean on self-report structure, the SWDD gains (and SWDD→* transfer) need a short raw-label or leave-correction-out control so the improvement is not partly an artifact of the audit that also defines the self-disclosure slice.

Circularity Check

1 steps flagged

No load-bearing circular derivation; only a minor by-construction flavor in the interpretability of routing under L_route.

specific steps
  1. self definitional [§2.4 Eq. (9) L_route; Abstract / §4.3 gate analyses]
    "Lroute aligns the first three expert gates with confident weak priors... Experiments on Chinese and English datasets show that WPG-MoE outperforms strong baselines with interpretable routing behavior."

    The training objective explicitly pulls the first three gate weights toward the weak priors π derived from Path-A layouts; reporting that average gates emphasize self-disclosure / episode / sparse views as “interpretable routing” is therefore partly by construction of L_route rather than an independent discovery. This is minor and non-load-bearing: ablations show w/o Route Loss retains most gains, and the main metric is depression-label prediction, not prior recovery.

full rationale

WPG-MoE is an empirical architecture paper, not a first-principles derivation. The central claims—post-screening evidence heterogeneity hurts flat detectors, and weak-prior-guided dense MoE under LUPI improves user-level depression detection while remaining deployable with only Path-B PHQ-9 screening—are evaluated against independent binary depression labels, held-out users, cross-dataset transfer, and external baselines (Table 4). Weak priors π = Prior(RA, B) and L_route that aligns the first three gates with confident priors are transparent training-time privileged supervision (LUPI), not a prediction that equals its fit by construction: the predicted quantity is the depression label y, not reconstruction of π. Human evidence-layout diagnostics are assigned from raw packets without Path-A scores (Appendix A.8), and ablations show residual gains after removing weak priors or route loss, so performance is not forced by the prior-alignment term. Self-citations (e.g., Tian et al.) appear in related work and are not uniqueness theorems that forbid alternatives. The only mild self-referential note is that “interpretable routing” matching self-disclosure/episode/sparse layouts is partly encouraged by L_route; that does not make the detection results circular. Score 1 for that minor interpretability-by-construction flavor only.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central claim rests on modeling assumptions about evidence layouts and a LUPI training/inference split, plus many engineering hyperparameters. No new physical entity is postulated; the invented pieces are architectural (expert views, weak-prior vector, dual-path scoring). Free parameters are loss and routing knobs chosen for training stability rather than fitted to claim the target metric by construction.

free parameters (5)
  • Loss weights α, β, γ and entropy schedule δt
    Eq. (9) jointly optimizes classification, route alignment, evidence supervision, balance, and entropy with α=0.3, β=0.2, γ=0.15 and entropy decay 0.1→0.02; these are hand-set training knobs that shape reported gains.
  • Routing-loss confidence thresholds
    Route loss applies only when max weak prior ≥0.6 and margin ≥0.1 (Appendix A.5); these cutoffs decide which users receive prior-aligned gate supervision.
  • Training perturbation rates
    Risk-source swap 0.5, META dropout 0.5, episode-block dropout 0.4, prior dropout 0.3 (Table 1) are chosen rates that align privileged and deployable paths.
  • Candidate budget K(n) and 12.5% screening ratio
    Top-K risk-post selection uses a history-length rule with nominal 12.5% ratio (Eq. 4; Appendix A.5), controlling which posts enter the model.
  • Path-A composite score weights wA
    Fixed non-negative weights map six LLM attributes to sA_i; even if not validation-tuned, they define the privileged ranking that induces priors and blocks.
axioms (5)
  • domain assumption Post-screening depressive risk is usefully summarized by soft evidence layouts (self-disclosure, episode-supported, sparse high-risk) rather than a single averaged user representation.
    Stated in §1 and operationalized in §2.3; human diagnostics support difficulty ordering but do not prove clinical subtypes.
  • domain assumption PHQ-9 symptom templates are an adequate deployable screening interface for Path B.
    Path B and inference depend on PHQ-9 template similarity (§2.2; Appendix A.2), following prior clinical-NLP practice.
  • domain assumption Privileged LLM-extracted structure can supervise routing at training time without being required at inference (LUPI).
    Core design in Abstract/§2; standard LUPI framing (Vapnik & Vashist, 2009) applied to this task.
  • ad hoc to paper Dense soft routing over five experts is preferable to hard subtype partitioning for overlapping evidence tendencies.
    Motivated by controlled mixing (§4.3) and used as the architectural commitment in Eq. (7).
  • domain assumption Standard supervised classification metrics on stratified user holdouts are valid proxies for early depression screening utility in this study.
    Evaluation protocol in §3; paper explicitly notes this is controlled comparison, not clinical diagnosis or official eRisk delay-aware scoring.
invented entities (3)
  • Weak prior vector π = [π_self, π_epis, π_sparse] independent evidence
    purpose: Soft training-time routing hints summarizing evidence-layout tendencies from Path A and blocks.
    Defined in Eq. (5); not a clinical label set, and not available at inference.
  • Five expert views (SD, EP, SP, MIX, G) no independent evidence
    purpose: Specialize user representations for self-disclosure, episode, sparse, mixed, and global residual history signals.
    Architectural invention in §2.3–2.4; gate analyses provide internal interpretability but no external clinical validation of the five-way split.
  • Path-A composite evidence scorer ScoreA / offline structured fields independent evidence
    purpose: Produce privileged post scores, blocks, and priors used only in training.
    External Qwen3-max schema in Appendix A.6; audited against humans but remains a paper-specific privileged pipeline.

