Reinforcement and inference in cross-situational word learning

Area of Science:

• Cognitive Science
• Computational Linguistics
• Developmental Psychology

Background:

• Cross-situational word learning relies on identifying commonalities across word uses to determine referents.
• Existing models often focus on single learning mechanisms, neglecting the interplay between different strategies.

Purpose of the Study:

• To propose and validate an adaptive learning algorithm that models cross-situational word learning.
• To explain observed human learning strategies as a trade-off between reinforcement and inference.

Main Methods:

• Developed an adaptive learning algorithm with parameters for reinforcement strength and inference regulation.
• Tuned model parameters to match data from diverse cross-situational word learning experiments.
• Analyzed the model's ability to predict human learning strategies under different experimental conditions.

Main Results:

• The algorithm successfully explains learning strategies by balancing reinforcement and inference.
• Demonstrated that inference dominates in fast mapping, while reinforcement is key in segregated contextual diversity.
• Showed that other conditions utilize a mix of both reinforcement and inference.

Conclusions:

• Human cross-situational word learning strategies arise from a dynamic trade-off between reinforcement and inference.
• The proposed algorithm provides a unified framework for understanding diverse word learning behaviors.
• The model's parameters offer insights into the cognitive mechanisms underlying word acquisition.