Mutual Exclusivity in Pragmatic Agents.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Developmental Psychology

Background:

• Word learning presents a challenge due to multiple potential referents in context.
• Children exhibit a mutual exclusivity (ME) bias, associating novel words with novel referents, aiding disambiguation.

Purpose of the Study:

• To investigate the conditions under which pragmatic agents exhibit a mutual exclusivity (ME) bias.
• To develop a computational framework for understanding the pragmatic basis of the ME bias in human word learning.
• To build artificial agents capable of displaying an ME bias.

Main Methods:

• Utilized the Rational Speech Act model combined with gradient-based learning for semantic learning in pragmatic agents.
• Analyzed agent behavior across various aspects of word learning, including online inference, long-term learning, and developmental effects.
• Tested different model implementations to identify crucial components for simulating the ME bias trajectory.

Main Results:

• Demonstrated striking parallels between the model's behavior and human word learning concerning the ME bias.
• Identified pragmatic online inference and incremental evidence collection for word-referent correspondences as key factors in modeling the ME bias's developmental trajectory.
• Developed a novel deep neural network architecture capable of processing naturalistic input without requiring indirect access to test inputs during training to display ME bias.

Conclusions:

• The developed pragmatic agent model successfully replicates key aspects of human mutual exclusivity (ME) bias in word learning.
• The findings highlight the importance of pragmatic inference and evidence accumulation in the developmental emergence of the ME bias.
• This research offers the first deep neural network model to exhibit an ME bias without task-specific training manipulations, advancing artificial intelligence in language acquisition.