Beyond Scale: Small Language Models are Comparable to GPT-4 in Mental Health Understanding

Feng Xia; Hong Jia; Shiya Fu; Ting Dang; Vassilis Kostakos

REVIEW 5 major objections 5 minor 1 cited by

Beyond Scale: Small Language Models are Comparable to GPT-4 in Mental Health Understanding

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Small language models come within two F1 points of GPT-4 on coarse mental-health classification, and a few examples let Phi-3 beat it on stress detection.

desk verdict New SLM scores on mental-health tasks are worth having, but the 'comparable to GPT-4' claim rests on borrowed baselines and inconsistent numbers. read the letter →

arxiv 2507.08031 v2 pith:NES2NJOU submitted 2025-07-09 cs.CL

Hong Jia , Shiya Fu , Feng Xia , Vassilis Kostakos , Ting Dang This is my paper · ORCID

classification cs.CL

keywords smalllanguagemodelsmentalhealthclassificationzero-shotpromptingfew-shotlearningstressdetectiondepressionseveritysuicideriskmacroF1

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

This paper asks whether small language models (SLMs) understand mental-health language well enough to serve as privacy-preserving alternatives to large models. On four binary tasks—stress, depression, suicidal ideation, and suicide risk—five current SLMs achieve a mean macro F1 of 0.64, within 0.02 of the 0.66 posted by the LLM baselines in zero-shot prompting. Few-shot prompting lifts small models further, with Phi-3 reaching 0.81 on stress detection and beating GPT-4's 0.78. Both model families lose a substantial and comparable share of F1 on multi-class severity tasks, suggesting the hard part of mental-health understanding is severity grading, not scale. If the comparison holds, mental-health screening can be done locally on small models without giving up much accuracy.

What carries the argument

The machinery is a six-task benchmark divided into binary condition detection and multi-class severity assessment, with prompting as the only adaptation mechanism. Zero-shot prompts concatenate the raw text, a context strategy (social-media origin, mental-health-expert framing, or both), a query, and strict output-format constraints, using five query variants per task with scores averaged. Few-shot prompts prepend balanced prompt-label example pairs, one to four per class, to the zero-shot template. Macro F1 is the reported metric because the datasets are class-imbalanced. This design lets the authors isolate model understanding from fine-tuning, and it is the comparison apparatus that produces both the near-parity binary result and the shared severity collapse.

What would settle it

Re-run the three LLM baselines with the exact five prompt variants, few-shot templates, temperature, and F1 script used for the SLMs; if the re-measured LLM mean on the four binary tasks exceeds the SLM mean by more than about 0.02, or if GPT-4's stress-detection F1 stays above Phi-3's in the few-shot setting with matched inputs, the central parity claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that parameter count is not a reliable predictor of mental-health understanding. Across six tasks built from Reddit data, SLMs—Phi-3 and Phi-3.5 in particular—perform close to or better than LLM baselines on binary classifications: zero-shot mean macro F1 is 0.64 for SLMs versus 0.66 for LLMs, and Phi-3.5-Mini is the second-best model on every task, ahead of FLAN-T5-XXL and Alpaca-7B. With few-shot examples, Phi-3 improves by 14.6% on stress detection to an F1 of 0.81, surpassing GPT-4's 0.78, while GPT-4's few-shot results are flat or slightly negative. On four- and five-level severity tasks, both model classes drop to much lower F1 scores, showing that fine-grained clinical grading is a shared bottleneck rather than a small-model shortfall. The authors read this as evidence that SLMs can be effective, adaptable, privacy-preserving tools for coarse-grained mental-health screening.

Load-bearing premise

The load-bearing premise is that the GPT-4, FLAN-T5-XXL, and Alpaca-7B scores quoted from the earlier Mental-LLM study were produced under the same prompts, decoding settings, and scoring rules as the SLM runs; if that protocol match does not hold, the two-point gap could be a measurement artifact.

