The MIT Linguistics Department is pleased to announce the first linguistics colloquium of the spring semester:
Speaker: T. Florian Jaeger (University of Rochester)
Title: Efficient Language Production?
Time: Friday, February 27th, 2009, 3:30pm
Place: 32-141
Abstract:
In this talk, I return to a question that has fascinated language researchers from various disciplines for a long time (Zipf, 1929; Mandelbrot, 1965; Givon, 1987; Hawkins, 1994; Hale, 2001; Bybee, 2002; Genzel & Charniak, 2002; Manin, 2006, among many others), although it has arguably never been at the heart of language research: To what extent is human language processing efficient?
More specifically, I ask to what extent speakers structure their utterances so as to
be communicatively efficient. I present a series of studies that test the Uniform
Information Density hypothesis (Jaeger, 2006; Levy & Jaeger, 2007; based on Genzel
& Charniak, 2002): Within the bounds defined by grammar, speakers prefer to
structure their utterances so that information is distributed uniformly across the
signal (information density; where the information content of a linguistic unit is
defined information theoretically, Shannon, 1948, as -log p(unit)). Where speakers
can choose between several variants to encode their message, they prefer the variant
with more uniform information density. Uniform Information Density is theoretically
optimal in that it maximizes the amount of successfully transferred information and
minimizes average processing load.
I discuss evidence from phonetic, phonological, morphosyntactic, and syntactic
reduction (word durations; weak vs. full vowels; t/d deletion; contractions such as he’s
vs. he is; whiz-deletion in passive subject-extracted relative clauses), as well as studies
on discourse planning beyond the level of the clause. I also present new experimental
evidence from the distribution of disfluencies and gestures in information dense
stretches of speech. The results of all these studies lend to support to the hypothesis
that language production is organized to be efficient. When encoding their intended
message into linguistic utterances, speakers are sensitive to the information density
of the variants they can choose from.
(In collaboration with: Susan Wagner Cook, Austin Frank, Carlos Gomez Gallo, Ting Qian, and Matt Post)