Model Merging versus Model Splitting Context-Free Grammar Induction

When comparing different grammatical inference algorithms, it becomes evident that generic techniques have been used in different systems. Several finite-state learning algorithms use state-merging as their underlying technique and a collection of grammatical inference algorithms that aim to learn context-free grammars build on the concept of substitutability to identify potential grammar rules. When learning context-free grammars, there are essentially two approaches: model merging, which generalizes with more data, and model splitting, which specializes with more data. Both approaches can be combined sequentially in a generic framework. In this article, we investigate the impact of different approaches within the first phase of the framework on system performance.