DepressionX: Knowledge Infused Residual Attention for Explainable Depression Severity Assessment

Tarique Anwar; Tommy Yuan; Yusif Ibrahimov

REVIEW 4 major objections 5 minor 50 references

DepressionX: Knowledge Infused Residual Attention for Explainable Depression Severity Assessment

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read DepressionX, a knowledge-infused residual-attention model, classifies Reddit posts into four depression-severity levels, reporting weighted F1 of 82.5% and 90.9% on imbalanced and balanced datasets, over 7% above state-of-the-art…

desk verdict A real architectural novelty, but the D2 result likely relies on a knowledge graph built from the full evaluation corpus, so the '>7% on both datasets' claim is not yet established. read the letter →

arxiv 2501.14985 v1 pith:CTV6YFPD submitted 2025-01-24 cs.LG

Yusif Ibrahimov , Tarique Anwar , Tommy Yuan This is my paper · ORCID

classification cs.LG

keywords depressionseveritydetectionexplainableAIknowledgegraphresidualmulti-headattentionordinalregressionsocialmediamentalhealthRedditneuralnetworks

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 introduces DepressionX, a model that classifies Reddit posts into four depression-severity levels (minimal, mild, moderate, severe) and explains each prediction. The model combines word-, sentence-, and post-level text embeddings with a depression-specific knowledge graph built from Wikipedia entities filtered by their similarity to BDI-II symptoms. On two Reddit benchmarks, one imbalanced and one balanced, the paper reports weighted F1 scores of 82.5% and 90.9%, respectively, exceeding state-of-the-art comparison models by more than 7%. The authors argue that explainability is the missing piece in existing deep-learning depression detectors, and they supply word-, sentence-, and graph-edge attention as the explanation mechanism.

What carries the argument

The carrying object is a two-stream architecture: a residual multi-head attention stack over text, and a knowledge-graph encoder over a depression-specific graph. Text is encoded at three granularities (FastText for words, SentenceTransformer for sentences and posts), and residual multi-head attention produces a post vector while emitting word- and sentence-level attention weights. In parallel, Wikipedia entities and relations extracted by REBEL are filtered by cosine similarity greater than 0.5 to BDI-II symptoms, embedded with SentenceTransformer, and processed by two GIN layers plus a GAT layer to yield a graph vector. Concatenating the text and graph vectors and applying an ordinal soft-label loss ties the two streams together; the same attention weights and a top-K edge subgraph supply the explanation.

What would settle it

Rerun the D1 experiment with the knowledge-graph module removed, keeping all hyperparameters and the 10-run protocol; the paper reports word-level-only F1 of 0.840 against the full model's 0.825, so a persistent, statistically significant gap would falsify the claim that knowledge infusion helps across datasets.

Watch

Extended reading notes

Core claim

The central claim is that a residual multi-head attention network over multi-level text encodings, infused with a domain knowledge graph, can simultaneously reach higher severity-classification accuracy than existing models and provide transparent rationales. Concretely, DepressionX fuses FastText word vectors, SentenceTransformer sentence and post embeddings, and a graph representation learned by two GIN layers plus one GAT layer from a Wikipedia-derived depression knowledge graph; the fused vector is passed through a feed-forward network trained with an ordinal soft-label cross-entropy loss. The paper reports weighted F1 of 0.825 on D1 and 0.909 on D2, beating the strongest prior baselines by over 7% on both, and uses attention heatmaps plus a top-K edge subgraph to show which words and knowledge-graph relationships influenced each decision.

Load-bearing premise

The load-bearing premise is that a knowledge graph assembled by keeping Wikipedia entities whose SentenceTransformer embeddings have cosine similarity above 0.5 to BDI-II symptoms genuinely helps severity classification; the paper's own ablation shows this help on the balanced dataset but a small decline on the imbalanced one relative to word-level text alone.