pith-pipeline@v1.1.0-grok45 · 40434 in / 3765 out tokens · 40208 ms · 2026-07-11T19:52:35.033866+00:00 · methodology

0 comments
read the original abstract

Online social media posts provide scalable signals for early depression screening, and recent studies mainly improve pre-classification evidence through risk-post selection, symptom grounding, and clinically informed feature construction. However, these screening-stage designs often leave final decisions to a single detector, overlooking how users heterogeneously express depressive risk after screening. A monolithic classifier must average across heterogeneous users, which may dilute localized evidence and cause misclassification, especially for non-self-disclosing users. To address this issue, we propose WPG-MoE, a weak-prior-guided dense mixture-of-experts framework built on a shared large language model (LLM) backbone. WPG-MoE derives user-level weak semantic priors to softly route users to experts matched to different evidence layouts. We formulate this process as learning using privileged information (LUPI): rich LLM-extracted structured evidence guides training-time routing, while inference retains only Patient Health Questionnaire-9 (PHQ-9) template screening and the deployable backbone. Experiments on Chinese and English datasets show that WPG-MoE outperforms strong baselines with interpretable routing behavior.

Figures

Figures reproduced from arXiv: 2607.04350 by Guoqing Wang, Xian Li, Yang Yang, Yan Song, Yuanhe Tian.

Figure 1
Figure 1. Figure 1: Heterogeneous evidence patterns create a mis [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall pipeline of WPG-MoE. Path A provides training-time privileged evidence from an external scorer, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Human-labeled evidence-layout diagnostic. Six flat detectors’ AUPRC/F1 on a blinded cross-dataset [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Controlled mixing on SWDD. Top row targets [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average training-time gate weights. E1–E5 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case-wise comparison on two anonymized SWDD users. The flat model denotes DeCapsNet; SP, EP, [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Large-scale rule-derived slice diagnostic on corrected SWDD. AUPRC and F1 are reported for flat [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

84 extracted references · 6 linked inside Pith

  1. [1]

    Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Improving the generalizability of depression detection by leveraging clinical questionnaires , author=. Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  2. [2]

    , title=

    Zhang, Zhiling and Chen, Siyuan and Wu, Mengyue and Zhu, Kenny Q. , title=. Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence,

  3. [3]

    Expert Systems with Applications , volume=

    Depression detection on online social network with multivariate time series feature of user depressive symptoms , author=. Expert Systems with Applications , volume=

  4. [4]

    Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence,

    Shen, Guangyao and Jia, Jia and Nie, Liqiang and Feng, Fuli and Zhang, Cunjun and Hu, Tianrui and Chua, Tat-Seng and Zhu, Wenwu , title=. Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence,

  5. [5]

    European Conference on Information Retrieval , pages=

    eRisk 2025: contextual and conversational approaches for depression challenges , author=. European Conference on Information Retrieval , pages=

    2025

  6. [6]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    Identifying fine-grained depression signs in social media posts , author=. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    2024

