MATCHA: A Chinese Multi-Disorder Mental Health Dataset with Cross-Lingual Analyses on Social Media

Bingrui Jin; Baihan Li; Zitong Li; Kunyao Lan; Shimeng Li; Kenny Q. Zhu; Mengyue Wu

cs.CLOct 26, 2026

MATCHA: A Chinese Multi-Disorder Mental Health Dataset with Cross-Lingual Analyses on Social Media

Authors: Bingrui Jin, Baihan Li, Zitong Li, Kunyao Lan, Shimeng Li, Kenny Q. Zhu, Mengyue Wu

Organizations: Shanghai Jiao Tong University, Shanghai, China · University of Texas at Arlington, Arlington, TX, USA

Abstract

Mental health research on social media has been dominated by English-language datasets and DSM-5-centered resources, limiting cross-lingual and cross-cultural understanding of online mental health expression. We introduce MATCHA, the first Chinese multi-disorder mental health dataset that links disorder-specific Weibo Super-Topic communities with users' broader public timelines. MATCHA covers seven DSM-5 disorders and a control group, comprising 46,025 cohort-level users and 1.8 million posts. The dataset distinguishes users with explicit self-disclosure evidence from broader disorder-related participants through an LLM-assisted annotation pipeline, and is released in a rigorously de-identified, controlled-access format. Using MATCHA, we analyze behavioral, linguistic, symptom-related, and life-event-related patterns across disorders, and conduct cross-lingual comparisons with Twitter-STMHD. Our analyses show that Weibo Super-Topics function as concentrated spaces for affective and health-related expression, while cross-platform differences between Weibo and Twitter often exceed cross-disorder variation. MATCHA provides a new resource for community-aware and cross-lingual mental health NLP research.

Figures & tables

Figure 1: Overview of the MATCHA Dataset Collection and LLM -assisted Annotation Pipeline.

Figure 2: Composition of MATCHA across disorder and control cohorts. Panel (a) reports cohort-level user counts and Panel (b) reports within-cohort deduplicated post counts. Across the eight cohorts, MATCHA contains 46,025 cohort-level users and 1,834,699 within-cohort unique posts. Because cohorts are not mutually exclusive, these aggregated counts may include users or posts appearing in more than one cohort.

Disorder	IP users [%]	IP posts [%]	STP/IP	STP/non-IP
ADHD	23.5	34.5	7.6	4.4
ASD	8.9	58.9	50.1	3.4
BD	18.8	43.6	7.6	2.3
DEP	19.4	50.3	12.5	3.0
GAD	28.6	61.7	10.3	2.6
OCD	23.3	53.6	10.3	2.7

Table 1: Posting activity comparison between CEP-defined IP and non-IP users within the Super-Topic Post (STP) population across disorders. IP denotes explicit textual self-disclosure evidence under the operational CEP criterion, not clinically verified diagnosis.

Disorder	Evening [%]	$\Delta$ Evening [pp]	Night Ratio
ADHD	$21.9$	$0.5$	$0.50$
ASD	$22.8$	$1.4$	$0.57$
BD	$27.2$	$5.8$	$1.18$
DEP	$27.7$	$6.3$	$1.20$
GAD	$26.8$	$5.4$	$0.72$
OCD	$26.2$	$4.8$	$0.94$

Table 2: User-normalized evening and night-time posting metrics for the full disorder cohorts, computed on the merged STP+HP corpus after note-ID deduplication. Evening is each user’s 20:00–23:00 posting share averaged over users;

\Delta

Evening is relative to the original control cohort; Night Ratio is the 22:00–05:00 share divided by the 10:00–17:00 share.

Panel of Figure 3 — Figure 3: Proxy-signal density, defined as the expected number of classifier-positive labels per post, for CEP-defined IP users’ STP and HP across disorders. The dashed line denotes the corresponding HP-control baseline. Classifier outputs are aggregate proxies rather than clinical assessments.

Figure 4: Cross-lingual Pearson correlation matrices for (a) LIWC feature log ratios, (b) posting hours, (c) life-event distributions, and (d) symptom distributions. Cross-platform symptom correlations in (d) are lower on average than within-platform correlations, indicating substantial platform- and language-associated variation in symptom composition. These correlations are descriptive and do not identify causal cultural effects. Numbers in the upper-right corner of each cell denote statistical significance levels: 1 = $p<0.001$ , 2 = $p<0.01$ , and 3 = $p<0.05$ .

Figure 5: Cross-platform differences in core LIWC features across BD, DEP, GAD, and OCD. Bars show within-platform cohort–control frequency differences (percentage points) for CEP-defined MATCHA IP cohorts and frozen Twitter-STMHD cohorts.

Figure 6: Cross-platform symptom-category distribution differences between MATCHA and Twitter-STMHD. Avoidance is excluded from STMHD due to unstable English classifier behavior.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Cohort	Weibo Super-Topic
DEP	抑郁症超话 ( yiyuzheng chaohua , ‘Depression Super-Topic’)
DEP	抑郁症患者超话 ( yiyuzheng huanzhe chaohua , ‘Depression Patient Super-Topic’)
DEP	抑郁超话 ( yiyu chaohua , ‘Depression Super-Topic’)
SZ	精神分裂症超话 ( jingshen fenliezheng chaohua , ‘Schizophrenia Super-Topic’)
GAD	焦虑症超话 ( jiaolvzheng chaohua , ‘Anxiety Disorder Super-Topic’)
ADHD	多动症超话 ( duodongzheng chaohua , ‘Hyperactivity Disorder Super-Topic’)

Appendix

Table 4: Disorder cohorts and their corresponding public Weibo Super-Topics. Chinese terms are accompanied by pinyin transliterations and English glosses.

	Symptom	Life Event
Number of labels	38	12
Annotated posts	13,723	13,723
Annotation records	22,153	22,153
Median positive rate	2.26%	1.47%
Positive rate range	0.78–16.39%	0.26–43.94%
Most frequent label	Depressed Mood	Health

Appendix

Table 5: Annotation statistics for the Chinese training data. The 22,153 records are shared post–annotator records annotated for both task families; the two columns are not additive.

Task	Macro AUC	Micro AUC	Macro F1
Symptom	0.9374	0.9494	0.5579
Life Event	0.9533	0.9673	0.6449

Appendix

Table 6: Chinese classifier evaluation on a held-out test set (

n=2{,}745

). Per-label decision thresholds were selected on the validation set and applied once to the held-out test set.

Task	Macro AUC	Micro AUC	Macro F1
Symptom	0.9910	0.9942	0.8852
Life Event	0.9769	0.9767	0.7669

Appendix

Table 7: English classifier evaluation results on held-out test sets.

#	Label	Positive (1)	Positive %	Mentioned (0+1)	Mentioned %
1	Irritable Mood	372	2.71	499	3.64
2	Anxious Mood	1,498	10.92	1,835	13.37
3	Autonomic Symptoms	834	6.08	899	6.55
4	Cardiovascular Symptoms	740	5.39	786	5.73
5	Catatonic Behavior	197	1.44	244	1.78
6	Fatigue or Loss of Energy	550	4.01	622	4.53

Appendix

Table 8: Per-label annotation statistics for symptom labels. Positive counts correspond to value 1, while mention counts include both 0 and 1, i.e., instances in which the corresponding category was mentioned in the post.

#	Label	Pos.	Pos. [%]	Mentioned	Mentioned [%]
1	Bereavement	90	0.66	196	1.43
2	Health	6,030	43.94	7,409	53.99
3	Relationship Change	303	2.21	658	4.79
4	Family Member Change	46	0.34	199	1.45
5	Relocation	116	0.85	228	1.66
6	Education	236	1.72	431	3.14

Appendix

Table 9: Per-label annotation statistics for life-event labels.

English Label	Semantic Gloss
Irritable Mood	Persistent irritability, anger, frustration, low emotional tolerance
Anxious Mood	Excessive worry, panic anticipation, fear of catastrophic outcomes
Autonomic Symptoms	Sweating, accelerated heart rate, dry mouth, shortness of breath
Cardiovascular Symptoms	Chest pain, palpitations, cardiovascular discomfort
Catatonic Behavior	Motor rigidity, stupor, mutism, abnormal movement behavior
Decreased Energy Tiredness Fatigue	Exhaustion, low energy, persistent tiredness

Appendix

Table 10: Complete bilingual symptom taxonomy used in MATCHA (Part 1).

English Label	Semantic Gloss
Diminished Emotional Expression	Reduced facial expression, monotone speech, emotional flattening
Do Things Easily Get Painful Consequences	Risk-taking, self-harm, dangerous impulsive behavior
Drastical Shift in Mood and Energy	Abrupt changes in mood and energy levels
Fear About Social Situations	Fear or avoidance of conversations, strangers, or public performance
Fear of Gaining Weight	Fear of weight gain, body-image disturbance, restrictive eating
Fears of Being Negatively Evaluated	Fear of criticism, embarrassment, or social judgment

Appendix

Table 11: Complete symptom taxonomy used in MATCHA (Part 2).

