On Restrictions in Computational Language Learning

In 1990 Fulk proved that partially set-drivenness (rearrangement-independence) does not weaken the power of unrestricted computational language learning. The question arises whether this result still holds if paired with various learning restrictions. We investigate the influence of two main categories of such restrictions, namely content-based and delayable ones. An adaption of Fulk’s theorem is verified for content-based learning and some delayable restrictions regarding U-shaped learning. On the other hand, we give an example criterion of delayable learning — explanatory learning from text by a strongly monotone scientist — for which partially set-drivenness does reduce the learning power. Additionally, the interdependence of these restrictions with several other interaction operators and success criteria are explored.