Degen, J. 1 , Hawkins, R. X. 1 , Graf, C. 2 & Goodman, N. D. 1
1 Stanford University
2 University of Osnabrueck
Speakers have a well-documented tendency to add redundant adjectives in referring expressions. This "overinformativeness" poses a challenge for theories of language production, especially those positing rational language use. We present a novel production model of referring expressions in the Rational Speech Acts framework that accounts for this asymmetry. Speakers are modeled as rationally trading off cost of additional adjectives with amount of information added about the target object. The innovation is assuming that adjectives' truth functions are differentially noisy -- specifically, that size adjectives are noisier than color adjectives.
Not only does the model capture the color-size asymmetry, it predicts fine-grained interactions between amount of visual variation among objects and the dimension sufficient for establishing reference (color vs size) on speakers' probability of redundant adjective use: increasing scene variation results in much more greatly increased redundant color than size adjective use. These predictions were borne out in an interactive reference game experiment where 60 pairs of participants referred to objects in varying contexts through a chat window over the web (R^2 between model posterior predictive and data=.94).
We conclude that the systematicity with which speakers redundantly use adjectives implicates a system geared towards communicative efficiency rather than towards wasteful overinformativeness.