PsihoRo: Depression and Anxiety Romanian Text Corpus

The paper introduces PsihoRo, the first open-source Romanian corpus for depression and anxiety, which was constructed using open-ended responses and standardized screening surveys (PHQ-9 and GAD-7) from 205 participants to address the lack of mental health resources for the Romanian language in NLP.

The Gist

Imagine you are trying to understand the mood of a whole country by listening to its people talk. For years, scientists have been able to do this easily in English, using vast libraries of text to spot signs of sadness or worry. But for Romanian speakers, this library was empty. There was no "Rosetta Stone" to translate their words into data about mental health.

This paper introduces PsihoRo, a new, open-source library designed to fill that gap. Think of it as building the first dedicated "mental health map" specifically for the Romanian language.

Here is a simple breakdown of how they built it and what they found:

1. How They Built the Map (The Data Collection)

Instead of just scrolling through social media (which is like listening to people shout their best moments at a party), the researchers invited 205 people to sit down and have a quiet, honest conversation.

• The Interview: They asked six open-ended questions. Some were about heavy topics like war, politics, and stress (the "stormy weather"), while others were about resilience and community (the "sunshine").
• The Check-Up: To make sure the data was accurate, they didn't just guess who was struggling. They gave everyone two standard medical "thermometers": the PHQ-9 (for depression) and the GAD-7 (for anxiety). These tools give a score to see how much someone is struggling.
• The Result: They ended up with a collection of 205 honest stories, paired with medical scores, all completely anonymous.

2. The Detective Work (The Analysis)

Once they had the stories, they used three different "magnifying glasses" to look for patterns:

• The Word Counter (LIWC): They used a special software that counts specific types of words.
  • The Big Surprise: In English, using the word "I" a lot is a classic sign of depression. But in Romanian, this didn't work. Romanian is a "pro-drop" language, meaning you can often drop the subject "I" from a sentence without it sounding weird. So, the researchers found that counting "I" is useless for Romanian mental health analysis.
  • What actually worked? They found that people with higher anxiety or depression scores used more words about their bodies (like "headache" or "stomach"), more words showing uncertainty ("maybe," "perhaps"), and fewer words about success or fun.
• The Emotion Radar: They built an AI to detect emotions like sadness, fear, or joy.
  • People at risk for depression spoke a lot about sadness.
  • People at risk for anxiety spoke a lot about fear and trust (often worrying about their families).
  • Interestingly, the emotion of "surprise" was completely missing from the data!
• The Theme Finder (Topic Modeling): They grouped the conversations by theme.
  • Everyone was worried about politics and the economy (a common theme in Romania right now).
  • However, the at-risk groups focused more on how these external stresses were draining their energy, affecting their sleep, and making them feel isolated from their communities.

3. Why This Matters

Think of this research as giving a translator to a doctor who speaks English but needs to treat a Romanian patient.

• No More Guessing: Before this, researchers had to assume English rules applied to Romanian. This paper proves that languages have different "fingerprints" for mental health.
• A First Step: While 205 people might sound like a small crowd, it's a massive step forward for a language that previously had zero resources. It's the foundation upon which better AI tools, better therapy apps, and more targeted support for Romanians can be built.

In a nutshell: The researchers built a new, honest library of Romanian stories about mental health. They discovered that to understand Romanian sadness and worry, you have to look at what they say about their bodies and their uncertainty, not just how often they say "I." It's a crucial first step toward helping the Romanian population feel seen and understood.

Technical Summary

Here is a detailed technical summary of the paper "PsihoRo: Depression and Anxiety Romanian Text Corpus":

1. Problem Statement

• Data Scarcity in Low-Resource Languages: While Natural Language Processing (NLP) for mental health is advanced in English, there is a significant lack of open-source, high-quality psychological corpora for Romanian. Existing resources are either non-existent, not publicly accessible, or derived from social media (which suffers from positivity bias and lack of clinical validation).
• Linguistic Generalizability Issues: Features predictive of mental health in English (e.g., the frequent use of the first-person pronoun "I") do not necessarily translate to other languages. Romanian is a "pro-drop" language (where subject pronouns are often omitted), making English-based linguistic markers unreliable for Romanian mental health analysis.
• Need for Validated Data: Social media scraping often lacks informed consent and clinical correlation. There is a need for a dataset collected via open-ended questions that is explicitly correlated with standardized clinical screening tools.

