Kanazawa has shown that several non-trivial classes of categorial grammars are learnable in Gold’s model. We propose in this article to adapt this kind of symbolic learning to natural languages. In order to compensate the combinatorial explosion of the learning algorithm, we suppose that a small part of the grammar to be learned is given as input. That is why we need some initial data to test the feasibility of the approach: link grammars are closely related to categorial grammars, and we use the English lexicon which exists in this formalism.