Editorial extensions

If this is right

Depression severity can be read off Reddit text at four ordered levels with weighted F1 of 0.825 and 0.909 on imbalanced and balanced benchmarks, respectively.
The reported margin of more than 7% over prior models suggests that combining multi-granularity text attention with domain knowledge is a viable direction for severity detection rather than binary detection.
Predictions come with inspectable word- and sentence-level attention plus a small knowledge subgraph, making the model more suited to healthcare-facing use where transparency matters.
The ordinal soft-label loss encodes the ordering of severity levels, so misclassifications between adjacent levels are penalized less than far-off errors.

Reading between the lines

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

Editorial extension: the ablation pattern suggests the multi-level residual attention, not the knowledge graph, may drive most of the gain, since word-level-only text already beats the full model on the imbalanced dataset D1.
Editorial extension: a testable extension is to gate knowledge infusion by post length or by the confidence of the text stream, which might recover the D1 gap while keeping the D2 gain.
Editorial extension: the same architecture could transfer to other mental-health outcomes, but only if the knowledge graph is rebuilt from a symptom lexicon appropriate to that condition.
Editorial extension: replacing the fixed 0.5 cosine-similarity threshold with a learned or per-entity threshold could make the knowledge component more robust across datasets.

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

Two things to know. First, the architecture is a genuine new combination: multi-level text encoding (FastText, Sentence Transformer) plus a Wikipedia/BDI-II-derived knowledge graph, processed with GIN/GAT, then ordinal regression, with attention-based explanation. That's worth taking seriously. Second, the headline claim—over 7% weighted F1 gain on both datasets—is not established, because the knowledge graph on the balanced dataset D2 is built from that exact dataset's full corpus, and the paper never says construction used only the training split.

Let me be specific. The KG section says they feed the Sampath and Durairaj dataset (which is D2) to REBEL for entity extraction. The only data-hygiene guarantee in the paper is that hyperparameter tuning used train/validation splits. That does not cover KG construction. So the fixed graph representation g likely encodes information from test posts. The ablation is exactly what you'd expect under leakage: adding the KG helps on D2 (0.885→0.909) but hurts on D1 (0.835→0.825). That pattern is too clean to ignore. Without the KG, D2 F1 is 0.885, which is still +5.0 points over MuLHiTA (0.835), not the claimed >7. The D1 >7% claim survives because the textual stack alone reaches 0.835 there. So the paper's central claim overstates what is currently shown.

Credit where due. The authors report 10-run medians with standard deviations, include many baselines, provide code and data, give an ablation, and got ethics approval. The explainability section is honest that they only show attention heatmaps and a subgraph—no human evaluation, no faithfulness metric, and the UMAP before/after figure is not backed by the F1 numbers. That's a modest gap, not fatal.

Two smaller concerns. First, BERT and MentalBERT show standard deviations of 0.001, which suggests they did not get the same Optuna treatment as the proposed model; the comparison isn't apples-to-apples. Second, the soft-label ordinal regression has a tuned severity scale beta=3.0, which is fine, but the absolute performance should be read with that in mind.

Bottom line: the work is a plausible contribution to social-media mental-health screening, and the textual-only part already beats the baselines on D1. But the D2 result and the ">7% on both" claim are in question until the KG is rebuilt from training data only and the experiments rerun. That's a major revision, not a rejection. I'd send it to reviewers who know graph-based leakage risks, and ask specifically for the train-only KG rerun and a real explainability check.

Referee Report

4 major / 5 minor

Summary. The paper proposes DepressionX, a model for explainable depression severity classification from social media posts. It combines multi-level textual encodings (FastText word embeddings, SentenceTransformer sentence/post embeddings) with a residual multi-head attention module, and infuses a domain-specific knowledge graph constructed by extracting Wikipedia entities/relations from depression-related datasets using REBEL and filtering by cosine similarity to BDI-II symptoms. The graph is encoded with two GIN layers and one GAT layer, and its pooled representation is concatenated with the text representation for ordinal-regression-based soft-label classification. Experiments on two Reddit datasets (D1, imbalanced; D2, balanced) report weighted F1 scores of 0.825 and 0.909, exceeding the best baselines by more than 7%. The paper also presents ablations, attention-based text explanations, a knowledge subgraph explanation, and parametric analyses.

