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

This paper proposes an interpretable modeling approach that integrates person-level psychological traits with situational context features derived from social media data to predict dynamic mental well-being, demonstrating that theory-driven methods offer competitive performance and greater human-understandable insights compared to standard language model embeddings.

The Gist

Imagine your mental health isn't a fixed statue in a garden, but rather a kaleidoscope. Every time you turn the tube (change your situation), the pattern shifts. Sometimes the colors are bright and harmonious; other times, they are chaotic and dark.

This paper is about building a machine that doesn't just look at the kaleidoscope and say, "That's a sad picture." Instead, it wants to understand why the picture looks that way by looking at two things: who you are and where you are.

Here is the breakdown of their work, translated into everyday language:

1. The Big Idea: It's Not Just "You," It's "You + The Moment"

Most old ways of checking mental health are like taking a photo of a person and labeling them "Happy" or "Sad" forever. But humans are messy! You might be a generally resilient person, but if you're stuck in traffic and your boss just yelled at you, you might feel terrible in that moment.

The authors argue that to understand mental health, you need to look at the interaction between:

• The Person: Your deep-seated traits (like your natural optimism, your hidden fears, or how you handle stress).
• The Situation: The specific context of the moment (Is it a fight? A celebration? A boring meeting?).

2. The Two Tools They Built

To solve this, the team built two different "lenses" to look at people's social media posts.

Lens A: The "Psychology Detective" (The Theory-Based Approach)

This is their "principled baseline." Instead of letting a computer guess blindly, they gave it a rulebook based on real psychology.

• The Person Part: They used AI to scan posts and guess things like "Is this person feeling anxious?" "Do they feel hopeful?" or "Are they using negative thinking traps?"
• The Situation Part: They used a framework called DIAMONDS (a fancy acronym for 8 types of situations: Duty, Intellect, Adversity, Mating, Positivity, Negativity, Deception, Sociality).
  • Analogy: Imagine a post says, "I'm so stressed about this deadline." The AI tags this as Adversity (high stress) and Duty (work).
• The Result: This lens is like a doctor who asks specific questions. It's very clear on why it thinks you are struggling, but it might miss some subtle nuances.

Lens B: The "Memory-Keeping Robot" (HaRT)

This is their advanced AI model. Think of this robot as a super-attentive friend who has read everything you've ever written.

• It doesn't just look at one sentence; it looks at your whole history. It knows that when you say "I'm fine," it might mean something different than when your friend says it.
• It creates a "digital fingerprint" of you that changes as you change over time.
• The Result: This lens is great at spotting patterns and predicting the future, but it's a bit of a "black box"—it's hard to explain exactly why it made a decision.

3. The Experiment: Which Lens is Better?

They tested both lenses on a dataset of 30 people's social media posts to see who could best predict:

1. How well the person is doing overall (Well-being).
2. Which specific sentences show they are coping well (Adaptive) vs. struggling (Maladaptive).

The Findings:

• The "Psychology Detective" (Lens A) did surprisingly well! It proved that if you combine what we know about human personality with what we know about situations, you get a very accurate picture. Plus, it's explainable. We can say, "The model thinks you're struggling because you're in a high-stress situation and you have a tendency toward negative thinking."
• The "Memory-Keeping Robot" (Lens B) was also very good, especially at spotting the specific moments of struggle. It was sensitive to the tiny shifts in how people talk over time.
• The Winner? A Hybrid Team. When they combined the clear rules of the Detective with the deep memory of the Robot, they got the best results.

4. Why This Matters (The "So What?")

The authors found some fascinating things:

• Context is King: Being "hyper-vigilant" (always on guard) might be a bad thing in a classroom, but it's a good thing for a soldier in combat. The AI learned that the same behavior can be healthy or unhealthy depending on the situation.
• The "Faith" Paradox: They found that for some people, relying on "a higher power" was actually linked to lower well-being in their data. Why? Because the AI interpreted it as "giving up control" in a situation where the person needed to take action. It's a reminder that AI needs human context to understand the meaning behind the words.

The Takeaway

This paper is a call to stop treating mental health like a static label (e.g., "Depressed"). Instead, we should treat it like a weather system.

• You are the climate (your personality).
• The Situation is the weather front (the storm or the sunshine).
• Mental Health is the current forecast.

By using AI that understands both the climate and the weather, we can build tools that don't just tell us "it's raining," but explain why it's raining and how we can best navigate the storm. This makes the technology not just smart, but human-understandable and safe to use.

Technical Summary

Here is a detailed technical summary of the paper "Who We Are, Where We Are: Mental Health at the Intersection of Person, Situation, and Large Language Models."

1. Problem Statement

Mental health is traditionally assessed using static diagnostic labels that assume stable traits. However, psychological theory (specifically interactionist and constructionist frameworks) posits that mental health is a dynamic process resulting from the interplay between an individual's dispositions (person) and their immediate context (situation).

The challenge addressed in this work is how to computationally model these dynamic self-states in longitudinal social media data. Specifically, the authors aim to:

1. Predict overall well-being scores for social media posts.
2. Identify specific spans of text that serve as evidence for adaptive (healthy) vs. maladaptive (unhealthy) self-states.
3. Move beyond "black box" deep learning models to create systems that are both accurate and interpretable, grounded in established psychological theory.