Editorial extensions

If this is right

Binary mental-health screening can plausibly run locally on a phone or edge device, since SLMs reach within about 0.02 macro F1 of GPT-4's zero-shot result without sending text to a cloud API.
Adding just one to four labeled examples per class is enough to specialize an SLM for a target condition; Phi-3's stress-detection F1 rises from 0.66 to 0.81, overtaking GPT-4's 0.78.
Multi-class severity assessment—depression or suicide-risk level—remains unreliable for both SLMs and LLMs, so automated triage should be limited to coarse yes/no signals until severity-level accuracy improves.
Model scale is not the main driver of mental-health understanding: a 3-billion-parameter Phi-3.5-Mini beats 7B and 11B baselines, and the binary-vs-severity gap is nearly identical across all model sizes.

Reading between the lines

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

Because the SLM few-shot gains are consistent while GPT-4's are flat or negative, a cost-conscious deployment could treat a small model plus a handful of hand-labeled examples as a cheaper substitute for a frontier API; the paper measures accuracy but not inference cost or latency.
The shared drop on severity tasks hints at a common training-data ceiling rather than a capacity limit; a direct test would be fine-tuning an SLM on severity labels and checking whether it closes that gap.
The paper's privacy argument becomes stronger if future work verifies that local inference keeps raw posts on-device end to end; current results only show that the smaller models have the accuracy, not that a deployed app actually avoids network transmission.
The longest inputs are user-level suicide-risk records averaging around 1,900 tokens; a natural stress test is whether near-parity with GPT-4 holds when inputs are truncated or noisy, which could change the model choice for real social-media feeds.

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.

Desk Editor's Note

read the letter

First, the useful part: the paper contributes new zero- and few-shot F1 numbers for five small models (Gemma-2B, Llama-3.2-3B, Qwen2.5-3B, Phi-3, Phi-3.5) on six standard mental-health text tasks. The best small model, Phi-3.5, is competitive with the two weaker LLM baselines and clearly behind GPT-4; few-shot prompting lifts Phi-3 above GPT-4 on the stress task. Those specific measurements are new, and the paper is clearly written.

The main weakness is that the LLM numbers are not re-run. They are imported from Mental-LLM [6], with only 'same experimental design' and 'similar' prompt settings as evidence of comparability. No prompt templates, decoding parameters, or evaluation scripts are provided. On skewed data, a 0.02 F1 gap can easily be explained by prompt wording or parsing differences. So the within-2% claim is not established by this paper.

The title overstates as well. The 0.64 vs 0.66 comparison is against the mean of three LLMs; GPT-4 alone sits around 0.755 on the four binary tasks, about 0.12 above the small models. 'Comparable to GPT-4' is only true for Phi-3.5 on the stress task after few-shot.

There are internal inconsistencies: the abstract says the multi-class drop is 'over 30%,' the results section says 'approximately 50%,' and the table supports the latter. The few-shot narrative says GPT-4 degraded 1.2% on stress; the table says -0.2%. The few-shot claim is also based on only two of the five SLMs, which the paper does not call out. And the citation to [4] in Section IV.B is wrong—that is a social media and mental health paper, not the LLM study they mean.

On the positive side, the task design is sensible, the datasets are the usual ones, and the degradation on severity tasks looks real even with the borrowed baselines. The paper would be more honest as 'small models are competitive with the average of older LLMs' rather than 'comparable to GPT-4.'

Who should read it: people tracking small-model progress on health text classification. A serious referee should not desk-reject it, but should require a controlled head-to-head or a much more careful framing of the borrowed baselines. I'd send it to review with heavy revision expected.

Referee Report

5 major / 5 minor

Summary. The paper evaluates five small instruction-tuned language models (Phi-3, Phi-3.5, Qwen2.5, Llama-3.2, Gemma2) on six mental-health text classification tasks covering stress, depression, suicidal ideation, suicide risk, depression severity, and suicide-risk severity. Using zero-shot and few-shot prompting, it compares macro-F1 scores against three LLMs (GPT-4, FLAN-T5-XXL, Alpaca-7B), where the LLM scores are taken from a prior study [6]. The authors report that SLMs reach mean performance within 2% F1 of the mean LLM performance on binary tasks, that both model families degrade similarly on multi-class severity tasks, and that few-shot prompting allows Phi-3 to exceed GPT-4 on stress detection (0.81 vs 0.78). They conclude that SLMs are promising for privacy-preserving, on-device mental-health screening.

