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

AI summary

Overview

• Research area: Natural Language Processing / computational mental health, with a focus on Chinese social media corpora and cross-lingual comparison.
• Technical level: Intermediate — readers should be comfortable with dataset construction, LLM-assisted annotation, LIWC lexicon features, and standard classification metrics (F1, AUROC, Cohen's kappa).
• Scope: The paper introduces MATCHA, a Chinese multi-disorder Weibo dataset covering seven DSM-5 disorders plus a control group, and uses it for behavioral, linguistic, symptom/life-event, cue-reduced, and cross-lingual analyses against the English Twitter-STMHD dataset.

What This Paper Is About

Research on mental health in social media has been dominated by English-language, DSM-5-centered resources such as SMHD and Twitter-STMHD, which limits understanding of how mental health is expressed in other languages and platform cultures. The authors build MATCHA, a Chinese Weibo dataset that links disorder-specific Super-Topic communities to users' broader public timelines, and use it to compare Chinese and English online mental health expression. The goal is to provide a resource that supports community-aware analysis and cross-lingual study rather than only single-language disorder detection.

Key Contributions

1. A new multi-disorder Chinese corpus. MATCHA covers seven DSM-5 disorders (ADHD, ASD, bipolar disorder, depressive disorders, generalized anxiety disorder, OCD, schizophrenia) plus a control group, comprising 46,025 cohort-level users and 1,834,699 within-cohort unique posts. It is described as the first Chinese multi-disorder mental health dataset that links disorder-specific Weibo Super-Topic communities with users' broader public timelines.

2. A dual-source retrieval design with user stratification. Posts are split by retrieval source into Super-Topic Posts (STP, concentrated disorder-centered discourse) and Homepage Posts (HP, broader public timelines). An LLM-assisted pipeline using a locally deployed Qwen3 model labels Clinical Evidence Posts (CEP) and defines Identified Positives (IP) separately from Disorder-Related Users (DRU), alongside DSM-5-grounded symptom and life-event classifiers.

3. A multi-level analytical framework. The authors connect community-level dynamics with individual linguistic and behavioral patterns, rather than restricting analysis to user- or post-level detection.

4. A cross-lingual comparison with Twitter-STMHD. Using the four cohorts with shared data and label mappings (BD, DEP, GAD, OCD), they compare linguistic patterns, posting hours, symptom distributions, and life-event distributions across Weibo and Twitter.

Main Findings

• Identified Positives post far more than other users. IP users' per-user activity is 1.7–14.7 times higher than non-IP users in STP. For ASD specifically, IPs make up under 9% of STP users but generate nearly 59% of posts.

• STP versus HP engagement varies by disorder. Across all disorders, HP volumes substantially exceed STP volumes, with STP/HP ratios ranging from 0.18 for ASD to 0.06 for schizophrenia.

• Elevated evening posting across all disorder cohorts. Relative to the original control cohort, disorder cohorts show +0.5 to +6.3 percentage-point higher user-normalized evening (20:00–23:00) posting, with the largest increases for DEP (+6.3), BD (+5.8), and GAD (+5.4). Night ratios (22:00–05:00 share divided by 10:00–17:00 share) exceed the control value of 0.88 only for BD (1.18), DEP (1.20), and OCD (0.94); the control baseline itself is 0.88.

• Consistent affective and cognitive shifts in Chinese LIWC. Most disorder groups show lower positive emotion and higher negative emotion in STPs, alongside substantially elevated cognitive-process categories (insight, discrepancy, tentative). For example, negative emotion in STP is +3.21 percentage points for OCD and +2.84 for GAD.

• Pronoun patterns differ for early-onset disorders. In the auxiliary English LIWC analysis, ADHD and ASD show relatively stronger third-person pronoun signals (e.g., she/he), compatible with caregiver-mediated narration, whereas the other cohorts show higher first-person pronoun use relative to controls, especially in STPs. Present-focus is elevated across cohorts; future-focus differences are small.

• Context amplifies proxy signals. Classifier-derived symptom and life-event densities are higher in STP than HP for both signal families in every disorder. Symptom density is highest in OCD and GAD STP; life-event density is relatively high in ADHD and GAD STP.

• Homepage posts retain detectable cohort signals. HP-only cohort prediction reaches an average Macro-F1 of 0.842 (AUROC 0.932) versus 0.953 (AUROC 0.995) for STP-only, despite explicit disorder or clinical cues appearing far less often in HP (4.64% vs. 19.74%). After masking these cues, Macro-F1 drops by 0.009 for HP and 0.017 for STP.

• Platform and language similarity exceeds disorder similarity. Across 73 LIWC categories, disorder groups are closer to other groups in the same language than to their disorder counterparts across platforms. The same clustering-by-platform pattern appears for posting hours, life-event distributions, and symptom distributions.

• Specific cross-platform lexical contrasts. MATCHA shows stronger increases in negative emotion, anxiety, sadness, and cognitive-processing categories than Twitter-STMHD, and stronger present-focus, whereas Twitter-STMHD shows relatively greater past-focus. Life-related categories such as leisure and ingest are lower in MATCHA. Avoidance was excluded from the Twitter-STMHD symptom comparison because its English classifier estimates were unstable.

• Annotation pipeline reliability. In a blinded audit of 200 posts (100 LLM-positive, 100 LLM-negative, balanced across seven disorders), the LLM achieved 88.0% precision, 83.0% recall, and 85.4% F1 against majority human reference. Human reliability was high (87.3% average pairwise agreement; average pairwise Cohen's kappa = 0.747; Fleiss' kappa = 0.746; 81.0% unanimous agreement), and on the unanimously labeled subset LLM precision was 95.8%. Errors were conservative: the 12 false positives clustered in caregiver-oriented or educational/informational posts, especially for ASD, while the 18 false negatives more often involved implicit self-reference, weak medication mentions, or ADHD/ASD family reports.

