Words Are the Voice of the Mind: Depression Severity Estimation from Social Traces via Mental-Guided Evidence Reasoning

Nada Xu; Xueqin Chen; Qiang Gao; Hongru Liang; Wenqiang Lei

cs.CLOct 24, 2026

Words Are the Voice of the Mind: Depression Severity Estimation from Social Traces via Mental-Guided Evidence Reasoning

Authors: Nada Xu, Xueqin Chen, Qiang Gao, Hongru Liang, Wenqiang Lei

Organizations: Sichuan University · Southwestern University of Finance and Economics

Abstract

Depression severity estimation is crucial for mental health assessment and intervention, yet traditional approaches depend on clinical interviews or self-report questionnaires that require users' active participation. With the widespread use of online social media, users leave rich social traces that may reflect their mental states, motivating the task of social media-based depression severity estimation. Existing efforts have progressed from textual-feature-based classification to LLM-based questionnaire completion over standardized scale items, yet they still fall short in user-specific mental priors modeling, diversified semantic retrieval, and fine-grained reasoning over mental evidence. To address these limitations, we propose UMer, a user-specific mental evidence reasoning framework for social traces-based depression severity estimation. UMer first distills a user-specific mental prior from historical social traces to capture the user's psychological tendencies, behavioral patterns, and self-expression style. It then performs scale-guided multi-query retrieval to construct evidence memories from item-relevant posts. Finally, it conducts coarse-to-fine evidence-aware reasoning to distinguish direct and indirect evidence and predict item-level scores with interpretable rationales. Extensive experiments on real-world benchmarks demonstrate that UMer not only achieves competitive performance over existing baselines, but also provides evidence-grounded interpretability and generalizes to other scale-based tasks.

Figures & tables

Figure 2: Overview of UMer, a three-stage LLM-based agentic framework. The model progressively constructs the user-specific mental prior from historical social traces, retrieves item-relevant posts to construct scale-aware mental evidence memories, and performs coarse-to-fine evidence-aware reasoning for interpretable item-level scoring.

		Questionnaire Metrics		Symptom Metrics
Collection	Model	DCHR % ( $\uparrow$ )	ADODL % ( $\uparrow$ )	AHR % ( $\uparrow$ )	ACR % ( $\uparrow$ )
eRisk 2020	BioInfo $\ddagger$	30.00	76.01	38.30	69.21
	iLab $\ddagger$	27.14	81.70	37.07	69.41
	RELAI $\ddagger$	34.29	83.15	36.39	68.32
	Sense2Vec $\ddagger$	37.14	82.61	38.97	70.10
	Voting $\dagger$	47.14	85.33	35.24	67.41

Table 1: Depression severity estimation on eRisk 2020 and eRisk 2021.

\dagger

: uses manual annotations;

\ddagger

: involves task-specific model training. The best results are highlighted in bold , and the second-best results are underlined .

Figure 3: The score distribution between proprietary LLM and Open-weight LLM

Dataset	Objective Accuracy	Completeness	Subjective Reasonableness
eRisk-Avg	4.53	4.02	4.14

Table 2: Average scores assigned by three human annotators for each mental prior quality metric.

Configuration	MAE ( $\downarrow$ )	Symptom Metrics		Questionnaire Metrics
Configuration	MAE ( $\downarrow$ )	AHR % ( $\uparrow$ )	ACR % ( $\uparrow$ )	ADODL % ( $\uparrow$ )	DCHR % ( $\uparrow$ )
UMer	0.8082	41.97	73.77	84.13	40.00
w/o $\mathcal{S}_{u}$	0.8707	38.57	70.97	81.72	28.57
w/o SMR	0.8361	40.27	72.13	82.56	31.43
w/o EaR	0.8721	39.12	70.93	81.68	38.57

Table 3: Ablation study on eRisk 2020.

Figure 4: Generalization evaluation on eRisk 2022 and eRisk 2023 EDE-Q scoring tasks. Lower values indicate better performance.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Content attention analysis. "Mixed" denotes reasoning that jointly relies on information from both the mental prior and retrieved posts.

Retriever	MAE $\downarrow$	Symptom Metrics		Questionnaire Metrics
Retriever	MAE $\downarrow$	AHR % ( $\uparrow$ )	ACR % ( $\uparrow$ )	ADODL % ( $\uparrow$ )	DCHR % ( $\uparrow$ )
BGE-M3	0.8082	41.97	73.77	84.13	40.00
GTE multilingual base	0.8354	38.91	72.15	83.13	44.29
Qwen3 embedding 0.6B	0.8517	39.66	71.61	82.18	38.29

Appendix

Table A1: Performance of different embedding encoders.

Configuration	MAE $\downarrow$	Symptom Metrics		Questionnaire Metrics
Configuration	MAE $\downarrow$	AHR % ( $\uparrow$ )	ACR % ( $\uparrow$ )	ADODL % ( $\uparrow$ )	DCHR % ( $\uparrow$ )
top-25	0.8082	41.97	73.77	84.13	40.00
top-10	0.8422	39.66	71.93	82.22	30.00
top-15	0.8395	39.46	72.02	82.45	28.57
top-20	0.8388	39.18	72.04	82.93	32.86
top-30	0.8279	39.93	72.40	82.84	34.29

Appendix

Table A2: Performance of different retrieval depths.

Figure A2: Case study for mental prior

Figure A3: Prompt template for constructing the accumulated mental prior from posts.

Figure A4: Prompt template for multi-query construction.

Figure A5: Prompt template for structured reasoning and final item-level self-rating simulation.

Figure A6: Prompt template for key-LLM-based query generation.

Figure A7: Prompt template for Generalization ability Experiment.

Depression Severity Estimation

Depression severity estimation infers symptom levels from clinical interviews, self-reports, speech, text, video, and other behavioral signals.

See all 35 papers

WPG-MoE: Weak-Prior-Guided Dense Mixture-of-Experts for User-Level Social Media Depression Detection
Jul 5, 2026Xian Li, Yuanhe Tian, Yang Yang +2Social Media Analysis Depression Detection
University of Electronic Science and Technology of China · Zhongguancun Academy · University of Science and Technology of China
Fine-tuning LLMs for Passive Depression Severity Estimation from AI Mental Health Dialogue
Jun 16, 2026Olivier Tieleman, Ziyi Zhu, Ting Su +3Depression Detection LLM Fine-Tuning
Depression Risk Assessment in Social Media via Large Language Models
Apr 21, 2026Giorgia Gulino, Manuel PetrucciSocial Media Analysis Depression Detection
1Guglielmo Marconi University, Department of Human Sciences, Rome, Italy

Jul 5, 2026J/K move · Enter open · S save

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

Xian Li, Yuanhe Tian, Yang Yang, Guoqing Wang+1

University of Electronic Science and Technology of China · Zhongguancun Academy · University of Science and Technology of China

Social Media Analysis Depression Detection Mixture-of-Experts Models

Open paper AskPDF arXiv

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.

Abstract

Depression Severity Estimation

Explore similar work

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

Fine-tuning LLMs for Passive Depression Severity Estimation from AI Mental Health Dialogue

Depression Risk Assessment in Social Media via Large Language Models