Significance. If the reported results are sound, DepressionX would be a useful contribution to explainable mental-health NLP: it targets severity levels rather than binary detection, provides both textual and graph-level explanations, and reports systematic comparisons against 11 baselines with 10-run medians. The public code release and the explicit statement that hyperparameter tuning used only training/validation splits are positive reproducibility features. However, the central claim of consistent >7% improvement depends critically on the D2 results, and the knowledge graph construction appears to use the full D2 dataset, creating a potential transductive leakage that the ablation pattern is consistent with. The comparison fairness for BERT/MentalBERT is also questionable. These issues are load-bearing and need to be resolved before the main claim can be accepted.

major comments (4)

[Domain-specific knowledge graph construction] The knowledge graph is built by feeding "depression-related datasets (Sampath and Durairaj 2022)" to REBEL, and D2 is exactly the Sampath and Durairaj dataset. The paper states only that hyperparameter optimization used training/validation splits; it never states that KG construction was restricted to the training split. If the full D2 corpus, including test posts, was used for entity extraction, the learned graph representation g encodes aggregate information from the test set, which is a form of transductive contamination. This concern is reinforced by the ablation: adding the KG improves D2 from 0.885 to 0.909 (+0.024) but reduces D1 from 0.835 to 0.825 (-0.010). The authors must clarify whether KG construction used only the training portions of D1/D2, and if it did not, they must rerun all experiments with a train-only KG and report whether the >7% gains on D2 persist.
[Experimental Settings / Table 1] The reported standard deviations for BERT and MentalBERT are exactly 0.001 for every metric on both datasets, which is implausibly small compared with the other baselines (e.g., LSTM has deviations of 0.02-0.09). This suggests that these transformer baselines may not have received the same Optuna tuning as the proposed model, or that the runs were not conducted under comparable conditions. The authors should specify the exact hyperparameter search space and the number of Optuna trials used for each baseline, and report whether identical tuning budgets were used. Without this, the relative ranking of DepressionX against BERT/MentalBERT is not fairly established.
[Ablation study, Table 2] Table 2 shows that the full DepressionX model obtains a weighted F1 of 0.825 on D1, which is lower than the word-level-only Block 1 (0.840) and lower than Block 1+2+3 without the KG (0.835). The text nonetheless states that the full model "shows the best overall performance" and attributes gains to domain knowledge. The authors should explicitly acknowledge that the knowledge graph hurts D1 performance, discuss why the KG benefit is dataset-specific, and tone down the claim of consistent improvement. This is particularly important because the D2-specific benefit is the main piece of evidence for the knowledge graph's value.
[Experimental Settings / Proposed model] The hyperparameter list and the parametric analysis (Figure 7b) include a "contrastive loss temperature" with an optimal value of 0.5, but the proposed model description in Section "Proposed model" does not define any contrastive loss. It is unclear whether this temperature applies to some unstated auxiliary objective, to the SentenceTransformer training, or to the final classification loss. The authors should specify the loss function in which this temperature is used, otherwise the experimental setup is not fully reproducible.

minor comments (5)

[Equation (14)] Equation (14) defines the cost as beta * |r_t - r_i|, but the notation r_t is not introduced; it should presumably be r_p (the actual severity of the post). Please correct this typo.
[Throughout] Several cross-references to sections and equations are empty in the arXiv version (e.g., "Section ", "Equation 7", "Section "). The paper should be compiled with proper labels so that all references resolve.
[Problem statement] The label set is introduced as Y = {minimum = 0, mild = 1, moderate = 2, severe = 3}, but the severity class in the problem statement is named "minimal", not "minimum". Please use consistent terminology.
[Experimental Results, D2 description] In the D2 dataset description, "non-depressed" and "non-depressed/minimum" are used interchangeably across datasets; clarifying whether "non-depressed" corresponds to the "minimal" class in D2 would improve readability.
[Explainability analysis] The explanatory subgraph approach is described as maximizing mutual information with the original graph, but the actual optimization minimizes Smooth L1 loss between representations. Please clarify the relationship between the stated objective and the implemented loss.

Circularity Check

1 steps flagged · score 6.0 of 10

The D2 knowledge-graph gain is not independent because the KG is built from the D2 corpus itself; aside from this self-input issue, the derivation is empirical and non-circular.

fitted input called prediction [Knowledge Graph Representation (Eqs. 10-12), Datasets, Experimental Settings]
"We build our depression KG by feeding depression-related datasets (Sampath and Durairaj 2022) to REBEL, extracting entities, relationships, hyperlinks, and summaries from Wikipedia data. ... D2 (Sampath and Durairaj 2022): Comprising 1985, 1000, and 902 posts in the non-depressed, moderate, and severe categories, respectively. ... All hyperparameter optimisation was performed exclusively on the training and validation splits of the datasets, with no access to the test data."

D2 is exactly the Sampath-Durairaj corpus named as the KG source, and the only stated data-hygiene restriction covers hyperparameter optimisation, not KG construction. The fixed representation g (Eq. 10) can therefore encode statistics from the full D2 corpus, including test posts. Since each prediction is f(pi)=argmax softmax(FFN(p'i ⊕ g)), D2 test posts are classified using a graph built from the same corpus. The ablation matches this: KG helps D2 (0.885→0.909) but hurts D1 (0.835→0.825), so the advertised D2 gain over MuLHiTA (0.909 vs 0.835) is at least partly an evaluation on the KG's construction input rather than a transfer of external knowledge. This is leakage/self-input rather than label fitting, but it makes the reported balanced-dataset prediction non-independent.

full rationale

Aside from the D2/KG self-input issue, the model's derivation is empirically grounded: textual encoders, residual multi-head attention, GIN/GAT, and ordinal soft-label loss are standard components with no self-referential definitions, and the explainability analysis is post hoc (attention heatmaps and a top-K subgraph) rather than an input to the prediction. Self-citations such as Ibrahimov, Anwar, and Yuan (2024) motivate severity levels and transparency but are not load-bearing mathematical support. The ordinal-regression scale β=3.0 is a tuned loss design choice, not a circular prediction. The one serious problem is the unstated KG split: if KG construction were restricted to the training split, the paper would have no circularity and the score would be 0. On the text as written, the balanced-dataset result is not fully independent, which supports the partial-circularity score.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim depends on only a few hand-picked settings: the cosine similarity threshold for knowledge graph construction, the severity scale in the ordinal loss, and the contrastive loss temperature. The model also relies on domain assumptions about the validity of BDI-II based filtering and about the faithfulness of attention as explanation. No new physical or conceptual entities are invented. Overall the account is empirically driven rather than derived from first principles.

free parameters (3)

cosine similarity threshold = 0.5
Retains knowledge graph entities whose embedding similarity to BDI-II symptoms exceeds 0.5; chosen by hand, directly determines the size and content of the depression knowledge graph.
severity scale beta = 3.0
Controls the penalty magnitude in the ordinal regression soft-label distribution (Equation 14); tuned via Optuna.
contrastive loss temperature = 0.5
Temperature in the contrastive loss used when fine-tuning Sentence Transformer embeddings; tuned via Optuna.

assumptions (3)

domain assumption BDI-II symptom similarity is a valid filter for depression-related entities.
The knowledge graph retains only entities with cosine similarity greater than 0.5 to BDI-II symptoms; this assumes BDI-II adequately represents the depression concept space.
domain assumption Attention weights and edge-attention scores correspond to the model's actual reasoning.
The explainability analysis relies on this assumption, but it is not validated with human or quantitative faithfulness evaluation.
standard math GIN and GAT architectures learn useful graph representations from node embeddings generated by SentenceTransformer.
This is standard graph neural network background, but the specific effectiveness on this knowledge graph is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DepressionX: Knowledge Infused Residual Attention for Explainable Depression Severity Assessment." pith.science (2026). https://pith.science/paper/CTV6YFPD