Significance. If the empirical claims were supported by a controlled comparison, the paper would be useful: it provides a systematic zero-shot and few-shot benchmark of current 2B–4B models on mental-health tasks, uses public datasets, reports macro-F1 rather than accuracy, and includes an explicit ethics statement. The practical implication for privacy-preserving screening is real and worth investigating. However, the headline claim of comparability to GPT-4 is not supported by the data as presented, because GPT-4 alone outperforms the SLMs by a wide margin in Table II and the few-shot comparison is cross-study rather than a direct, controlled evaluation.

major comments (5)

[Section III and Table II] The central comparison is uncontrolled: the GPT-4, FLAN-T5-XXL, and Alpaca-7B scores in Table II and Table III are taken from prior study [6] rather than re-run with this paper's prompts, decoding parameters, and evaluation script. Section III states only that these are 'baseline results from previous research' and Section II.B says the prompts are 'similar to [6]', but no prompt templates, five query variants, few-shot examples, or parsing rules are provided. Because macro-F1 on these datasets is sensitive to prompt wording, temperature, output parsing, and few-shot contents, the 0.02-point gap in the binary tasks and the 0.81-vs-0.78 few-shot result could be protocol artifacts rather than model-capability differences. This limitation is not disclosed in Section V. Without re-running the baselines or otherwise demonstrating protocol equivalence, the paper's central quantitative claims are not established.
[Abstract vs. Section IV.A.2] The claimed degradation on multi-class tasks is internally inconsistent: the abstract says both model categories experience 'a drop of over 30%', while Section IV.A.2 states 'a drop of approximately 50%' and gives values consistent with the 50% figure (e.g., binary SLM mean 0.64 vs. multi-class mean 0.34). The abstract needs correction, and the discrepancy should be resolved before revision.
[Section II.B and Tables II–III] No measures of uncertainty are reported. Section II.B states that five query variants are averaged per task, but Tables II and III present only point estimates, with no standard deviations, confidence intervals, or significance tests. Consequently, the 0.64-vs-0.66 mean difference and the 0.81-vs-0.78 few-shot difference cannot be distinguished from run-to-run or prompt-variant noise.
[Title and Table II] The title's claim that SLMs are 'Comparable to GPT-4' is not supported by the data shown. In Table II, GPT-4's mean binary F1 across T1, T2, T4, and T5 is approximately 0.755, while the SLM mean is 0.64; the within-2% statement in the abstract refers to the mean of all three LLMs, which is heavily influenced by weaker models (Alpaca-7B and FLAN-T5-XXL). The paper should either compare directly against GPT-4 or rephrase the headline to refer to the LLM mean.
[Section IV.B and Table III] The prose misreports Table III. The text says GPT-4 has a 'slight performance degradation (1.2%)' on Task 1, but the table shows -0.2%; it says Alpaca-7B gains '10.2% on T2', but the table shows +0.7%; and it says FLAN-T5 improves on one task while declining on another, but the table shows positive relative improvements on all three tasks for FLAN-T5. The results narrative needs to be reconciled with the table.

minor comments (5)

[Table I] The DepSeverity class percentages sum to 90% in both the training and test rows (73% + 17% and 72% + 18%), which is likely a typographical error.
[Section II.C] The word 'varing' should be 'varying' in the sentence describing the number of shots.
[Section IV.B] The sentence 'following similar experimental settings as previous LLM studies [4]' cites reference [4], which is a behavioral-science paper on social media and mental health, not a language-model few-shot study; reference [6] appears to be the intended citation.
[Section I] The phrase 'speech-based cues' is misleading for a study that analyzes text posts; it should be 'linguistic' or 'text-based' cues.
[Table I] The average token lengths for DepSeverity are identical to those of Dreaddit (114±41 and 113±37), which may be a copy-paste error and should be verified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the SLM results are new empirical measurements and the LLM baselines are imported from an external prior study, not from the authors' own fitted outputs.

