Who We Are, Where We Are: Mental Health at the Intersection of Person, Situation, and Large Language Models

Nikita Soni, August Håkan Nilsson, Syeda Mahwish, Vasudha Varadarajan, H. Andrew Schwartz, Ryan L. BoydNAACL 2025

TL;DR

This paper introduces a novel approach to understanding mental health by integrating psychological theory with large language models (LLMs). They developed interpretable models that predict well-being and identify adaptive/maladaptive self-states from social media data, showing that theory-driven features offer competitive performance and greater explainability than purely data-driven LLM embeddings.

What Problem Does This Paper Address?

Understanding mental health has traditionally relied on static assessments or diagnostic labels. However, psychological theories increasingly emphasize that mental states are dynamic, constantly shaped by the interplay between an individual's inherent traits and the situational contexts they experience. This paper tackles the challenge of modeling these dynamic mental states, particularly within digital communication where personal expression offers a rich window into internal states. The authors aim to move beyond simply predicting mental health outcomes to providing interpretable insights into how and why these states fluctuate, bridging the gap between computational modeling and psychological theory.

Methods

The authors propose a two-pronged approach:

  1. A Principled Baseline: This involves combining:
  2. Situational Context Features: Using the Situational 8 DIAMONDS (S8D) framework, they annotated social media posts for eight psychosocial aspects of the situation (e.g., duty, intellect, adversity). This was done using a large language model (Deepseek-R1) with few-shot prompting.
  3. Person-Level Traits (PLT): They extracted 19 features representing implicit motives (achievement, affiliation, power) and mental health dimensions (valence, harmony, satisfaction, depression, anxiety, resilience, cognitive distortions) using pre-existing, fine-tuned models.
  4. Human-Centered Language Model (HaRT): They fine-tuned a human-centered language model (HaRT) trained on the Human Language Modeling (HuLM) task. HaRT incorporates a latent user representation derived from their temporal language history, enabling person-contextualized embeddings that capture individual-specific patterns over time.

These approaches were evaluated for predicting continuous well-being scores and identifying adaptive/maladaptive self-states.

Datasets

The primary dataset used is from the CLPsych 2025 shared task. This dataset consists of approximately 343 posts from 30 users, with 199 posts annotated for adaptive and maladaptive self-states (as text spans) and overall well-being scores.

Key Insights and Ideas

  • Theory-Driven Features Matter: The paper strongly advocates for integrating psychological theory into computational models. Their "principled" baseline, combining S8D and PLT features, not only performed well but also offered significantly greater interpretability.
  • Situational Context is Crucial: The S8D features, representing the environment in which a person's expression occurs, were found to be predictive of well-being, highlighting that mental states are not solely individual but are deeply intertwined with context.
  • Interpretability Enhances Understanding: While LLM embeddings (like HaRT) can offer high predictive accuracy, the authors demonstrate that simpler, theory-grounded features provide more transparent insights into why certain patterns are associated with mental health outcomes.
  • Nuance in Resilience: The "daily lived experience" facet of resilience showed a negative association with well-being, suggesting that simply "pushing through" might mask underlying difficulties rather than indicate robust well-being in certain contexts. Similarly, "belief in a higher power" was negatively weighted, possibly due to its interpretation as passive external control rather than active mastery.

Key Results

  • Well-being Prediction: The theory-driven baseline (S8D + PLT) achieved competitive performance in predicting continuous well-being scores, as shown in Table 1 and Table A2.
  • Adaptive/Maladaptive States: HaRT models performed well in classifying adaptive and maladaptive self-states, with additional gains when combined with PLT features (Table 3 and Table A4).
  • Qualitative Analysis: The top features positively correlated with well-being included 'satisfaction with life', 'positivity' (S8D), and 'harmony in life'. Conversely, 'higher power belief', 'depression', and overall 'resilience score' were negatively correlated. The ridge regression model highlighted positive importance for 'sociality', 'positivity', and 'intellect' (S8D), and negative importance for 'daily lived experience' and 'need for affiliation'.

Limitations and Future Work

  • Dataset Size: The CLPsych 2025 dataset is relatively small (30 users, <200 posts), limiting the generalizability of the findings. Future work should evaluate models on larger, more diverse datasets.
  • Interpretability of Labels: The ground truth labels themselves are interpretive, raising questions about the extent to which linguistic expressions are definitive proxies for psychological states.
  • Model Biases: Pre-trained models used for feature extraction may carry latent biases or be insensitive to cultural nuances.
  • Correlational Nature: The models identify linguistic markers but cannot definitively speak to underlying mechanisms, causal relationships, or interventions. Longitudinal clinical assessments are needed for validation.

The authors suggest future work should explore how different contexts modulate trait adaptiveness and how language-based systems can support more flexible, resilient self-states over time. They emphasize the importance of integrating theory and computation for meaningful and human-understandable outputs.

Summary updated

BibTeX
@inproceedings{soni2025who,
  author = {Nikita Soni and August Håkan Nilsson and Syeda Mahwish and Vasudha Varadarajan and H. Andrew Schwartz and Ryan L. Boyd},
  title = {{Who We Are, Where We Are: Mental Health at the Intersection of Person, Situation, and Large Language Models}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure