Abstract

Much recent research into rule and grammar learning has focussed on the potential role of statistical regularities in recognising grammatically determined patterns. Considerable emphasis has been placed on transition probabilities in driving the perception of segmentation in strings of syllables generated by Markovian grammars and the subsequent ability to discriminate between “legal” and “illegal” strings. An alternative approach has stressed the ability to recognize and extract algebraic (variable dependent) regularities in similar artificial strings. In this talk I will present the results of a recent series of investigations into our ability to recognise and discriminate grammatical series generated by recursive grammars which generate strings with no local transition probabilities and whose output cannot be easily characterized in algebraic terms. The observation that humans are very good at these tasks raises an number of important questions about the nature of statistical analyses being commonly appealed to and the principles underlying grammatical learning.