full rationale

The paper's central claim is an empirical comparison, not a derivation. The SLM numbers (Phi-3, Phi-3.5, Qwen2.5, Llama-3.2, Gemma2) are newly run evaluations using the specified zero-shot and few-shot prompt structures; there is no fitted parameter that is later renamed as a prediction, and no equation in which an output is defined in terms of the target result. The LLM columns in Tables II and III are taken from the external Mental-LLM study [6] ('we include baseline results from previous research [6]'), which is an independent published benchmark, not a quantity derived by this paper from its own definitions. A protocol-mismatch or uncontrolled-comparison concern about using 'similar' prompts without shared code would be a validity threat to the comparison, but it is not circularity: the prior numbers do not depend on the present paper's outputs. The only self-citation traceable to the authors ([5], a background citation on multimodal LLMs in health) is non-load-bearing. Accordingly, no circular step can be exhibited with the required quotation-and-reduction evidence, and the score is 0.

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

The paper introduces no fitted parameters or invented entities. Its empirical claims rest on three assumptions: that the reused LLM baselines from [6] are directly comparable to the authors' SLM runs, that macro F1 over five prompt variants is a reliable measure of understanding, and that the chosen Reddit datasets are valid proxies for mental health states. None of these assumptions is validated inside the paper.

assumptions (4)

domain assumption LLM F1 scores from Mental-LLM [6] are directly comparable to the authors' SLM runs under identical protocols.
The paper reuses external baselines instead of re-running the LLMs; Section III.B 'we include baseline results from previous research [6]' and Section II 'follow the same experimental design'.
domain assumption Macro F1 averaged over five prompt variants is a stable measure of mental health understanding.
Section II.B describes five query variants averaged per task; no variance or significance is reported.
domain assumption The six public Reddit datasets are valid proxies for mental health understanding tasks.
Section II.A lists Dreaddit, DepSeverity, SDCNL, and CSSRS-Suicide as the task sources; no validity argument beyond prior publication.
domain assumption Zero-shot F1 reflects knowledge internalized during pre-training.
Introduction states zero-shot ability 'indicates that the model has internalized relevant knowledge during pre-training' without supporting evidence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beyond Scale: Small Language Models are Comparable to GPT-4 in Mental Health Understanding." pith.science (2026). https://pith.science/paper/NES2NJOU

@misc{pith2026250708031,
  author       = {Pith},
  title        = {Pith review of: Beyond Scale: Small Language Models are Comparable to GPT-4 in Mental Health Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NES2NJOU}},
  note         = {Machine review of arXiv:2507.08031}
}

read the original abstract

The emergence of Small Language Models (SLMs) as privacy-preserving alternatives for sensitive applications raises a fundamental question about their inherent understanding capabilities compared to Large Language Models (LLMs). This paper investigates the mental health understanding capabilities of current SLMs through systematic evaluation across diverse classification tasks. Employing zero-shot and few-shot learning paradigms, we benchmark their performance against established LLM baselines to elucidate their relative strengths and limitations in this critical domain. We assess five state-of-the-art SLMs (Phi-3, Phi-3.5, Qwen2.5, Llama-3.2, Gemma2) against three LLMs (GPT-4, FLAN-T5-XXL, Alpaca-7B) on six mental health understanding tasks. Our findings reveal that SLMs achieve mean performance within 2\% of LLMs on binary classification tasks (F1 scores of 0.64 vs 0.66 in zero-shot settings), demonstrating notable competence despite orders of magnitude fewer parameters. Both model categories experience similar degradation on multi-class severity tasks (a drop of over 30\%), suggesting that nuanced clinical understanding challenges transcend model scale. Few-shot prompting provides substantial improvements for SLMs (up to 14.6\%), while LLM gains are more variable. Our work highlights the potential of SLMs in mental health understanding, showing they can be effective privacy-preserving tools for analyzing sensitive online text data. In particular, their ability to quickly adapt and specialize with minimal data through few-shot learning positions them as promising candidates for scalable mental health screening tools.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

HealthSLM-Bench: Benchmarking Small Language Models for Mobile and Wearable Healthcare Monitoring
cs.AI 2025-09 conditional novelty 5.0 of 10

Small language models rival large ones on several wearable health prediction tasks, with large efficiency gains, but suffer from class imbalance and poor calorie regression.