• Human annotation of symptoms and life events. 21 annotators labeled 13,723 unique Weibo posts across 38 symptom and 12 life-event categories, with macro-averaged Cohen's kappa of 0.72 across labels among multiply annotated posts.

Methodology in Plain English

The team used Weibo's official APIs under an approved research protocol. For each disorder they identified one to three public Super-Topics (for example, the depression, schizophrenia, or ADHD Super-Topic) and pulled two streams of posts: posts directly from those disorder communities (Super-Topic Posts), and posts from the public personal timelines of the users found in those communities (Homepage Posts). A control group was assembled separately using 20 everyday keywords such as "breakfast," "shopping," and "weather," which produced 3,992 raw anchor posts before expansion and deduplication.

Because joining a Super-Topic is not a diagnosis, they added a strict annotation step. A locally deployed Qwen3 model flagged posts as Clinical Evidence Posts if they contained an explicit diagnosis, a prescription medication mention, a direct self-disclosure or family declaration, or evidence of hospitalization or ongoing treatment. For early-onset conditions (ADHD, ASD) family-member reports were retained; for other disorders only first-person disclosures counted. Users with at least one such post became Identified Positives; everyone else found through the Super-Topics remained Disorder-Related Users. Three researchers then blind-audited 200 posts to check the LLM against human judgment.

For finer-grained signals, 21 annotators with psychology, medicine, or linguistics backgrounds were trained on DSM-5-based guidelines and labeled posts across 38 symptom categories and 12 life-event categories, with labels aggregated before classifier training. Separate English and Chinese classifiers were then trained for symptom and life-event detection so the Chinese Weibo data and English Twitter-STMHD data could be compared on the same categories.

Analysis proceeded in layers. Posting behavior covered volume, engagement ratios, and hourly patterns. Language was measured with the Chinese LIWC lexicon (SC-LIWC) on the original text, plus English LIWC2015 on machine-translated posts as an auxiliary check on pronouns and time orientation. Classifier outputs were converted into per-post densities, defined as the number of positive labels divided by the number of posts, and treated as aggregate proxies rather than clinical assessments. A cue-reduced experiment tested whether cohort membership could still be predicted from Homepage Posts after masking explicit disorder and clinical cues. Finally, cross-platform comparisons computed cohort-versus-control log ratios over 73 LIWC categories, along with correlations of posting-hour, symptom, and life-event distributions.

Temporal coverage was handled with care: STP collection spans May 13, 2014 to January 14, 2025, while disorder-cohort HP reach back to April 10, 2010, and control HP span October 17, 2009 to December 29, 2024. Of 1,559,045 disorder-cohort HP, 1,551,872 (99.5%) fall inside the STP range, and restricting HP to that range retains 30,843 of 30,851 users (over 99.9%).

Why This Matters

Impact on research. Existing mental health NLP resources are overwhelmingly English and built around explicit self-disclosure on Reddit or Twitter. MATCHA adds a Chinese, community-structured alternative and shows that platform and language differences can outweigh disorder differences on standard measures — a caution for anyone assuming that a depression signal learned in English transfers cleanly. The dual STP/HP design also offers a controlled way to study how much of a mental-health signal depends on explicit clinical vocabulary.

Real-world applications:

• Platform community support: Super-Topics act as high-density spaces for affective and health-related expression, which could inform how platforms surface support resources within disorder-related communities.
• Chinese-language screening research: A multi-disorder, cohort-labeled corpus enables development and benchmarking of Chinese cohort-level models, where previously only depression-focused resources such as WU3D existed.
• Caregiver-aware analysis: The finding that ADHD and ASD content shows stronger third-person pronoun use, and that ASD IPs generate a disproportionate share of posts, points to designing tools that account for family members posting on someone else's behalf.
• Cross-cultural clinical communication: The descriptive contrasts in symptom composition and time orientation between Weibo and Twitter offer hypotheses for clinicians and researchers studying how distress is narrated in different language environments.

Industry relevance. The controlled-access, de-identified release format and the LLM-assisted annotation audit results are directly useful to teams building moderation, escalation, or resource-recommendation systems for Chinese-language platforms. The cue-reduced benchmark is especially relevant: it quantifies how much predictive signal survives once explicit disorder keywords are removed, which is closer to realistic deployment conditions than Super-Topic-saturated evaluation.

Future Directions

• Modeling comorbidity and progression. The paper states that comorbidity, symptom overlap, and temporal disorder progression are not explicitly modeled in the current release, leaving these as open modeling problems.
• Broadening disorder coverage. MATCHA covers seven DSM-5 disorders and does not span the full range of psychiatric or neurodevelopmental conditions; extending the cohort set is a natural next step.
• Stronger privacy-preserving release mechanisms. The limitations section explicitly calls for continued work on privacy-preserving release, alongside culturally adaptive annotation frameworks and clinically informed evaluation protocols.
• Disentangling culture from platform. The design cannot separate cultural effects from platform affordances, sampling, demographics, classifier transfer effects, and residual cross-lingual annotation mismatches. Building matched sampling or shared-label protocols across platforms is an unresolved question the authors raise.

Target Audience

Researchers in computational linguistics and NLP working on mental health, dataset builders interested in non-English and community-structured corpora, and annotation-methodology researchers interested in LLM-assisted labeling with human audits. It is also relevant to clinical and public-health researchers who want descriptive, aggregate-level evidence about how mental health is discussed on Chinese social media, and to platform trust-and-safety or digital mental health teams evaluating Chinese-language cohort modeling — provided they read the paper's repeated caveat that all signals are proxies, not diagnoses.

Authors’ 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.