Category	Representative Sub-events
Health	Hospitalization, medication use, surgery, self-harm, recovery from addiction
Relationship Changes	Breakup, divorce, marriage, family conflict, parenting difficulties
Career	Job loss, promotion, workplace discrimination, unemployment
Education	Examinations, school transition, academic stress, dropping out
Financial	Debt, bankruptcy, major financial difficulty, home purchase
Relocation	Moving cities, immigration, homelessness, leaving parents’ home

Appendix

Table 12: Representative life-event categories in MATCHA.

Source	Posts	Dis. [%]	Clin. [%]	Either [%]
HP-control	153,704	0.17	0.42	0.57
STP	169,508	13.88	8.22	19.74
HP	1,559,045	3.72	1.31	4.64

Appendix

Table 13: Post-level prevalence of explicit disorder and clinical cues across sources.

Panel A: Original versus fully masked text
Source	F1 orig	F1 masked	$\Delta$ F1	AUROC orig	AUROC masked
STP	$0.953\pm 0.006$	$0.936\pm 0.007$	$-0.017$	0.995	0.992
HP	$0.842\pm 0.012$	$0.833\pm 0.012$	$-0.009$	0.932	0.927
Panel B: Tiered cue masking
Condition	A	B	C	$\Delta$ (A–B)	$\Delta$ (A–C)
HP-only	$0.842\pm 0.012$	$0.833\pm 0.012$	$0.833\pm 0.012$	0.009	0.009

Appendix

Table 14: Cue-masking performance. Panel A compares original and fully masked text; Panel B reports tiered masking results. Values are mean

\pm

standard deviation over five seeds.

	HP			STP
Disorder	Original	Masked	$\Delta$	Original	Masked	$\Delta$
ADHD	0.858	0.847	0.011	0.953	0.910	0.043
ASD	0.874	0.856	0.018	0.967	0.937	0.030
BD	0.850	0.849	0.001	0.952	0.949	0.003
DEP	0.789	0.783	0.006	0.956	0.952	0.004
GAD	0.808	0.805	0.002	0.977	0.972	0.005

Appendix

Table 15: Per-disorder Macro-F1 before and after masking disorder-name and clinical cues.

\Delta

denotes the absolute decrease from the original condition.

Mental Health

Mental health research studies psychological well-being, distress, disorders, their causes, assessment, prevention, and treatment.

See all 164 papers

Mental Health Disorder Detection Beyond Social Media: A Systematic Review of Available Datasets
Jul 3, 2026Sadiya Sayara Chowdhury Puspo, Ana-Maria Bucur, Stevie Chancellor +2Mental Health
George Mason University, USA · Università della Svizzera italiana, Switzerland · University of Minnesota, USA
A Benchmark Suite of Reddit-Derived Datasets for Mental Health Detection
Apr 25, 2026Khalid Hasan, Jamil SaquerMedical Diagnosis Mental Health
Missouri State University, Springfield, MO, USA
Understanding the Sociocultural Dimensions of Mental Health Discourse in Arabic-Language X Communities
Jun 6, 2026Amal Alqahtani, Rana Salama, Mona DiabSocial Media Analysis Mental Health
King Saud University, Riyadh, Saudi Arabia · King Saud University, Riyadh, KSA · Cairo University, Egypt +3

Jul 3, 2026J/K move · Enter open · S save

Mental Health Disorder Detection Beyond Social Media: A Systematic Review of Available Datasets

Sadiya Sayara Chowdhury Puspo, Ana-Maria Bucur, Stevie Chancellor, Özlem Uzuner+1

George Mason University, USA · Università della Svizzera italiana, Switzerland · University of Minnesota, USA

Mental Health

Open paper AskPDF arXiv

Detecting mental health disorders in a timely manner is an important societal challenge. NLP and machine learning (ML) methods used to assist with detection rely on data collected primarily from social media. However, such datasets often have sampling biases and inherent ethical and privacy issues. One avenue to overcome these limitations is non-social media data. We present the first comprehensive review of non-social media, free-text datasets for mental health research. We use the PRISMA methodology to conduct our survey and we review datasets available in multiple languages. We find that non-social media free-text based datasets are predominantly focused on English and on detecting depression. These datasets also vary in demographics, platforms, data types, annotation techniques, and methodologies. This systematic review also reveals key gaps and highlights opportunities to develop more diverse, reliable and clinically-relevant resources.

Abstract

Mental Health

Explore similar work

Mental Health Disorder Detection Beyond Social Media: A Systematic Review of Available Datasets

A Benchmark Suite of Reddit-Derived Datasets for Mental Health Detection

Understanding the Sociocultural Dimensions of Mental Health Discourse in Arabic-Language X Communities