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

Nikita Soni, August Nilsson+4 more|arXiv (Cornell University)|Mar 6, 2026

Mental Health via Writing0 citations

TL;DR

The paper proposes theory-grounded, interpretable models that combine person-level traits with Situational 8 DIAMONDS features and HaRT-based embeddings to predict well-being and detect adaptive/maladaptive self-states in longitudinal social media data. It compares principled baselines to psychometrically informed language models and analyzes predictive features for interpretability.

ABSTRACT

Mental health is not a fixed trait but a dynamic process shaped by the interplay between individual dispositions and situational contexts. Building on interactionist and constructionist psychological theories, we develop interpretable models to predict well-being and identify adaptive and maladaptive self-states in longitudinal social media data. Our approach integrates person-level psychological traits (e.g., resilience, cognitive distortions, implicit motives) with language-inferred situational features derived from the Situational 8 DIAMONDS framework. We compare these theory-grounded features to embeddings from a psychometrically-informed language model that captures temporal and individual-specific patterns. Results show that our principled, theory-driven features provide competitive performance while offering greater interpretability. Qualitative analyses further highlight the psychological coherence of features most predictive of well-being. These findings underscore the value of integrating computational modeling with psychological theory to assess dynamic mental states in contextually sensitive and human-understandable ways.

Motivation & Objective

• Bridge psychological theory with NLP to model dynamic mental health states in context.
• Integrate person-level traits with Situational 8 DIAMONDS features as a principled baseline.
• Leverage HaRT-based, person-contextual embeddings to capture temporal dynamics in language.
• Evaluate predictive performance for well-being and adaptive/maladaptive self-states.
• Provide qualitative insights into which psychological features most strongly relate to well-being.

Proposed method

• Construct a principled baseline combining S8D situational features with four PLT subdomains (implicit motives, mental health, resilience, cognitive distortions).
• Annotate posts with eight S8D dimensions using Deepseek-R1 few-shot prompting.
• Compute 19 PLT features at sentence and post levels using RoBERTa-Large, Valence/Harmony/Satisfaction/Anxiety/Depression indices, and ReLM resilience facets.
• Fine-tune HaRT to generate person-contextual embeddings from temporal histories for post- and sentence-level predictions.
• Evaluate at post level for continuous well-being (ridge regression) and at sentence level for adaptive/maladaptive labels (logistic regression).
• Use nested 5-fold cross-validation to assess models and thresholds to extract evidence-level annotations.

Figure 1: Distribution of probabilities to predict adaptive state for a given sentence. On the top is using HaRTWB-FT + PLT features, and the bottom is using PLT features in Logistic Regression models.

Experimental results

Research questions

• RQ1 Can a theory-grounded baseline combining S8D and PLT features predict well-being as effectively as or better than psychometric-language models?
• RQ2 Do HaRT person-contextual embeddings provide additional predictive power for well-being and self-state evidence detection beyond principled baselines?
• RQ3 Which psychological features most strongly correlate with well-being and maladaptive/adaptive self-states in longitudinal social media data?

Key findings

• Situational characteristics inferred from posts (S8D) are predictive of annotated well-being scores.
• A principled baseline combining S8D with PLT features improves performance over individual feature groups.
• HaRT-based person-contextual embeddings outperform baselines on internal validation for well-being prediction, with combined feature sets achieving strongest performance.
• HaRT-based models effectively identify adaptive and maladaptive evidence spans, showing higher sensitivity to language variation than baseline features.
• Qualitative analysis highlights positive associations of well-being with satisfaction with life, harmony in life, and positivity, and negative associations with certain resilience and cognitive distortion indicators.

Figure 2: Qualitative analysis of features in our principled baseline consisting of psychological characteristics of the situation and person-level traits. Left: Pearson correlation coefficients; Right: Ridge regression beta coefficients for predicting well-being with the S8D and PLT features.