Why is language well-designed for communication? (Commentary on Christiansen and Chater: 'Language as shaped by the brain')

Selection through iterated learning explains no more than other non-functional accounts, such as universal grammar, why language is so well-designed for communicative efficiency. It does not predict several distinctive features of language like central embedding, large lexicons or the lack of iconicity, that seem to serve communication purposes at the expense of learnability. Christiansen and Chater rightfully observe that communicatively arbitrary principles, such as UG, are unable to explain why language is adequate for communication. The same criticism can be addressed, however, to their own account. If the main driving force that led to language emergence is learnability rather than communicative efficiency, language should be locally optimal for the former and not for the latter. Evidence suggests that, in several respects, the exact opposite is the case. What would language be like if, as the authors claim, the cultural selection of learnable languages were “stronger” than the biological selection of brains designed for efficient communication? If language can compare with a “viral” entity that gets selected for its ability to resist vertical cultural transmission, we predict for instance iconic signifiers, especially gestures, to win the contest. Yet, although analogue resemblance makes learning almost trivial, linguistic evolution shows that non-iconic signifiers tend to prevail, even in sign languages. The “viral” theory of language does not explain the size of lexicons either. Ideally, an expressive code is easiest to learn, and resists iterated transmission best, if words are limited in number and have separate and unambiguous meanings. Yet, real vocabularies include tens of thousands of words, massive near synonymy and many rare unpredictable word combinations (Briscoe, 2006). Such evidence suggests that there may be some “viral” cause for the existence of plethoric lexicons, but its action is opposite to what is expected from selection for learning efficiency. Language, as mainly shaped by selection through repeated learning, is supposed to mirror the general human induction bias. Efficient induction systems (Solomonoff, 1978), including human learning (Chater, 1999) and analogy making (Cornuéjols, 1996) are guided by a complexity minimization principle. If languages were the bare expression of a simplicity-based induction device looping on itself, we should expect the complexity of languages to converge to a minimal amount. A similar claim is that general-purpose learning devices, except in rote learning mode, produce only “good shapes” (Gestalten),