@misc{pith2026250114985,
  author       = {Pith},
  title        = {Pith review of: DepressionX: Knowledge Infused Residual Attention for Explainable Depression Severity Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CTV6YFPD}},
  note         = {Machine review of arXiv:2501.14985}
}

abstract

In today's interconnected society, social media platforms have become an important part of our lives, where individuals virtually express their thoughts, emotions, and moods. These expressions offer valuable insights into their mental health. This paper explores the use of platforms like Facebook, $\mathbb{X}$ (formerly Twitter), and Reddit for mental health assessments. We propose a domain knowledge-infused residual attention model called DepressionX for explainable depression severity detection. Existing deep learning models on this problem have shown considerable performance, but they often lack transparency in their decision-making processes. In healthcare, where decisions are critical, the need for explainability is crucial. In our model, we address the critical gap by focusing on the explainability of depression severity detection while aiming for a high performance accuracy. In addition to being explainable, our model consistently outperforms the state-of-the-art models by over 7% in terms of $\text{F}_1$ score on balanced as well as imbalanced datasets. Our ultimate goal is to establish a foundation for trustworthy and comprehensible analysis of mental disorders via social media.

Figures

Figures reproduced from arXiv: 2501.14985 by the authors.

**Figure 1.** Proposed model DEPRESSIONX. It predicts depression severity of a post and generates explanation with word- and sentence-level attentions and a knowledge subgraph. The prediction and the explanatory components are highlighted in yellow. post pi is represented as SentenceTransformer(si,j ) ∈ R 768 , ∀i ∈ {1, N}, ∀j ∈ {1, L}, and our post-level encoding pi is represented as SentenceTransformer(pi) ∈ R 768 , ∀i ∈ {1, N… view at source ↗

**Figure 2.** Our domain-specific knowledge graph Knowledge graphs (KGs) are powerful in many retrieval applications by integrating external knowledge bases. Our method for depression severity detection involves constructing a domain-specific KG with relevant depression-related information and relationships. The KG construction starts with entity extraction using the Relation Extraction By Endto-end Language Generation (REBEL) … view at source ↗

**Figure 3.** Two-dimensional plots of post representations [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Explainability with sentence- and word-level im [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Attention heatmap of the most and the least important words [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Explanatory knowledge subgraph cific properties of the dataset, such as the distribution of labels. We observe that adjustments to the number of heads in the GAT architecture have a lesser effect compared to the multi-head attention model. Increasing the number of hea…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 36 canonical work pages

[1]

K.; Shatte, A.; Liu, C.; and Sukunesan, S

Abuhassan, M.; Anwar, T.; Fuller-Tyszkiewicz, M.; Jarman, H. K.; Shatte, A.; Liu, C.; and Sukunesan, S. 2023 a . Classification of Twitter users with eating disorder engagement: Learning from the biographies. Computers in Human Behavior, 140: 107519

work page 2023
[2]

K.; and Fuller-Tyszkiewicz, M

Abuhassan, M.; Anwar, T.; Liu, C.; Jarman, H. K.; and Fuller-Tyszkiewicz, M. 2023 b . EDNet: Attention-Based Multimodal Representation for Classification of Twitter Users Related to Eating Disorders. In Proceedings of the ACM Web Conference 2023, 4065--4074

work page 2023
[3]

American Psychiatric Association, D.; American Psychiatric Association, D.; et al. 2013. Diagnostic and statistical manual of mental disorders: DSM-5, volume 5. American psychiatric association Washington, DC

work page 2013
[4]

Amini, H.; and Kosseim, L. 2020. Towards explainability in using deep learning for the detection of anorexia in social media. In Natural Language Processing and Information Systems: 25th International Conference on Applications of Natural Language to Information Systems, NLDB 2020, Saarbr \"u cken, Germany, June 24--26, 2020, Proceedings 25 , 225--235. Springer

work page 2020
[5]

Anwar, T.; Nepal, S.; Paris, C.; Yang, J.; Wu, J.; and Sheng, Q. Z. 2022. Tracking the evolution of clusters in social media streams. IEEE Transactions on Big Data, 9(2): 701--715

work page 2022
[6]

