Trying to understand human language by constructing robots that have language necessarily implies an embodied view of language, where the meaning of linguistic expressions is derived from the physical interactions of the organism with the environment. The paper describes a neural model of language according to which the robot's behaviour is controlled by a neural network composed of two sub-networks, one dedicated to the non-linguistic interactions of the robot with the environment and the other one to processing linguistic input and producing linguistic output. We present the results of a number of simulations using the model and we suggest how the model can be used to account for various language-related phenomena such as disambiguation, the metaphorical use of words, the pervasive idiomaticity of multi-word expressions, and mental life as talking to oneself. The model implies a view of the meaning of words and multi-word expressions as a temporal process that takes place in the entire brain and has no clearly defined boundaries. The model can also be extended to emotional words if we assume that an embodied view of language includes not only the interactions of the robot's brain with the external environment but also the interactions of the brain with what is inside the body.