The question ``What does it mean to learn language?'' isone of the greatest topics of scientific inquiry. The problem hasattracted many researchers in linguistics, cognitive science, and computerscience and attempts to answer this question vary greatly from disciplineto discipline.In this talk, I will consider language learning as the problem oflearning grammars which capture syntax and semantics and which allowaccess to meaning during language processing (i.e., parsing andgeneration). I will introduce a new grammar formalism whichintegrates syntax and semantics, and is learnable. I will present acomputationally efficient model for learning this type of grammars fromrepresentative data, which consists of pairs of strings and their semanticrepresentations. Last but not least, I will discuss the linguisticrelevance of these grammars, and the usefulness of this learning model forthe acquisition of knowledge from text.