Reference graph

Works this paper leans on

26 extracted references · 20 canonical work pages · cited by 1 Pith paper

[6]

Mental-llm: Leveraging large language models for mental health prediction via online text data,

X. Xu, et al., “Mental-llm: Leveraging large language models for mental health prediction via online text data,” Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 8, no. 1, pp. 1-32, 2024

work page 2024
[4]

Social media and mental health: benefits, risks, and opportunities for research and practice,

J. A. Naslund, et al., “Social media and mental health: benefits, risks, and opportunities for research and practice,” J. Technol. Behav. Sci., vol. 5, pp. 245-257, 2020

work page 2020
[1]

The global economic burden of noncommunicable diseases,

D. E. Bloom, et al., “The global economic burden of noncommunicable diseases,” World Economic Forum, 2012

work page 2012
[2]

The State of Mental Health in America 2024,

M. Reinert, et al., “The State of Mental Health in America 2024,” Mental Health America, 2024

work page 2024
[3]

How mental health care should change as a consequence of the COVID-19 pandemic,

C. Moreno, et al., “How mental health care should change as a consequence of the COVID-19 pandemic,” The Lancet Psychiatry, vol. 7, no. 9, pp. 813-824, 2020

work page 2020
[5]

Dang, et al

T. Dang, et al. ”Multimodal Large Language Models in Human-Centered Health: Practical Insights.” IEEE Pervasive Computing 23.4 (2025): 87- 93

work page 2025
[7]

A taxonomy of ethical tensions in inferring mental health states from social media,

S. Chancellar, et al., “A taxonomy of ethical tensions in inferring mental health states from social media,” in Proc. FAT, 2019, pp. 79-88

work page 2019
[8]

Beyond LDA: exploring supervised topic modeling for depression-related language in Twitter,

P. Resnik, et al., “Beyond LDA: exploring supervised topic modeling for depression-related language in Twitter,” in Proc. CLPsych, 2015, pp. 99-107

work page 2015

Show all 26 references

[9]

AER-LLM: Ambiguity-aware emotion recognition leveraging large language models,

X. Hong, et al., “AER-LLM: Ambiguity-aware emotion recognition leveraging large language models,” in ICASSP 2025, pp. 1–5

2025
[10]

”Token-Level Logits Matter: A Closer Look at Speech Foundation Models for Ambiguous Emotion Recognition.” arXiv preprint arXiv:2505.18484 (2025)

Halim, Jule Valendo, et al. ”Token-Level Logits Matter: A Closer Look at Speech Foundation Models for Ambiguous Emotion Recognition.” arXiv preprint arXiv:2505.18484 (2025)

2025 arXiv
[11]

Privacy-preserving deep learning,

R. Shokri, et al., “Privacy-preserving deep learning,” in Proc. CCS, 2015, pp. 1310-1321

2015
[12]

”Efficient and personalized mobile health event prediction via small language models.” Proceedings of the 30th Annual International Conference on Mobile Computing and Networking

Wang, Xin, et al. ”Efficient and personalized mobile health event prediction via small language models.” Proceedings of the 30th Annual International Conference on Mobile Computing and Networking. 2024

2024
[13]

Phi-3 technical report: A highly capable language model locally on your phone,

M. Abdin, et al., “Phi-3 technical report: A highly capable language model locally on your phone,” arXiv preprint arXiv:2404.14219, 2024

2024 arXiv
[14]

Gemma: Open models based on gemini research and technology,

Gemma Team, et al., “Gemma: Open models based on gemini research and technology,” arXiv preprint arXiv:2403.08295, 2024

2024 arXiv
[15]

Federated learning for mobile keyboard prediction,

A. Hard, et al., “Federated learning for mobile keyboard prediction,” arXiv preprint arXiv:1811.03604, 2018

2018 arXiv
[16]

A primer on neural network models for natural language processing,

A. Rogers, et al., “A primer on neural network models for natural language processing,” J. Artif. Intell. Res., vol. 57, pp. 345-420, 2016

2016
[17]