  7. [7]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages=

    Diverse perspectives, divergent models: Cross-cultural evaluation of depression detection on Twitter , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages=

    2024

  8. [8]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    Depression detection on social media with large language models , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    2025

  9. [9]

    Proceedings of the 9th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2024) , pages=

    Explainable depression detection using large language models on social media data , author=. Proceedings of the 9th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2024) , pages=

    2024

  10. [10]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    End-to-End Learnable Psychiatric Scale Guided Risky Post Screening for Depression Detection on Social Media , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  11. [11]

    proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , pages=

    Depressionnet: learning multi-modalities with user post summarization for depression detection on social media , author=. proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , pages=

  12. [12]

    Proceedings of the international AAAI conference on web and social media , volume=

    Predicting depression via social media , author=. Proceedings of the international AAAI conference on web and social media , volume=

  13. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Depression detection via capsule networks with contrastive learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  14. [14]

    Neural networks , volume=

    A new learning paradigm: Learning using privileged information , author=. Neural networks , volume=

  15. [15]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Cooperative multimodal approach to depression detection in twitter , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  16. [16]

    Information Sciences , volume=

    Online social network individual depression detection using a multitask heterogenous modality fusion approach , author=. Information Sciences , volume=

  17. [17]

    PLoS medicine , volume=

    The global gap in treatment coverage for major depressive disorder in 84 countries from 2000--2019: A systematic review and Bayesian meta-regression analysis , author=. PLoS medicine , volume=

    2000

  18. [18]

    World mental health report: Transforming mental health for all , author=

  19. [19]

    2025 , month = aug, url =

    Depressive Disorder (Depression) , author =. 2025 , month = aug, url =

    2025

  20. [20]

    Neural computation , volume=

    Adaptive mixtures of local experts , author=. Neural computation , volume=

  21. [21]

    arXiv preprint arXiv:1701.06538 , year=

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer , author=. arXiv preprint arXiv:1701.06538 , year=

    Pith/arXiv arXiv

  22. [22]

    IEEE Latin America Transactions , volume=

    Mental health prediction from social media text using mixture of experts , author=. IEEE Latin America Transactions , volume=

  23. [23]

    IEEE Access , year=

    Mixture of experts for depression and anxiety disorder prediction from textual and non-textual social media data , author=. IEEE Access , year=

  24. [24]

    Proceedings of the 2017 conference on empirical methods in natural language processing , pages=

    Depression and self-harm risk assessment in online forums , author=. Proceedings of the 2017 conference on empirical methods in natural language processing , pages=

    2017

  25. [25]

    Proceedings of the 32nd Pacific Asia conference on language, information and computation , year=

    Feature attention network: interpretable depression detection from social media , author=. Proceedings of the 32nd Pacific Asia conference on language, information and computation , year=

  26. [26]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Are LLMs effective psychological assessors? Leveraging adaptive RAG for interpretable mental health screening through psychometric practice , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  27. [27]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Explainable Depression Detection in Clinical Interviews with Personalized Retrieval-Augmented Generation , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  28. [28]

    Journal of Machine Learning Research , volume=

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity , author=. Journal of Machine Learning Research , volume=

  29. [29]

    arXiv preprint arXiv:1511.03643 , year=

    Unifying distillation and privileged information , author=. arXiv preprint arXiv:1511.03643 , year=

    Pith/arXiv arXiv

  30. [30]

    Proceedings of the workshop on computational linguistics and clinical psychology: From linguistic signal to clinical reality , pages=

    Quantifying mental health signals in Twitter , author=. Proceedings of the workshop on computational linguistics and clinical psychology: From linguistic signal to clinical reality , pages=

  31. [31]

    Proceedings of the 2nd workshop on computational linguistics and clinical psychology: from linguistic signal to clinical reality , pages=

    From ADHD to SAD: Analyzing the language of mental health on Twitter through self-reported diagnoses , author=. Proceedings of the 2nd workshop on computational linguistics and clinical psychology: from linguistic signal to clinical reality , pages=

  32. [32]

    Proceedings of the international AAAI conference on web and social media , volume=

    Mental health discourse on reddit: Self-disclosure, social support, and anonymity , author=. Proceedings of the international AAAI conference on web and social media , volume=

  33. [33]

    Sensitive Self-disclosures, Responses, and Social Support on Instagram: The Case of

