Language Augmented Prediction

of a sentence containing words that begin with those letters, or learning to tie a knot by thinking of a rabbit going in and out of a hole), and (2) explicit verbal mediation, i.e., “thinking in words.” Indeed, this introspection of thinking in words is often so strong that it leads researchers to conflate that feeling of talking to oneself with the format of conceptual representations (Ryle, 1968; e.g., Carruthers, 2002; Levinson, 1997 for discussion). This confusion can be clarified by considering the role language can play in generating top-down predictions (Lupyan, 2012a,b for discussion). A growing body of work suggests that language interfaces directly with the surprisal-reducing machinery at the core of predictive-coding models. Consider a task in which one hears an auditory cue (e.g., a barking sound) and then sees a picture (e.g., a dog). The goal is to respond “yes” if the cue and picture match at a conceptual level, and “no” otherwise (e.g., a car following a barking sound). The better the match between the top-down predictive signal and the bottom-up activation produced by the probe, the faster (or more accurately) subjects can respond. Lupyan and Thompson-Schill (2012) found that linguistic cues (“dog”) were more effective than non-linguistic cues (e.g., a barking sound, a car horn), even though both cue types were judged as equally predictive and unambiguous of the associated category. As the delay between the cue and probe was increased, the difference between the verbal and non-verbal-cue conditions also increased. Under the influence of the label (through hypothesized top-down effects), the resultant representations appeared to become more similar across subjects with increasing delays in a way that they did not on trials without the verbal label. This provides a basic demonstration of how verbal A commentary on