2. Methodology

The authors developed PsihoRo, the first open-source corpus for depression and anxiety in Romanian, using a rigorous data collection and analysis pipeline:

• Data Collection:
  • Participants: 205 anonymous respondents collected over an 8-month period (March–October 2025).
  • Instrument: A three-part survey:
    1. Open-ended Questions: 6 questions (3 positive, 3 negative) regarding social media, resilience, politics, community support, war/economic crises, and stress recovery.
    2. Depression Screening: PHQ-9 (Patient Health Questionnaire).
    3. Anxiety Screening: GAD-7 (Generalized Anxiety Disorder).
  • Ethics: Fully anonymous, informed consent obtained, no personal data (IP, names) stored. Approved by the University of Bucharest Ethics Committee.
• Population Stratification: Participants were categorized into three groups based on clinical cut-off scores (≥10):
  • Depression Risk Group: 37 respondents (18%).
  • Anxiety Risk Group: 36 respondents (17.56%).
  • No Risk Group: 28 respondents (13.66% scored 0 on at least one scale).
• Analysis Techniques:
  • LIWC (Linguistic Inquiry and Word Count): Used the Romanian version (Ro-LIWC2015) to analyze 98 psychological and linguistic categories.
  • Classification: A LightGBM classifier was trained on LIWC features to predict binary labels for depression and anxiety (cutoff ≥10).
  • Explainability: SHAP (SHapley Additive exPlanations) was used to identify feature importance.
  • Emotion Detection: A Romanian BERT model was fine-tuned on the REDv2 dataset to perform multi-label emotion classification on the corpus.
  • Topic Modeling: Structural Topic Modeling (STM) using the stm R package to identify thematic patterns across the three population groups.

3. Key Contributions

• First Open-Source Romanian Mental Health Corpus: PsihoRo provides 205 annotated texts linked to validated clinical scores (PHQ-9 and GAD-7), filling a critical gap in Romanian NLP resources.
• Linguistic Validation: The study empirically demonstrates that English-centric markers (specifically the pronoun "I") are not predictive of depression in Romanian, confirming the necessity of language-specific research.
• Novel Linguistic Markers: Identified specific Romanian linguistic patterns correlated with mental health, such as:
  • Tentativeness: High correlation with both anxiety and depression.
  • Body-related words: Strong positive correlation with both conditions (somatic awareness).
  • Negations: Interestingly, higher symptom scores correlated with fewer negations.
  • Achievement/Leisure: Negative correlation with depression/anxiety scores.
• Public Availability: The dataset is publicly released on Hugging Face to facilitate further research.

4. Results

• Descriptive Statistics: The dataset shows a left-skewed distribution, with most respondents having minimal to mild symptoms. A strong correlation ($r=0.761$) exists between PHQ-9 and GAD-7 scores.
• LIWC Findings:
  • Pronoun "I": No significant correlation with depression/anxiety scores, validating the pro-drop hypothesis.
  • Sadness & Death: "Sadness" words increased with depression; "Death" category words correlated significantly with depression but not anxiety.
  • Family: "Family" words correlated positively with anxiety.
  • Classification Performance: The LightGBM model achieved 0.85 accuracy for depression and 0.83 accuracy for anxiety using LIWC features.
• Emotion Analysis:
  • Surprise was absent from the data.
  • Depression Risk Group: Showed a high prevalence of sadness.
  • Anxiety Risk Group: Showed elevated levels of sadness, fear, and trust.
  • Question Type: Pessimistic questions elicited more anger and sadness; optimistic questions elicited more joy and trust.
• Topic Modeling: Distinct themes emerged per group:
  • No Risk: Focused on political critique and social balance.
  • Depression Risk: Focused on fatigue, routine exhaustion, and the negative impact of social networks.
  • Anxiety Risk: Focused on external instability (war, crisis), the link between politics and stress, and the need for community resilience.

5. Significance

• Cultural Specificity: The paper underscores that mental health NLP models cannot be directly transferred across languages. It provides a blueprint for creating culturally and linguistically adapted mental health resources.
• Clinical & Research Utility: By combining open-ended text with standardized clinical scores, PsihoRo enables the development of more accurate screening tools and intervention strategies for the Romanian-speaking population.
• Foundation for Future Work: The dataset serves as a baseline for future studies on Romanian emotional language, allowing researchers to move beyond English-centric assumptions and develop inclusive, multilingual mental health AI.

Limitations & Future Work: The authors note the relatively small sample size (205) and plan to expand the dataset via crowdsourcing in future iterations. They also emphasize the need for multi-perspective approaches beyond self-reported text to fully understand mental health patterns.