I cannot find a citation quickly, but I recall years ago reading a paper about simulated agents "evolving" a language in a game context where agents had to indicate items to one another, by sending messages which were subject to a noisy channel. Items had multiple attributes (think "small red square", "big green triangle" etc), and experimenters could vary both the noise in the channel, and the entropy of the distribution over items. Naturally if "small red square" is 99% of the things you have to communicate, and there is low noise, agents invent an abbreviation for it. If there's a huge amount of noise and a relatively even distribution over items, then "small small green green triangle triangle" or similar becomes more likely. Languages very naturally reflect both the things people discuss and the environment in which they discuss them.