Distributional learning of recursive structures is constrained by structural representation

This work investigates the mechanism of learning recursive structures. The ability of recursion is considered crucial for language and universally available, but there are considerable with- and cross-linguistic differences regarding the rules for recursive embedding, which poses a learnability problem. Previous research has shown that adults can learn the recursivity of linear structures in an artificial language from distributional cues in non-recursively embedded data. However, an important open question is participants’ structural representation of the grammar, which is considered crucial for linguistic recursion. In this study, we examined the hypothesis that representation of the structural head is necessary for the distributional learning of recursive structures. Adult participants were exposed to one of the two artificial languages which had identical linear orders but different heads. At test, participants were asked to rate test strings which examined their knowledge of the head and recursion. As predicted, we found that the learning mechanism is constrained: participants who learned the language with the correct head were more likely to allow recursive embedding. The findings suggest that human learners need structural representations beyond surface-level distributional cues to acquire recursive structures.