Pragmatically Framed Cross-Situational Noun Learning Using Computational Reinforcement Models.

Area of Science:

• Artificial intelligence
• Cognitive science
• Computational linguistics

Background:

• Word learning is crucial for cognitive development.
• Cross-situational learning and social pragmatic theories explain word-object association.
• Reinforcement learning offers a computational framework for modeling learning mechanisms.

Purpose of the Study:

• Investigate reinforcement's role in early word learning for an artificial agent.
• Model word-object pairing using both cross-situational learning and social pragmatic cues.
• Compare the performance of different reinforcement learning algorithms.

Main Methods:

• Utilized neural network-based reinforcement learning (Q-NN, NFQ, DQN) over table-based methods.
• Incorporated joint attention and prosodic cues as social signals.
• Trained models on the CHILDES dataset simulating mother-infant interactions.
• Applied the "novel words to novel objects" constraint for reward computation.

Main Results:

• Neural network models demonstrated better generalization and faster convergence than table-based models.
• Attentional-prosodic models outperformed attentional models in word learning.
• The Attentional-prosodic Deep Q-Network (DQN) achieved superior performance compared to existing models.

Conclusions:

• Social cues, particularly prosody combined with joint attention, significantly enhance artificial agent word learning.
• Deep Q-Networks provide an effective architecture for sophisticated word-learning models.
• This research advances computational models of early language acquisition.