Learning Regular Tree Languages from Correction and Equivalence Queries

Inspired by the results obtained in the string case, we present in this paper the extension of the correction queries to regular tree languages. Relying on Angluin's and Sakakibara's work, we introduce the algorithm LRTLC and we show that regular tree languages are learnable from equivalence and correction queries when the set of contexts is ordered by a Knuth-Bendix order. Moreover, a subclass of regular tree languages, called injective languages, is learned without equivalence queries. This can be extended for other subclasses and may have some practical relevance in fields like machine translation, pattern and speech recognition, building XML documents.