    Andalibi, Nazanin and. Sensitive Self-disclosures, Responses, and Social Support on Instagram: The Case of. Proceedings of the 2017

    2017

  34. [34]

    Proceedings of the 2018 EMNLP workshop SMM4H: the 3rd social media mining for health applications workshop & shared task , pages=

    Identifying depression on reddit: The effect of training data , author=. Proceedings of the 2018 EMNLP workshop SMM4H: the 3rd social media mining for health applications workshop & shared task , pages=

    2018

  35. [35]

    Proceedings of the fifth workshop on computational linguistics and clinical psychology: from keyboard to clinic , pages=

    Deep learning for depression detection of twitter users , author=. Proceedings of the fifth workshop on computational linguistics and clinical psychology: from keyboard to clinic , pages=

  36. [36]

    Proceedings of the fourth workshop on computational linguistics and clinical psychology—from linguistic signal to clinical reality , pages=

    Monitoring tweets for depression to detect at-risk users , author=. Proceedings of the fourth workshop on computational linguistics and clinical psychology—from linguistic signal to clinical reality , pages=

  37. [37]

    Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic , pages=

    Rsdd-time: Temporal annotation of self-reported mental health diagnoses , author=. Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic , pages=

  38. [38]

    Proceedings of the 28th international conference on computational linguistics , pages=

    Identifying depressive symptoms from tweets: Figurative language enabled multitask learning framework , author=. Proceedings of the 28th international conference on computational linguistics , pages=

  39. [39]

    NPJ digital medicine , volume=

    Methods in predictive techniques for mental health status on social media: a critical review , author=. NPJ digital medicine , volume=

  40. [40]

    Proceedings of the 11th international workshop on health text mining and information analysis , pages=

    Detection of mental health from reddit via deep contextualized representations , author=. Proceedings of the 11th international workshop on health text mining and information analysis , pages=

  41. [41]

    Proceedings of the tenth international workshop on health text mining and information analysis (LOUHI 2019) , pages=

    Multi-task, multi-channel, multi-input learning for mental illness detection using social media text , author=. Proceedings of the tenth international workshop on health text mining and information analysis (LOUHI 2019) , pages=

    2019

  42. [42]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    Classifying social media users before and after depression diagnosis via their language usage: A dataset and study , author=. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    2024

  43. [43]

    Computational Linguistics , volume=

    Comparing Selective Masking Methods for Depression Detection in Social Media , author=. Computational Linguistics , volume=

  44. [44]

    Journal of medical Internet research , volume=

    Ai for analyzing mental health disorders among social media users: Quarter-century narrative review of progress and challenges , author=. Journal of medical Internet research , volume=

  45. [45]

    Proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2025) , pages=

    Datasets for depression modeling in social media: An overview , author=. Proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2025) , pages=

    2025

  46. [46]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    ReDepress: A Cognitive Framework for Detecting Depression Relapse from Social Media , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  47. [47]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    Mentalhelp: A multi-task dataset for mental health in social media , author=. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    2024

  48. [48]

    Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Modeling task relationships in multi-task learning with multi-gate mixture-of-experts , author=. Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  49. [49]

    Neural Networks , volume=

    Learning using privileged information: SVM+ and weighted SVM , author=. Neural Networks , volume=

  50. [50]

    and Williams, Janet B

    Kroenke, Kurt and Spitzer, Robert L. and Williams, Janet B. W. , title=. Journal of General Internal Medicine , volume=. 2001 , issn=

    2001

  51. [51]

    Advances in Neural Information Processing Systems , volume=

    Mixture-of-experts with expert choice routing , author=. Advances in Neural Information Processing Systems , volume=

  52. [52]

    Proceedings of the 9th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2024) , pages=

    Using large language models (llms) to extract evidence from pre-annotated social media data , author=. Proceedings of the 9th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2024) , pages=

    2024

  53. [53]

    and Jackson, Thomas W

    Chen, Xuetong and Sykora, Martin D. and Jackson, Thomas W. and Elayan, Suzanne , title=. Companion of The Web Conference 2018,

    2018

  54. [54]

    Proceedings of the 2nd workshop on computational linguistics and clinical psychology: from linguistic signal to clinical reality , pages=

