A neural network architecture for learning word-referent associations in multiple contexts.

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Psycholinguistics

Background:

• Understanding how humans learn word meanings from context is a key challenge in cognitive science.
• Existing models often struggle to account for the contextual ambiguity inherent in word acquisition.

Purpose of the Study:

• To propose a biologically inspired neurocomputational architecture for learning word-referent associations.
• To model the incremental learning of word meanings in varied contexts.

Main Methods:

• Developed a multi-layered architecture processing object images and word phonemes.
• Utilized a Self-Organizing Map for incremental association node creation, adjustment, and removal.
• Incorporated contextual information to disambiguate word meanings.

Main Results:

• Achieved up to 78% accuracy in word-referent association in ambiguous situations.
• Simulated learning rates closely approximated human learning rates from cross-situational word learning experiments.
• Demonstrated similar learning patterns across different experimental conditions.

Conclusions:

• The proposed architecture effectively models cross-situational word learning.
• The model's ability to handle ambiguity and mimic human learning rates offers insights into language acquisition.
• This neurocomputational approach provides a framework for further research in computational psycholinguistics.