This thesis focuses on a challenging and long-standing problem of learning from language, in other words, how humans or machines may use language to share and acquire knowledge. The work has three distinct parts. First, I review how different disciplines define and approach the problem of learning from language and argue that a number of areas in Cognitive Science and Computer Science research have recently advanced enough to begin to tackle this challenge. Second, I present a series of three behavioral experiments studying the problem of learning from language in the context of pedagogical category communication. The experiments demonstrate the flexibility of verbal communication as a means for sharing category knowledge, as well as the advantage of mixing communication media (verbal and exemplar-based) as opposed to relying on any one isolated channel. In the last part of the dissertation, I focus on the question of how modern AI architectures can be adapted and applied to the problem of lifelong learning from language. In particular, I identify the types of operations that the model should be able to make, and propose a training procedure and an architecture that support learning such operations in an end-to-end fashion. I test the architecture on a number of simulated non-linguistic domains, leaving its NLP applications to future research. Although it is only a small step towards creating a fully functioning learning from language model, I still believe that this step is important.