Pattern Extraction Improves Automata-Based Syntax Analysis in Songbirds

Abstract

We propose an automata induction approach to modeling birdsongs on the basis of Angluin’s induction algorithm, which ensures that k-reversible languages can be learned from positive samples in polynomial time. There are similarities between Angluin’s algorithm and the vocal learning of songbirds; for example, during a critical period, songbirds learn songs from positive samples of conspecific birds. In our previous method, we could not construct song syntaxes for complex songs. In this paper, we introduce a pattern extraction method to improve our previous method and propose a new birdsong modeling method. We estimate the robustness and properness of our method by using artificial song data, and demonstrate that the song syntaxes of the Bengalese finch can be successfully represented as reversible automata. As a result, almost all Bengalese finchs’ song syntaxes can be represented with lower k-reversibility; further, one song has 3-reversibility song syntax showing the highest reversibility.