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. Boyd

cs.CLMar 6, 2026

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

Authors: Nikita Soni, August Håkan Nilsson, Syeda Mahwish, Vasudha Varadarajan, H. Andrew Schwartz, Ryan L. Boyd

Organizations: Department of Computer Science, Stony Brook University · Oslo Metropolitan University, Oslo Business School · University of Texas at Dallas, Dept. of Psychology

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.

Figures & tables

	5-fold Ridge CV
	$r$ ↑	$MSE$ ↓
S8D	0.528	2.556
Dist	0.365	3.059
ReLM	0.533	2.538
PLT	0.629	2.149
S8D + ReLM + Dist	0.623	2.178

Table 1: Pearson correlation (r) and Mean Squared Error (MSE) when training a ridge regression model using different “principled” baseline features to predict continuous Well-being scores using nested 5-fold cross-validation.

	Internal Val Set		5-fold Ridge CV
	$r$ ↑	$MSE$ ↓	$r$ ↑	$MSE$ ↓
HaRTWB-FT	0.684	1.828	0.876	0.828
HaRTWB-FT + S8D + ReLM + Dist	-	-	0.883	0.787
HaRTWB-FT + S8D + PLT	-	-	0.884	0.783

Table 2: Pearson correlation (r) and Mean Squared Error (MSE) when fine-tuning HaRT using internal train and validation splits and further training a ridge regression model using resulting embeddings and principled baseline features. Note: Pearson

r

values for the 5-fold ridge CV numbers are likely inflated due to partial data contamination across the fine-tuning and cross-validation datasets. The internal validation set was used while finetuning the HaRT model, after which the weights were separately used as inputs for well-being task. The internal validation numbers have been omitted for the post-finetuned model.

Panel of Figure 1 — 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.

	5-fold Log.reg. CV
	$F1_{macro}$	$AUC$
Dist adaptive	0.54	0.75
HaRTWB-FT	0.50	0.74
HaRTWB-FT + ReLM + Dist adaptive	0.53	0.76
PLT adaptive	0.48	0.66
HaRTWB-FT + PLTadaptive	0.52	0.76

Table 3: Macro F1 and AUC results when training a logistic regression model to predict binary adaptive and maladaptive labels separately over sentences split from posts.

Panel of Figure 2 — 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.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

	A.1 Adaptive/Maladaptive
	recall			weighted recall
	overall	adaptive	maladaptive	overall	adaptive	maladaptive
HaRTWB-FT + ReLM + Dist (LinSVC)	0.108	0.099	0.116	0.103	0.097	0.109
HaRTWB-FT + Log.reg.	0.105	0.077	0.132	0.102	0.075	0.129
HaRTadaptive-FT	0.276	0.245	0.308	-	-	-
HaRTmaladaptive-FT	-	-	-	0.236	0.238	0.235

Appendix

Table A1: Task A.1 Adaptive and Maladaptive Evidence task. We found that finetuning HaRT for independent adaptive and maladaptive evidence classification can yield significant boosts over traditional principled baselines.

	A.2 Well-being
	timeline MSE ↓				F1
	overall	minimal impairment	impaired	serious impairment	macro
HaRTWB-FT	3.73	2.76	1.95	5.92	0.15
HaRTWB-FT + Ridge	3.42	2.95	1.33	5.1	0.17
HaRTWB-FT + S8D + ReLM + Dist (Ridge)	3.27	2.63	1.38	4.98	0.19
HaRTWB-FT + S8D + PLT (Ridge)	3.22	1.40	2.60	4.86	0.19

Appendix

Table A2: Task A.2 Wellbeing task. We found that unlike Task A.1, finetuning HaRT to the wellbeing prediction task need not consistently offer boosts, instead, principled and theoretical methods can offer significant advantages, with a small number of interpretable dimensions without compromising on the accuracy. The first 3 rows in this table were our official submissions while others are presented for additional analysis. We note that using HaRt fine-tuned for the respective adaptive and maladaptive binary classifications may provide benefits over using HaRTWB-FT in all combinations, however, due to time constraints we do not have empirical results for the same.

	Internal Val Set
	$F1_{macro}$	$F1_{wtd}$	$AUC$	$Acc$
Dist adaptive	0.53	0.85	0.75	0.89
HaRTadaptive-FT	0.67	0.89	0.66	0.89
Dist maladaptive	0.56	0.82	0.76	0.84
HaRTmaladaptive-FT	0.60	0.82	0.60	0.82

Appendix

Table A3: Task A.1 Additional results on internal validation set when predicting binary adaptive and maladaptive labels separately over sentences split from posts.