On the dangers of stochastic parrots: Can language models be too big?,

E. Bender, et al., “On the dangers of stochastic parrots: Can language models be too big?,” in Proc. FAccT, 2021, pp. 610-623

2021
[18]

A discourse-aware attention model for abstractive summarization of long documents,

A. Cohan, et al., “A discourse-aware attention model for abstractive summarization of long documents,” in Proc. NAACL-HLT, 2018, pp. 615-621

2018
[19]

Language models are few-shot learners,

T. Brown, et al., “Language models are few-shot learners,” in Proc. NeurIPS, 2020, pp. 1877-1901

2020
[20]

CLPsych 2019 shared task: Predicting the degree of suicide risk in Reddit posts,

K. Zirikly, et al., “CLPsych 2019 shared task: Predicting the degree of suicide risk in Reddit posts,” in Proc. CLPsych, 2019, pp. 24-33

2019
[21]

Dreaddit: A reddit dataset for stress analysis in social media,

E. Turcan and K. McKeown, “Dreaddit: A reddit dataset for stress analysis in social media,” arXiv preprint arXiv:1911.00133, 2019

1911 arXiv
[22]

Early Detection of Depression Severity Levels on Reddit using a Multi-aspect-based Model,

U. Naseem, et al., “Early Detection of Depression Severity Levels on Reddit using a Multi-aspect-based Model,” in Proc. ECIR, 2022, pp. 245-252

2022
[23]

Deep Learning for Suicidal Ideation Detection and Classification from Social Media Texts,

A. Haque, et al., “Deep Learning for Suicidal Ideation Detection and Classification from Social Media Texts,” in Proc. IEEE Big Data, 2021, pp. 2781-2790

2021
[24]

Knowledge-aware assessment of severity of suicide risk for early intervention,

M. Gaur, et al., “Knowledge-aware assessment of severity of suicide risk for early intervention,” in Proc. WWW, 2019, pp. 514-525

2019
[25]

Xu, et al

J. Xu, et al. ”Qwen2. 5-omni technical report.” arXiv preprint arXiv:2503.20215 (2025)

2025 arXiv
[26]

Grattafiori, et al

A. Grattafiori, et al. ”The llama 3 herd of models.” arXiv preprint arXiv:2407.21783 (2024)

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [6]

Mental-llm: Leveraging large language models for mental health prediction via online text data,

X. Xu, et al., “Mental-llm: Leveraging large language models for mental health prediction via online text data,” Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 8, no. 1, pp. 1-32, 2024

work page 2024

[2] [4]

Social media and mental health: benefits, risks, and opportunities for research and practice,

J. A. Naslund, et al., “Social media and mental health: benefits, risks, and opportunities for research and practice,” J. Technol. Behav. Sci., vol. 5, pp. 245-257, 2020

work page 2020

[3] [1]

The global economic burden of noncommunicable diseases,

D. E. Bloom, et al., “The global economic burden of noncommunicable diseases,” World Economic Forum, 2012

work page 2012

[4] [2]

The State of Mental Health in America 2024,

M. Reinert, et al., “The State of Mental Health in America 2024,” Mental Health America, 2024

work page 2024

[5] [3]

How mental health care should change as a consequence of the COVID-19 pandemic,

C. Moreno, et al., “How mental health care should change as a consequence of the COVID-19 pandemic,” The Lancet Psychiatry, vol. 7, no. 9, pp. 813-824, 2020

work page 2020

[6] [5]

Dang, et al

T. Dang, et al. ”Multimodal Large Language Models in Human-Centered Health: Practical Insights.” IEEE Pervasive Computing 23.4 (2025): 87- 93

work page 2025

[7] [7]

A taxonomy of ethical tensions in inferring mental health states from social media,

S. Chancellar, et al., “A taxonomy of ethical tensions in inferring mental health states from social media,” in Proc. FAT, 2019, pp. 79-88

work page 2019

[8] [8]

Beyond LDA: exploring supervised topic modeling for depression-related language in Twitter,

P. Resnik, et al., “Beyond LDA: exploring supervised topic modeling for depression-related language in Twitter,” in Proc. CLPsych, 2015, pp. 99-107

work page 2015