Abstract
Human-centric artificial intelligence needs systems that are accurate but also adaptive, computation-efficient, sustainable and applicable to real-world settings. This paper discusses Adaptive Cognitive Psychointelligence as an adaptive natural language processing system for detecting psycho-affective states expressed in text. The work investigates a fine-tuned DistilBERT language model on a multi-class mental well-being text data set with seven classes: normal, depression, suicidal thoughts, anxiety, stress, bipolar disorder and personality disorder. It proposes an adaptive cognitive framework treating affective state detection as a psychointel-ligent process where language, cognition, affective risk and responsible application are considered together. Empirical results confirm that the lightweight transformer network provides 81.99 percent accuracy and weighted F1 score of 0.819 after about 18 minutes of training time, showing that distilled language models can achieve impressive predictive performance while significantly decreasing computational load.