University of Buffalo

Abstract:
Elements of linguistic messages that are more predictable, either overall or in context, tend to have shorter forms than elements of linguistic messages that are longer. Zipf (1949) proposed that this tendency facilitates communication by giving typical messages shorter linguistic forms than atypical messages. More recent work has argued that this tendency reflects adaptation towards efficient communication in the sense of the Noisy Channel Theorem. However, others have argued that these effects are, in the terminology of Gould and Lewontin (1979), “spandrels:” non-adaptive side effects of other processes. In this talk, I will present work that develops this information-theoretic framing more fully. First, I will show that it is unlikely that natural language actually gets close to the information-theoretic efficiency bounds. Second, I will show how increased efficiency in the sense of the Noisy Channel theorem requires very different coding operations. Third, I will present corpus studies that find evidence for these very different coding operations. Because these operations are very different, if they are spandrels, they would probably be independent spandrels, an unlikely coincidence. Thus, while previous work has been too optimistic in proposing that predictability effects actually achieve theoretical efficiency bounds, we do find evidence for adaptation in the direction of theoretical efficiency bounds.