Exploring the Feasibility of Intent Recognition for Pre-Schoolers – Boluwaji Fagbuaro

One of the primary barriers to developing AI systems capable of interpreting toddler communication is the scarcity of ethically usable child-language datasets. Toddlers between the ages of one and four rarely communicate through full grammatical structures; instead, meaning is conveyed through short phrases, repeated sounds, vocal tone variation, and pauses. From a computational perspective, this represents a fundamentally different challenge from conventional conversational AI. The core problem is not dialogue generation, but intent recognition under high linguistic noise.

To explore the feasibility of intent recognition in early childhood communication, a lightweight transformer-based classifier was trained on a synthetic child-language dataset designed to approximate pre-linguistic and early-linguistic patterns. The dataset consisted of approximately 55,000 short, fragmented utterances mapped to coarse-grained intent categories such as need, comfort, imitation, and affirmation. The objective was not to simulate real child speech, but to test whether intent-level abstractions could be learned under simplified and noisy conditions.

Preliminary experiments produced a baseline validation accuracy of approximately 73% within this constrained synthetic setting. While these results do not imply readiness for real-world use, they suggest that intent-level modelling may be a viable abstraction for early childhood communication, even when language is incomplete. Further investigation using ethically sourced real-world child-language data would be required to evaluate robustness, generalisation, and safety.

Any consideration of applied child-facing systems must account for developmental and ethical risks, including anthropomorphism, emotional attachment, and the potential for misplaced trust. These concerns necessitate cautious, transparent design rather than claims of autonomy or social equivalence.

Existing child-facing robotic systems have demonstrated the value of embodied interaction in education and social development. For example, Moxie, a now-discontinued social robot, was designed to support emotional, social, and cognitive skill development through structured conversation, play, and expressive behaviours such as facial expressions and body movement. However, such systems primarily targeted older children, typically aged five years and above, whose linguistic abilities are comparatively stable. This distinction highlights an underexplored research space: AI systems centred on intent inference rather than dialogue flow, designed specifically for pre-linguistic and early linguistic age groups.