Synthesizing theories of human language with Bayesian program induction.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Cognitive Science

Background:

• Automated scientific model construction is a key challenge in AI.
• Understanding morpho-phonology, the system for word forms from sounds, is crucial for language analysis.

Purpose of the Study:

• To develop a framework for algorithmically synthesizing interpretable models of human language morpho-phonology.
• To integrate Bayesian inference, program synthesis, and cognitive models for linguistic discovery.

Main Methods:

• The system uses Bayesian inference and program synthesis to create morpho-phonological models.
• It incorporates linguistic theory and cognitive models of learning and discovery.
• The framework is tested on 70 datasets from 58 diverse languages.

Main Results:

• The framework successfully synthesizes human-interpretable morpho-phonological models for various languages.
• A meta-model capturing cross-language typological tendencies is automatically generated.
• The algorithm demonstrates few-shot learning capabilities, acquiring new rules from minimal examples.

Conclusions:

• The developed framework enables machine-assisted discovery of interpretable models in linguistics.
• This approach suggests new avenues for AI-driven scientific discovery in various domains.
• The system's ability to handle diverse languages and few-shot learning highlights its potential.