K.; Chen, P.-Y.; Dhurandhar, A.; Hind, M.; Hoffman, S

Arya, V.; Bellamy, R. K.; Chen, P.-Y.; Dhurandhar, A.; Hind, M.; Hoffman, S. C.; Houde, S.; Liao, Q. V.; Luss, R.; Mojsilovi \'c , A.; et al. 2019. One explanation does not fit all: A toolkit and taxonomy of ai explainability techniques. arXiv preprint arXiv:1909.03012

arXiv 2019
[7]

T.; Steer, R

Beck, A. T.; Steer, R. A.; Ball, R.; and Ranieri, W. F. 1996. Comparison of Beck Depression Inventories-IA and-II in psychiatric outpatients. Journal of Personality Assessment, 67(3): 588--597

work page 1996
[8]

Bojanowski, P.; Grave, E.; Joulin, A.; and Mikolov, T. 2017. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5: 135--146

work page 2017

Show all 50 references

[9]

a ckinger, E.; and Shah, R. 1993. Signature verification using a

Bromley, J.; Guyon, I.; LeCun, Y.; S \"a ckinger, E.; and Shah, R. 1993. Signature verification using a" siamese" time delay neural network. Advances in neural information processing systems, 6

1993
[10]

H.; and Navigli, R

Cabot, P.-L. H.; and Navigli, R. 2021. REBEL: Relation extraction by end-to-end language generation. In Findings of the Association for Computational Linguistics: EMNLP 2021, 2370--2381

2021
[11]

Chung, J.; Gulcehre, C.; Cho, K.; and Bengio, Y. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555

2014 arXiv
[12]

De Choudhury, M.; Gamon, M.; Counts, S.; and Horvitz, E. 2013. Predicting depression via social media. In Proceedings of the international AAAI conference on web and social media, volume 7, 128--137

2013
[13]

Devlin, J. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

2018 arXiv
[14]

Galassi, A.; Lippi, M.; and Torroni, P. 2020. Attention in natural language processing. IEEE transactions on neural networks and learning systems, 32(10): 4291--4308

2020
[15]

Ghosh, S.; and Anwar, T. 2021. Depression intensity estimation via social media: A deep learning approach. IEEE Transactions on Computational Social Systems, 8(6): 1465--1474

2021
[16]

Han, S.; Mao, R.; and Cambria, E. 2022. Hierarchical Attention Network for Explainable Depression Detection on T witter Aided by Metaphor Concept Mappings. In Proceedings of the 29th International Conference on Computational Linguistics, 94--104. Gyeongju, Republic of Korea: I...

2022
[17]

Hochreiter, S. 1997. Long Short-term Memory. Neural Computation MIT-Press

1997
[18]

Ibrahimov, Y.; Anwar, T.; and Yuan, T. 2024. Explainable AI for Mental Disorder Detection via Social Media: A survey and outlook. arXiv preprint arXiv:2406.05984

2024 arXiv
[19]

S.; and Goldberg, Y

Jacovi, A.; Shalom, O. S.; and Goldberg, Y. 2018. Understanding convolutional neural networks for text classification. arXiv preprint arXiv:1809.08037

2018 arXiv
[20]

Ji, S.; Zhang, T.; Ansari, L.; Fu, J.; Tiwari, P.; and Cambria, E. 2022. M ental BERT : Publicly Available Pretrained Language Models for Mental Healthcare. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, 7184--7190. Marseille, France: European L...

2022
[21]

Jia, Y.; McDermid, J.; Lawton, T.; and Habli, I. 2022. The role of explainability in assuring safety of machine learning in healthcare. IEEE Transactions on Emerging Topics in Computing, 10(4): 1746--1760

2022
[22]

Lewis, M.; Liu, Y.; Goyal, N.; Ghazvininejad, M.; Mohamed, A.; Levy, O.; Stoyanov, V.; and Zettlemoyer, L. 2020. BART : Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. In Proceedings of the 58th Annual Meeting of the...

