What Homophones Say about Words

Area of Science:

• Cognitive Science
• Developmental Psychology
• Linguistics

Background:

• Word learning requires restricting the vast hypothesis space of potential meanings.
• Current models often use constraints like convexity, assuming concepts and word meanings are continuous.

Purpose of the Study:

• To investigate how adults handle non-convex sets of objects when learning novel labels.
• To determine if adults infer homophony or disjunctive meanings in such cases.
• To inform computational models of word learning.

Main Methods:

• Two experiments were conducted presenting participants with novel labels for objects.
• Experimental conditions manipulated the conceptual space between learning exemplars to create gaps.
• The intervention of other lexical items in these gaps was also manipulated.

Main Results:

• Adults overwhelmingly preferred to postulate homophony (one word, two meanings) when presented with disjoint sets of objects.
• This preference occurred whether the gap was in conceptual space or due to lexical intervention.
• Participants avoided inferring disjunctive (non-convex) meanings for a single word form.

Conclusions:

• The convexity constraint in word learning is too strict and needs revision.
• Homophony is a preferred interpretation strategy for non-convex meaning spaces.
• Word learning algorithms must incorporate findings on homophony and avoid over-specialization on limited data.