Context-Aware Classification of Verbal Operants in Children with ASD Using Deep Learning

Abstract

Verbal operant assessment plays a critical role in autism spectrum disorder intervention planning, yet current manual evaluation methods suffer from subjectivity and time constraints. This study presents a context-aware deep learning framework for automatic classification of verbal operants (mand, tact, echoic, and intraverbal) in therapeutic speech recordings of children with ASD. The proposed multi-task learning architecture integrates contextual features including antecedent stimuli, functional consequences, and prosodic patterns through attention mechanisms. Experiments on 1,847 annotated speech samples from 52 children demonstrate classification accuracy of 83.7% for operant type identification and 89.2% for spontaneous versus prompted language discrimination. The framework successfully identifies atypical language patterns including delayed echolalia and scripted language with 81.4% precision. Results indicate that contextual feature integration improves classification performance by 12.3% compared to text-only baselines, providing objective support for language assessment and intervention planning in clinical practice.