Very efficient learning of structured classes of subsequential functions from positive data

Adam Jardine (1) , Jane Chandlee (1) , Rémi Eyraud (2) , Jeffrey Heinz (1)

1 Dept. of Linguistics and Cognitive Science
2 QARMA - éQuipe AppRentissage et MultimediA [Marseille]

In this paper, we present a new algorithm that can identify in polynomial time and data using positive examples any class of subsequential functions that share a particular finite-state structure. While this structure is given to the learner a priori, it allows for the exact learning of partial functions, and both the time and data complexity of the algorithm are linear. We demonstrate the algorithm on examples from natural language phonology and morphology in which the needed structure has been argued to be plausibly known in advance. A procedure for making any subsequential transducer onward without changing its structure is also presented.

Mots clés: transduction; onwardness; identification with polynomial bounds

Domaines: Apprentissage [cs.LG]

Adam Jardine, Jane Chandlee, Rémi Eyraud, Jeffrey Heinz. Very efficient learning of structured classes of subsequential functions from positive data. The 12th International Conference on Grammatical Inference, Sep 2014, Kyoto, Japan. pp.94-108. ⟨hal-01193057⟩