Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language

(Kirby et al., 2008) at PNAS

Today’s paper relates to yesterday’s by looking at humans in the process. It talks about language as a culturally constructed artifact. Every time we produce an utterance, we convey not only the meaning behind the utterance, but also details about the artifact—the nature of the language.

One claim from yesterday’s paper is that repeated transmission, with small population, leads to more complex structures. Today’s paper asserts that prior work makes the claim without qualifying the population.

The authors focus on diffusion again, but instead of label propagation, they use humans practicing “diffusion chains”: they observe a behavior, then demonstrate it for another participant, who demonstrates for another. It’s like a game of telephone.

The humans are exposed to visual stimuli: objects in motion, paired with words in an “alien language” to describe it. If the language does adapt as intended, it will become more learnable, and it will become more structured.

They use Levenshtein distance as their metric of distance, but they normalize it between 0 and 1. This…this should be impossible, unless they’re under-specifying something like an assumption about word lengths.

Anyway, it sounds very familiar: they need to avoid the situation of rote memorization à la Funes. People wound up using morphological concatenation. (Hooray, people found something in something designed to show them exactly that. I don’t like tests which get shocked that you found something, when the test is designed for you to find it.)

Written on March 6, 2018