2. Methodology

The authors propose a hybrid approach that integrates theory-driven feature engineering with person-contextualized language modeling.

A. Theoretical Framework

The approach is built on the Interactionist Perspective, combining:

• Person-Level Traits (PLT): Stable psychological anchors (e.g., resilience, cognitive distortions, implicit motives).
• Situational Context: The specific psychosocial characteristics of the environment where the text was written.

B. Feature Extraction

The model utilizes two primary categories of features:

1. Theory-Grounded Baseline Features:

• Situational 8 DIAMONDS (S8D): Using the DeepSeek-R1 LLM with few-shot prompting, the authors annotated each post for eight situational dimensions: Duty, Intellect, Adversity, Mating, Positivity, Negativity, Deception, and Sociality. Scores range from 1 (absent) to 9 (highly present).
• Person-Level Traits (PLT): Extracted using pre-trained, validated models to estimate:
  • Implicit Motives: Achievement, affiliation, and power (via fine-tuned RoBERTa-Large).
  • Mental Health Dimensions: Valence, life harmony, life satisfaction, anxiety, and two depression indices.
  • Resilience: Nine facets (e.g., optimism, coping toolkit) using the ReLM (Resilience through Language Modeling) framework.
  • Cognitive Distortions: Levels of distorted thinking patterns.

2. Person-Contextualized Embeddings (HaRT):

• The authors utilized HaRT (Human-aware Recurrent Transformer), a model trained on the Human Language Modeling (HuLM) task.
• Unlike standard LLMs, HaRT incorporates a latent user representation derived from the user's temporal historical language. This allows the model to understand language patterns specific to an individual over time, capturing the "person" aspect of the interaction.

C. Modeling Strategy

The team evaluated these features using 5-fold nested cross-validation on a dataset of 30 users (343 posts, 199 annotated) from the CLPsych 2025 Shared Task.

• Task 1 (Well-being): Predicting continuous well-being scores using Ridge Regression.
• Task 2 (State Identification): Predicting binary labels (Adaptive vs. Maladaptive) for sentence spans using Logistic Regression.
• Hybrid Configurations: They tested the theory-driven features alone, the HaRT embeddings alone, and combinations of both.

3. Key Contributions

1. Principled Baseline Proposal: Introduced a theory-driven baseline combining S8D situational features and PLT, offering a highly interpretable reference point for mental health modeling.
2. Integration of Person-Contextual Embeddings: Demonstrated the utility of HaRT, which captures individual-specific temporal patterns, aligning with psychometric theories of trait stability.
3. Interpretability vs. Performance Trade-off: Showed that while deep learning embeddings (HaRT) offer high predictive power, the theory-driven features provide superior interpretability and competitive performance, especially when combined.
4. Qualitative Validation: Conducted a qualitative analysis confirming that the most predictive features align with established psychological literature (e.g., "positivity" and "life satisfaction" correlating with well-being).

4. Results

• Well-being Prediction:
  • The S8D + PLT baseline achieved a Pearson correlation ($r$) of 0.629 and MSE of 2.149.
  • The HaRT model (fine-tuned) showed higher internal validation performance ($r=0.684$) but signs of overfitting in cross-validation.
  • Key Finding: The theory-driven baseline (S8D + PLT) outperformed the HaRT-only variant in the official evaluation, proving that psychological theory can rival or exceed data-driven embeddings in specific contexts.
• Adaptive/Maladaptive State Detection:
  • HaRT-based models were more sensitive to language variation and better at identifying evidence spans than PLT features alone.
  • Hybrid Approach: Combining HaRT embeddings with PLT features yielded the best results (Macro F1 $\approx$ 0.58, AUC $\approx$ 0.77 for maladaptive states), outperforming either method in isolation.
• Feature Analysis:
  • Positive Correlates: "Satisfaction with life," "Positivity" (situational), and "Harmony in life."
  • Negative Correlates: "Depression scale," "Cognitive distortions," and surprisingly, "Belief in a higher power" (interpreted as a potential sign of passive reliance rather than active coping in this specific context).

5. Significance and Implications

• Bridging Theory and Computation: The paper successfully demonstrates that integrating psychological theory (interactionism) with NLP leads to models that are not only accurate but also human-understandable. This is crucial for clinical applications where "black box" predictions are insufficient.
• Dynamic Assessment: By moving from static traits to dynamic person-situation interactions, the model captures the nuance that a behavior can be adaptive in one context (e.g., hypervigilance in combat) but maladaptive in another (e.g., in a classroom).
• Ethical Considerations: The authors emphasize the need for caution regarding privacy, consent, and bias. They argue that while these tools are powerful, they must be deployed with human oversight and an understanding that linguistic proxies are not definitive clinical diagnoses.
• Future Direction: The work suggests a path toward systems that view individuals not as fixed entities but as contextually situated and dynamically evolving, supporting more flexible and resilient mental health interventions.

In summary, this paper provides a robust framework for mental health analysis that respects the complexity of human psychology, proving that interpretable, theory-grounded features are essential for advancing computational social science beyond mere prediction accuracy.