Our research investigates if the human parser is sensitive to the underlying structure governing the simplest Lindenmayer grammar [1] consisting of two rewrite rules : 01, 101. This grammar is called Fibonacci (Fib) because at each generation the number of symbols (0s and 1s) follows the Fibonacci sequence (0, 1, 1, 2, 3, 5, 8,. . . ). The analysis of the formal properties of Fib-generated strings shows that whereas some symbols are fully ambiguous and thus cannot be linearly anticipated, others (called k-points) can be predicted if the parser relies on the hierarchical structure of this grammar [2]. In three experiments we explored whether humans are able to learn Fib and more particularly to track kpoints. We used a serial reaction time paradigm in which participants had to press as quickly as possible the button corresponding to the symbol on the screen. In experiment 1, participants learned the Fib grammar. The learning slope of k-points was significantly steeper than that of ambiguous symbols. However, since k-points were more frequent than ambiguous symbols, this advantage may reflect the tracking of surface statistical regularities rather than structural learning. Experiment 2 used the Skip grammar (001; 101101) in which the statistical distribution of k-points was reversed such that they were less frequent than ambiguous symbols.

The learning slopes were reversed, showing a significant advantage for ambiguous symbols. Experiment 3 tested how participants switch from learning one grammar in the first part of the experiment (Fib or Skip) to learning the other grammar in the second part. Different curves were observed for the two groups, suggesting that unlearning Fib is more difficult than unlearning Skip. This finding suggests that the advantage for k-points in Experiment 1 cannot be attributed only to the tracking of surface statistics, and that the parser extracts, to some extent, the underlying structure of the grammar.