	5-fold Log.reg. CV
	$F1_{macro}$	$F1_{wtd}$	$AUC$	$Acc$
Dist adaptive	0.54	0.87	0.75	0.90
HaRTWB-FT	0.50	0.87	0.74	0.90
HaRTWB-FT + ReLM + Dist adaptive	0.53	0.87	0.76	0.91
PLT adaptive	0.48	0.86	0.66	0.91
HaRTWB-FT + PLTadaptive	0.52	0.87	0.76	0.90

Appendix

Table A4: Additional results when training a logistic regression model to predict binary adaptive and maladaptive labels separately over sentences split from posts.

Panel of Figure 10 — Figure A1: Distribution of probabilities to predict maladaptive 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.

Panel of Figure 11 — Figure A2: Distribution of probabilities to predicting adaptive state for a given sentence. On the top is using HaRTWB-FT + ReLM + Dist adaptive features, and the bottom is using ReLM + Dist adaptive features in Logistic Regression models.

Panel of Figure 12 — Figure A3: Distribution of probabilities to predicting maladaptive state for a given sentence. On the top is using HaRTWB-FT + ReLM + Dist maladaptive features, and the bottom is using ReLM + Dist maladaptive features in Logistic Regression models.

Interpretable ML

Interpretable machine learning develops models and explanations that make predictions, evidence, and decision processes understandable to people.

See all 306 papers

Natural Language Processing Psychometrics
Aug 7, 2026Edoardo Sebastiano De Duro, Emma Franchino, Massimo StellaClinical Outcome Prediction Psychometric Validation
CogNoscoLab, Department of Psychology and Cognitive Science, University of Trento, Italy
psytechlab at CLPsych 2026: Utilising Natural Language Processing methods and Large Language Models for Social Media Text Analysis
Jul 3, 2026Igor Buyanov, Nafisa Valieva, Ekaterina MazurinaSocial Media Analysis Text Summarization
Psytechlab
DreamerNLplus: Interpretable Modeling of Mental Health Dynamics from Social Media Timelines using Hybrid Rule-Based and RAG Methods
May 21, 2026Maryia Zhyrko, Daisy Monika Lal, Erik van Mulligen +1Social Media Analysis Mental Health
Leiden Institute of Advanced Computer Science (LIACS), Leiden University, NL · School of Computing and Communications (SCC), Lancaster University, UK · Department of Medical Informatics, Erasmus University Medical Center Rotterdam, NL +1

Aug 7, 2026J/K move · Enter open · S save

Natural Language Processing Psychometrics

Edoardo Sebastiano De Duro, Emma Franchino, Massimo Stella

CogNoscoLab, Department of Psychology and Cognitive Science, University of Trento, Italy

Clinical Outcome Prediction Psychometric Validation Computational Psycholinguistics Clinical NLP

Open paper AskPDF arXiv

Natural Language Processing (NLP) models predicting mental health outcomes rarely specify what they measure: contextual knowledge, emotional content, or syntactic structure. NLP Psychometrics treats psychological prediction from text as a psychometric problem, linking scores to interpretable linguistic evidence and testing beyond the training text format. Nine LLMs, conditioned on controlled personas (cognitive digital shadows), completed psychometric questionnaires with textual explanations per item. We extracted emotional profiles and syntactic-semantic structure via textual forma mentis networks, combined with personality and sociodemographic variables in ablated random forest (RF) regressors, using SHAP to identify which features drove performance and in which direction. Full RF models explained up to 70.8% of variance in life satisfaction (SWLS), 55.7% in depression (PHQ-9), and, for DASS-21, 68.5% depression, 76.0% anxiety, 72.4% stress. Sociodemographics alone explained no meaningful variance in depression, anxiety, or stress, but did so for life satisfaction, where emotion features and income were the strongest predictors; neuroticism and network topology instead dominated depression and anxiety, reversing direction between them. Without retraining, RF models separated diaries from low- and high-score personas (

r

up to 0.91) and, using only network/emotion features, classified clinical from control participants in real transcripts with up to 68% accuracy. These results show the promise and limits of synthetic data: LLM personas can expose model biases, recover patterns consistent with clinical rumination, and support psychometric prediction from human text without a matched questionnaire, but cannot substitute for human validation. NLP Psychometrics makes these distinctions explicit, measurable, and testable through interpretable AI and network/emotional features.

Abstract

Interpretable ML

Explore similar work

Natural Language Processing Psychometrics

psytechlab at CLPsych 2026: Utilising Natural Language Processing methods and Large Language Models for Social Media Text Analysis

DreamerNLplus: Interpretable Modeling of Mental Health Dynamics from Social Media Timelines using Hybrid Rule-Based and RAG Methods