2020
[23]

Liu, S.; Vahedian, F.; Hachen, D.; Lizardo, O.; Poellabauer, C.; Striegel, A.; and Milenkovi \'c , T. 2021. Heterogeneous network approach to predict individuals’ mental health. ACM Transactions on Knowledge Discovery from Data (TKDD), 15(2): 1--26

2021
[24]

Liu, Y. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364

2019 arXiv
[25]

E.; Crestani, F.; and Parapar, J

Losada, D. E.; Crestani, F.; and Parapar, J. 2017. eRISK 2017: CLEF lab on early risk prediction on the internet: experimental foundations. In Experimental IR Meets Multilinguality, Multimodality, and Interaction: 8th International Conference of the CLEF Association, CLEF 2017...

2017
[26]

E.; Crestani, F.; and Parapar, J

Losada, D. E.; Crestani, F.; and Parapar, J. 2018. Overview of eRisk: early risk prediction on the internet. In Experimental IR Meets Multilinguality, Multimodality, and Interaction: 9th International Conference of the CLEF Association, CLEF 2018, Avignon, France, September 10...

2018
[27]

Luo, D.; Cheng, W.; Xu, D.; Yu, W.; Zong, B.; Chen, H.; and Zhang, X. 2020. Parameterized explainer for graph neural network. Advances in neural information processing systems, 33: 19620--19631

2020
[28]

Mihov, I.; Chen, H.; Qin, X.; Ku, W.-S.; Yan, D.; and Liu, Y. 2022. Mentalnet: Heterogeneous graph representation for early depression detection. In 2022 IEEE International Conference on Data Mining (ICDM), 1113--1118. IEEE

2022
[29]

G.; Kim, J.; and Khushi, M

Naseem, U.; Dunn, A. G.; Kim, J.; and Khushi, M. 2022 a . Early identification of depression severity levels on reddit using ordinal classification. In Proceedings of the ACM Web Conference 2022, 2563--2572

2022
[30]

Naseem, U.; Khushi, M.; Kim, J.; and Dunn, A. G. 2022 b . Hybrid text representation for explainable suicide risk identification on social media. IEEE transactions on computational social systems

2022
[31]

Olfson, M.; Blanco, C.; and Marcus, S. C. 2016. Treatment of adult depression in the United States. JAMA internal medicine, 176(10): 1482--1491

2016
[32]

Park, M.; Cha, C.; and Cha, M. 2012. Depressive moods of users portrayed in Twitter. In Proceedings of the 18th ACM International Conference on Knowledge Discovery and Data Mining, SIGKDD 2012, 1--8

2012
[33]

Park, M.; McDonald, D.; and Cha, M. 2013. Perception differences between the depressed and non-depressed users in twitter. In Proceedings of the international AAAI conference on web and social media, volume 7, 476--485

2013
[34]

Pirayesh, J.; Chen, H.; Qin, X.; Ku, W.-S.; and Yan, D. 2021. Mentalspot: Effective early screening for depression based on social contagion. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, 1437--1446

2021
[35]

Reimers, N.; and Gurevych, I. 2019. Sentence- BERT : Sentence Embeddings using S iamese BERT -Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJ...

2019
[36]

Sadeque, F.; Xu, D.; and Bethard, S. 2018. Measuring the latency of depression detection in social media. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, 495--503

2018
[37]

Sampath, K.; and Durairaj, T. 2022. Data set creation and empirical analysis for detecting signs of depression from social media postings. In International Conference on Computational Intelligence in Data Science, 136--151. Springer

2022
[38]

Sawhney, R.; Joshi, H.; Gandhi, S.; and Shah, R. R. 2021. Towards ordinal suicide ideation detection on social media. In Proceedings of the 14th ACM international conference on web search and data mining, 22--30

2021
[39]

Shader, R. I. 2020. COVID-19 and depression. Clinical therapeutics, 42(6): 962--963

2020
[40]

Shen, G.; Jia, J.; Nie, L.; Feng, F.; Zhang, C.; Hu, T.; Chua, T.-S.; Zhu, W.; et al. 2017. Depression detection via harvesting social media: A multimodal dictionary learning solution. In IJCAI, 3838--3844

2017
[41]

