Dynamic construction grammar (DCG) is a neurcomputational framework for learning and generalizing sentence-to-meaning mappings. It is inspired by the cue competition hypothesis of Bates and MacWhinney, and learns regularities in the ordering of open and closed class words and the corresponding mapping to semantic roles for the open class words. The structure of meaning is a crucial aspect of these form to meaning mappings. Here we describe the DCG framework, and the evolution of meaning representations that have been addressed. The first and most basic meaning representation is a predicate-argument form indicating the predicate, agent, object and recipient. We developed an action recognition system, that detected simple actions and used naïve subjects' narration to train the model to understand. The DCG comprehension model was then extended to address sentence production. The resulting models were then integrated into a cooperative humanoid robotic platform. We then demonstrated how observed actions could be construed from different perspectives, and used the production model to generate corresponding sentences. In order to allow the system to represent and create meaning beyond the single sentence, we introduce the notions of narrative construction and narrative function word. In the same way that grammatical function words operate on relations between open class elements in the sentence, narrative function words operate on events across multiple sentences in a narrative. This motivates the need for an intermediate representation of meaning in the form of a situation model that represents multiple events and relations between their constituents. In this context we can now begin to address how narrative can enrich perceived meaning as suggested by Bruner.