Cross-cultural differences in language markers of depression online

Kate Loveys, Jonathan Torrez, Alex Fine, Glen Moriarty, Glen CoppersmithNAACL 2018

TL;DR

This study investigates how cultural background influences the language used by individuals discussing depression online. By analyzing text from a peer support community, the researchers found that different cultural groups express their experiences of depression using distinct linguistic patterns, highlighting the need for culturally sensitive approaches in automated mental health assessments.

What Problem Does This Paper Address?

Depression is a global mental health crisis, yet current automated tools for its detection often rely on language analysis that may not account for cultural variations. This is problematic because cultural norms shape how individuals express distress, what symptoms they report, and their beliefs about mental illness. Without considering these cultural nuances, machine learning models trained on one population's language might misinterpret or misdiagnose depression in others, leading to ineffective or even harmful treatment outcomes. This paper aims to bridge this gap by exploring cross-cultural differences in online depression expression.

Methods

The researchers employed a mixed-methods approach to analyze online text data. They utilized:

  • Linguistic Inquiry and Word Count (LIWC): This established software analyzes text to quantify the percentage of words related to various psychological constructs (e.g., emotion, cognition) and life domains (e.g., health, social). This allowed for a quantitative comparison of language use across cultural groups.
  • Topic Modeling (Latent Dirichlet Allocation - LDA): This technique was used to identify emergent themes and topics within the discussions. LDA helps uncover patterns in user preoccupations that might not be immediately obvious through manual analysis.
  • Jaccard Similarity Coefficient: This metric was used to quantify the overlap in terms between topics identified by LDA across different cultural groups, providing a measure of how similar their thematic discussions are.

The study focused on exploratory analyses, meaning they aimed to identify potential differences rather than test pre-defined hypotheses. Statistical tests were not performed on the LIWC results due to the exploratory nature and the potential for misleading interpretations without specific hypotheses.

Datasets

The study utilized a dataset of written dialogue from 7 Cups of Tea, an online peer support community for emotional distress. The data comprised text generated by 1,593 users who self-reported depression as their primary concern at sign-up. Participants were categorized into four main cultural groups: White, Asian or Pacific Islander, Black or African American, and Hispanic or Latino. Users who did not indicate their culture or selected "Other" were excluded.

Key Insights and Ideas

  • Culture Matters in Online Depression Expression: The most significant takeaway is that cultural background demonstrably influences how individuals discuss their depression online. This challenges the assumption that language markers of depression are universal.
  • Emotion Expression Varies: Different cultural groups exhibit distinct patterns in expressing emotions. For instance, Asian or Pacific Islander users showed more inhibition of negative emotions, while White and Black or African American users expressed more negativity. Hispanic or Latino users showed a high amount of both positive and negative emotion.
  • Cognitive and Functional Differences: Cultural groups also differed in how they discussed cognitive aspects of depression and their functioning. For example, cognitive impairment markers were less evident in the language of Asian or Pacific Islander users. White users were more likely to discuss health and death, while Asian or Pacific Islander users were less open about these topics but more present with social terms.
  • Topic Overlap is Limited: The topic modeling revealed minimal overlap in the specific terms used within topics across cultural groups, suggesting that while some broad concerns like "friend" (loneliness) might be universal, the specific thematic discussions are culturally shaped.
  • Implications for AI: These findings underscore the critical need to develop culturally sensitive automated depression assessment tools. Without such adaptations, these tools risk misclassification and exacerbating existing health disparities.

Key Results

  • LIWC Analysis:
    • Emotion (Tone): Asian or Pacific Islander users showed more inhibition of negative emotion; White and Black or African American users expressed more negativity; Hispanic or Latino users expressed a high amount of both positive and negative emotion.
  • Cognition: Cognitive effects of depression were less evident in the language of Asian or Pacific Islander users.
  • Functioning: White users discussed health and death/self-destruction more, while being less social. Asian or Pacific Islander users were less open about health/death but discussed social terms more. Black or African American users discussed social terms highly, were less likely to discuss death, but more willing to talk about health. Hispanic or Latino users had low mentions of social terms and were less willing to discuss death, self-destruction, religion, or health.
  • Topic Modeling: While broad themes like "friend" were common across groups, the specific terms and their arrangement into topics showed significant variation, indicating distinct thematic preoccupations. Jaccard similarity coefficients between topics across cultural groups were generally low (around 0.1), with few instances of high overlap.

Limitations and Future Work

The authors acknowledge several limitations:

  • Platform Specificity: The data was collected from a peer support community. It's unclear if these findings generalize to other online spaces like social media, where discussions about mental health are not explicitly encouraged and anonymity is more prevalent.
  • Missing Data: Information on participants' country of residence and extent of acculturation was not collected. This makes it difficult to fully understand how deeply individuals were entrenched in their cultural beliefs versus Western norms. For example, the study couldn't distinguish between a white American and a white German, or a third-generation Korean American and a Korean citizen living in the US.
  • Group Labeling: The study used ethnic and racial labels determined by the peer support community, which are inherently imperfect and can be a mix of categories.

Future work suggested includes:

  • Replication and Broader Cultural Groups: Replicating these findings with more diverse cultural groups and in different online contexts.
  • Acculturation Studies: Investigating the role of acculturation in shaping language use related to mental health.
  • Hypothesis-Driven Research: Using these exploratory findings to guide future, hypothesis-driven research with statistical testing.
  • Extending to Other Conditions: Applying similar analyses to other mental health conditions beyond depression.
  • Intervention and Policy Development: Using these insights to inform the development of culturally sensitive interventions and mental health policy.

Summary updated

BibTeX
@inproceedings{loveys2018cross,
  author = {Kate Loveys and Jonathan Torrez and Alex Fine and Glen Moriarty and Glen Coppersmith},
  title = {{Cross-cultural differences in language markers of depression online}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure