Speakers' preferences at choice points in language production facilitate efficient communication (at a reasonable cost) - (talk recording available)
Dr. T. Florian Jaeger
Tuesday, April 13, 2010, 01:00pm - 02:00pm
University of Rochester, Brain and Cognitive Sciences Department
In order to convey a message, speakers need to encode it linguistically. This process involves many choice points, ranging from the packaging of the message into sentences, word order and other syntactic choices (e.g. optional that-mentioning, optional object drop, etc.), via morpho-syntactic choices (e.g. contraction, as in he's vs. he is), to the phonological and phonetic realization of the message (e.g. final t/d-deletion, speech rate modulation, and even the distribution of energy across the frequency spectrum is subject to variation). The question then is how speakers decide which variant to choose whenever several variants are available.
I propose the hypothesis of Uniform Information Density, which holds that speakers' decisions at choice points during language production are affected by a preference to distribute information uniformly across the speech signal, trading-off redundancy against the amount of signal. Uniform information density is a theoretical optimal solution, in that it maximizes the amount of successfully transferred information per time (where information is defined with reference to probability distributions in information theoretic terms, Shannon 1948).
I present from phonetic reduction, the distribution of disfluencies ("uh", "I mean") and gestures, morphological contractions ("I didn't" vs. "I did not"), optional head-marking morphology in Yucatec Maya, function word omission ("I like the way (that) it vibrates."), argument drop ("German lost (the finals)"), pronominalization, and inter-clausal planning ("Gimme the beer" vs. "Do you see the beer? Please give it to me"). As different as these cases are, they all provide converging evidence for a computational principle like Uniform Information Density and hence that language production is organized to be efficient. I discuss the relation of Uniform Information Density to the notion of audience design, and in particular to ambiguity avoidance (which also plays a crucial role in bi-directional Optimality Theory).
Since information is defined by reference to probability distributions, this raises questions as to how speakers learn, maintain, and update probability distributions of linguistic structures. If time permits, I discuss some of our recent attempts to address these questions. These include studies on cue integration in spontaneous speech and experiments on short-term as well as intermediate-term adaptation of linguistic representations in response to unexpected input.
Note: If you would like to receive email announcements about the colloquium series, please contact the Business Office to have your name added to our announce lists at email@example.com.