A Multimodal Model of Child Language Acquisition at the One-Word Stage

We present a multimodal neural network model of child language acquisition at the one-word stage that is inspired by current views on brain processing which suggest that information in the brain is ultimately stored in a common amodal set of conceptual representations. Our model takes into account a child's perceived communicative intention and simulates both ostensive and non-ostensive utterances using the same unsupervised network. This is contrary to some models of child language acquisition, and in line with current psycholinguistic views on child language processing at the one-word stage. When trained on natural child language data, our model learns to categorise the speech utterances and to generalise to new utterances. In addition, our model exhibits a U-shaped learning trajectory as well as a vocabulary spurt, in agreement with observations made in studies of child language acquisition at the one-word stage.