Learning Object Names in Real Time with Little Data

We present an online learning model of early cross-situational word learning which maps words to objects from context with relatively sparse input. The model operates by rewarding and penalizing probabilities of possible word-to-object mappings based on real-time observation, and using those probabilities to determine a lexicon. We integrate prosodic and gestural cues and allow the learner to evaluate lexical entries. These enrichments allow efficient learning with minimal computational effort, producing results comparable to that of more complex models.