    CLPsych 2015 shared task: Depression and PTSD on Twitter , author=. Proceedings of the 2nd workshop on computational linguistics and clinical psychology: from linguistic signal to clinical reality , pages=

    2015

  55. [55]

    Proceedings of the National Academy of Sciences , volume=

    Facebook language predicts depression in medical records , author=. Proceedings of the National Academy of Sciences , volume=

  56. [56]

    Current Opinion in Behavioral Sciences , volume=

    Detecting depression and mental illness on social media: an integrative review , author=. Current Opinion in Behavioral Sciences , volume=

  57. [57]

    Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology , pages=

    Then and now: Quantifying the longitudinal validity of self-disclosed depression diagnoses , author=. Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology , pages=

  58. [58]

    JMIR AI , volume=

    Enabling early health care intervention by detecting depression in users of web-based forums using language models: longitudinal analysis and evaluation , author=. JMIR AI , volume=

  59. [59]

    EPJ Data Science , volume=

    Instagram photos reveal predictive markers of depression , author=. EPJ Data Science , volume=

  60. [60]

    IEEE transactions on knowledge and data engineering , volume=

    Utilizing neural networks and linguistic metadata for early detection of depression indications in text sequences , author=. IEEE transactions on knowledge and data engineering , volume=. 2018 , publisher=

    2018

  61. [61]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    ESDM: Early sensing depression model in social media streams , author=. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages=

    2024

  62. [62]

    JMIR mental health , volume=

    Deep learning with anaphora resolution for the detection of tweeters with depression: Algorithm development and validation study , author=. JMIR mental health , volume=

  63. [63]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Towards identifying fine-grained depression symptoms from memes , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  64. [64]

    2019 , address=

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle=. 2019 , address=

    2019

  65. [65]

    He, Pengcheng and Gao, Jianfeng and Chen, Weizhu , booktitle=

  66. [66]

    2026 , eprint=

    Ministral 3 , author=. 2026 , eprint=

    2026

  67. [67]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    Named entity recognition for social media texts with semantic augmentation , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  68. [68]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    medikal: Integrating knowledge graphs as assistants of llms for enhanced clinical diagnosis on emrs , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

  69. [69]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Dialogue summarization with mixture of experts based on large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  70. [70]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Multimodal aspect-based sentiment analysis with plugin-enhanced large language models , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

  71. [71]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Chimed-gpt: A chinese medical large language model with full training regime and better alignment to human preferences , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  72. [72]

    Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    ZEN: Pre-training Chinese text encoder enhanced by n-gram representations , author=. Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    2020

  73. [73]

    Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    Topic memory networks for short text classification , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    2018

  74. [74]

    Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers , pages=

    Learning user embeddings from emails , author=. Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers , pages=

  75. [75]

    Proceedings of the ACM Web Conference 2023 , pages=

    Hashtag-guided low-resource tweet classification , author=. Proceedings of the ACM Web Conference 2023 , pages=

    2023

  76. [76]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Resemo: A benchmark chinese dataset for studying responsive emotion from social media content , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  77. [77]

    A Gold Standard Dataset and Evaluation Framework for Depression Detection and Explanation in Social Media using

    Bolegave, Prajval and Bhattacharya, Pushpak , year=. A Gold Standard Dataset and Evaluation Framework for Depression Detection and Explanation in Social Media using. 2507.19899 , archivePrefix=

    Pith/arXiv arXiv

  78. [78]

    Generating Medically-Informed Explanations for Depression Detection using

    Chen, Xiangyong and Lin, Xiaochuan , year=. Generating Medically-Informed Explanations for Depression Detection using. 2503.14671 , archivePrefix=

    Pith/arXiv arXiv

  79. [79]

    Interpretable Depression Detection from Social Media Text Using

    Kim, Samuel and Imieye, Oghenemaro and Yin, Yunting , year=. Interpretable Depression Detection from Social Media Text Using. 2506.06616 , archivePrefix=

    Pith/arXiv arXiv

  80. [80]

    2510.11233 , archivePrefix=

    Xu, Jinyuan and Lan, Tian and Yu, Xintao and He, Xue and Zhang, Hezhi and Wang, Ying and Magistry, Pierre and Valette, Mathieu and Li, Lei , year=. 2510.11233 , archivePrefix=

    Pith/arXiv arXiv

Showing first 80 references.