Trotzek, M.; Koitka, S.; and Friedrich, C. M. 2018. Utilizing neural networks and linguistic metadata for early detection of depression indications in text sequences. IEEE Transactions on Knowledge and Data Engineering, 32(3): 588--601

2018
[42]

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

2019 arXiv
[43]

Vaswani, A. 2017. Attention is all you need. Advances in Neural Information Processing Systems

2017
[44]

Velickovic, P.; Cucurull, G.; Casanova, A.; Romero, A.; Lio, P.; Bengio, Y.; et al. 2017. Graph attention networks. stat, 1050(20): 10--48550

2017
[45]

WHO. 2024. Depression. [Online; accessed 20-Jan-2024]

2024
[46]

Xia, L.; Feng, Y.; Guo, Z.; Ding, J.; Li, Y.; Li, Y.; Ma, M.; Gan, G.; Xu, Y.; Luo, J.; et al. 2022. MuLHiTA: A novel multiclass classification framework with multibranch LSTM and hierarchical temporal attention for early detection of mental stress. IEEE transactions on neural...

2022
[47]

Xu, K.; Hu, W.; Leskovec, J.; and Jegelka, S. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826

2018 arXiv
[48]

Zogan, H.; Razzak, I.; Jameel, S.; and Xu, G. 2021. Depressionnet: learning multi-modalities with user post summarization for depression detection on social media. In proceedings of the 44th international ACM SIGIR conference on research and development in information retrieva...

2021
[49]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...
[50]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

K.; Shatte, A.; Liu, C.; and Sukunesan, S

Abuhassan, M.; Anwar, T.; Fuller-Tyszkiewicz, M.; Jarman, H. K.; Shatte, A.; Liu, C.; and Sukunesan, S. 2023 a . Classification of Twitter users with eating disorder engagement: Learning from the biographies. Computers in Human Behavior, 140: 107519

work page 2023

[2] [2]

K.; and Fuller-Tyszkiewicz, M

Abuhassan, M.; Anwar, T.; Liu, C.; Jarman, H. K.; and Fuller-Tyszkiewicz, M. 2023 b . EDNet: Attention-Based Multimodal Representation for Classification of Twitter Users Related to Eating Disorders. In Proceedings of the ACM Web Conference 2023, 4065--4074

work page 2023

[3] [3]

American Psychiatric Association, D.; American Psychiatric Association, D.; et al. 2013. Diagnostic and statistical manual of mental disorders: DSM-5, volume 5. American psychiatric association Washington, DC

work page 2013

[4] [4]

Amini, H.; and Kosseim, L. 2020. Towards explainability in using deep learning for the detection of anorexia in social media. In Natural Language Processing and Information Systems: 25th International Conference on Applications of Natural Language to Information Systems, NLDB 2020, Saarbr \"u cken, Germany, June 24--26, 2020, Proceedings 25 , 225--235. Springer

work page 2020

[5] [5]

Anwar, T.; Nepal, S.; Paris, C.; Yang, J.; Wu, J.; and Sheng, Q. Z. 2022. Tracking the evolution of clusters in social media streams. IEEE Transactions on Big Data, 9(2): 701--715

work page 2022

[6] [6]

K.; Chen, P.-Y.; Dhurandhar, A.; Hind, M.; Hoffman, S

Arya, V.; Bellamy, R. K.; Chen, P.-Y.; Dhurandhar, A.; Hind, M.; Hoffman, S. C.; Houde, S.; Liao, Q. V.; Luss, R.; Mojsilovi \'c , A.; et al. 2019. One explanation does not fit all: A toolkit and taxonomy of ai explainability techniques. arXiv preprint arXiv:1909.03012

arXiv 2019

[7] [7]

T.; Steer, R

Beck, A. T.; Steer, R. A.; Ball, R.; and Ranieri, W. F. 1996. Comparison of Beck Depression Inventories-IA and-II in psychiatric outpatients. Journal of Personality Assessment, 67(3): 588--597

work page 1996

[8] [8]

Bojanowski, P.; Grave, E.; Joulin, A.; and Mikolov, T. 2017. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5: 135--146

work page 2017