Bayesian tree substitution grammars as a usage- based approach.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Grammar Formalisms

Background:

• Context-free grammars (CFGs) have limitations in representing linguistic structures.
• Tree substitution grammars (TSGs) generalize CFGs by allowing larger rewrite fragments.
• Usage-based approaches in linguistics require computational models for validation.

Purpose of the Study:

• To develop a computational model for learning Tree substitution grammars (TSGs).
• To address the challenge of fragment set size in TSG learning.
• To evaluate the performance of learned TSGs in linguistic tasks.

Main Methods:

• A model-based approach using Gibbs sampling with a non-parametric prior was employed.
• The method controls fragment size, favoring smaller fragments while allowing larger ones.
• Grammars were learned and evaluated on parsing accuracy and grammaticality classification.

Main Results:

• Bayesian TSGs achieved excellent performance on parsing and grammaticality tasks.
• The learned grammars balance fragment size effectively.
• Performance was competitive with existing TSG and Data-Oriented Parsing models.

Conclusions:

• The proposed Bayesian TSG learning method is effective for computational modeling of linguistic theories.
• This approach provides a practical solution to the fragment selection problem in TSGs.
• The results demonstrate the utility of TSGs in capturing linguistic generalizations.