Grammatical Inference as Class Discrimination

Grammatical inference is typically defined as the task of finding a compact representation of a language given a subset of sample sequences from that language. Many different aspects, paradigms and settings can be investigated, leading to different proofs of language learnability or practical systems. The general problem can be seen as a one class classification or discrimination task. In this paper, we take a slightly different view on the task of grammatical inference. Instead of learning a full description of the language, we aim to learn a representation of the boundary of the language. Effectively, when this boundary is known, we can use it to decide whether a sequence is a member of the language or not. An extension of this approach allows us to decide on membership of sequences over a collection of (mutually exclusive) languages. We will also propose a systematic approach that learns language boundaries based on subsequences from the sample sequences and show its effectiveness on a practical problem of music classification. It turns out that this approach is indeed viable.