Interpretable neuro-affective language understanding for psychological and psychometric assessment towards clinical-support applications

Human language conveys interacting semantic, affective, behavioral, psychological and risk-related information that is commonly modeled through isolated natural-language-processing tasks. This paper presents NALU (Neuro-Affective Language Understanding), an interpretable multi-task language-representation framework designed to jointly model multiple complementary dimensions of psychologically relevant language. The framework combines a shared Transformer encoder with specialized prediction heads, element-level supervision masks, validation-based threshold calibration and an auditable deterministic interpretation layer. The study was developed in two experimental stages. The initial configuration used 205,109 samples and eight specialized prediction heads comprising exactly 45 scalar outputs. The framework was then extended and evaluated on an audited corpus containing 589,155 unique samples, including 395,417 Brazilian-Portuguese and 193,738 English records. The expanded multilingual configuration uses XLM-RoBERTa and twelve specialized heads covering sentiment, emotion, valence–arousal–dominance, psychological constructs, behavior, mental-health risk, neurodevelopmental-language indicators, psychiatric-language indicators, addiction/eating/cognition-related language, electronic-risk indicators, psychometric items and psychometric scores. For the initial eight-head configuration, the calibrated model achieved a composite test score of 0.761 with a 95% bootstrap confidence interval of [0.754, 0.767]. Macro-F1 reached 0.874 for psychological constructs, 0.783 for mental-health risk and 0.769 for clinical-language markers. Validation-tuned thresholds significantly improved emotion, psychological, behavioral, clinical and overall composite performance relative to a fixed 0.5 threshold ( p < 0.001), whereas the improvement for the risk head was positive but not statistically significant ( p = 0.096). The expanded 589,155-sample XLM-R configuration achieved a sentiment Macro-F1 of 0.7912 and emotion Macro-F1 of 0.6858. Threshold-free discrimination was high for several multi-label heads, with AUROC values of 0.9659 for emotion, 0.9838 for psychological constructs, 0.9694 for behavior, 0.9810 for mental-health risk, 0.9899 for psychiatric-language indicators and 0.9963 for addiction/eating/cognition-related language. Continuous prediction errors were MAE = 0.1009 for valence–arousal–dominance, MAE = 0.0109 for psychometric items and MAE=0.0280 for psychometric scores. Outcomes without sufficient positive held-out support are explicitly reported as not estimable rather than interpreted as validated zero performance. NALU is not presented as a diagnostic instrument. Psychometric outputs are interpreted as questionnaire-associated or language-derived latent estimates, and clinical-support outputs are intended for screening-oriented research and human-in-the-loop analysis.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-13
DOI
https://doi.org/10.1038/s41598-026-73224-0
Primary Topic
Mental Health via Writing
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Interpretable neuro-affective language understanding for psychological and psychometric assessment towards clinical-support applications

Diego R. Faria
2 links
Scientific Reports
Mental Health via Writing
article

Interpretable neuro-affective language understanding for psychological and psychometric assessment towards clinical-support applications

Diego R. Faria
article en

Abstract

Human language conveys interacting semantic, affective, behavioral, psychological and risk-related information that is commonly modeled through isolated natural-language-processing tasks. This paper presents NALU (Neuro-Affective Language Understanding), an interpretable multi-task language-representation framework designed to jointly model multiple complementary dimensions of psychologically relevant language. The framework combines a shared Transformer encoder with specialized prediction heads, element-level supervision masks, validation-based threshold calibration and an auditable deterministic interpretation layer. The study was developed in two experimental stages. The initial configuration used 205,109 samples and eight specialized prediction heads comprising exactly 45 scalar outputs. The framework was then extended and evaluated on an audited corpus containing 589,155 unique samples, including 395,417 Brazilian-Portuguese and 193,738 English records. The expanded multilingual configuration uses XLM-RoBERTa and twelve specialized heads covering sentiment, emotion, valence–arousal–dominance, psychological constructs, behavior, mental-health risk, neurodevelopmental-language indicators, psychiatric-language indicators, addiction/eating/cognition-related language, electronic-risk indicators, psychometric items and psychometric scores. For the initial eight-head configuration, the calibrated model achieved a composite test score of 0.761 with a 95% bootstrap confidence interval of [0.754, 0.767]. Macro-F1 reached 0.874 for psychological constructs, 0.783 for mental-health risk and 0.769 for clinical-language markers. Validation-tuned thresholds significantly improved emotion, psychological, behavioral, clinical and overall composite performance relative to a fixed 0.5 threshold ( p < 0.001), whereas the improvement for the risk head was positive but not statistically significant ( p = 0.096). The expanded 589,155-sample XLM-R configuration achieved a sentiment Macro-F1 of 0.7912 and emotion Macro-F1 of 0.6858. Threshold-free discrimination was high for several multi-label heads, with AUROC values of 0.9659 for emotion, 0.9838 for psychological constructs, 0.9694 for behavior, 0.9810 for mental-health risk, 0.9899 for psychiatric-language indicators and 0.9963 for addiction/eating/cognition-related language. Continuous prediction errors were MAE = 0.1009 for valence–arousal–dominance, MAE = 0.0109 for psychometric items and MAE=0.0280 for psychometric scores. Outcomes without sufficient positive held-out support are explicitly reported as not estimable rather than interpreted as validated zero performance. NALU is not presented as a diagnostic instrument. Psychometric outputs are interpreted as questionnaire-associated or language-derived latent estimates, and clinical-support outputs are intended for screening-oriented research and human-in-the-loop analysis.

Scientific Reports
Loughborough University (GB)
Good health and well-being
Mental Health via Writing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon