Nouns refer to shapes, adjectives to properties: emergent biases in early vocabulary development

Althaus, N. & Plunkett, K.

University of Oxford

How do infants learn that nouns refer to object kinds, whereas adjectives refer to properties (e.g. Landau, Smith & Jones, 1992)? Computational modeling showed that the tendency to map nouns to objects based on shape could emerge from word-object co-occurrences (Samuelson, 2002).
However, it is unclear whether input correlations can account for a bias that causes even novel words to be mapped to shapes or properties accordingly.
We propose a computational approach that combines acquisition of syntactic categories in a simple recurrent network (SRN; Elman, 1990) and Hebbian learning of links between self-organizing maps (SOMs; Kohonen, 1982) of lexicon, shape and properties.
An SRN was trained with simple phrases (e.g. “The black cat sleeps”), presenting one word at a time, and using the next word as the target output (Elman, 1990). Syntactic representations emerged across the hidden units due to the given sequential input probabilities. The word representations obtained this way were then presented to a network of interconnected self-organizing maps, which simultaneously received property and shape information for the words’ referents. This network consisted of three SOMs for words, shapes and properties, fully interconnected by Hebbian links. Besides SOM learning - establishing clustering of words, properties and shapes in the individual maps - Hebbian links were strengthened between coactivated units.
After training, Hebbians correctly mapped words onto their referents - adjectives to properties, nouns to shapes. In addition, the model also generalized these links to novel nouns and adjectives, which were distinguished as such merely by their syntactic context.
This shows that pathways linking adjectives to properties, nouns to shape, can emerge based on the infants’ linguistic and visual environment.
We will discuss the implications of these findings both with regard to the role of words as object features and the emergence of a lexical/semantic network on the basis of co-occurrence.