Information density and the rational speaker: How do speakers manipulate the information density?

Jiří Zámečník

Hermann Paul School of Linguistics Basel – Freiburg (i.Br.)

In their everyday communication, speakers are constantly aware of the communicative situation. They adjust their output in response to their interlocutor or the environment they are in; moreover, two speakers in an identical situation may not necessarily shape a given message in the same way. Knowing that speakers constantly manipulate their output, we would expect to find some regularity governing this manipulation. Ideally, this regularity would seem rational to us in the sense of optimally using the communicative channel to transmit information.

Thus, a rational speaker would not only smoothen the information density in their output as suggested by the Uniform Information Density hypothesis. They would also ensure that the average information density is as high as the channel allows, but at the same time avoid exceeding its capacity.
Given this assumption, I expect the speakers to manipulate the average information density in correspondence to the mode of communication (writing should allow for higher information density than speech) or to their interlocutor (some interlocutors should be capable of receiving more densely packed messages than others). In the present talk, I will explore the first assumption on the basis of corpus data using the integrated measure (Mitchell 2011) to estimate the information density.

In addition to this exploration, a follow-up study will be presented, searching to provide insight into the small-scale information density manipulation employed by the speakers in the form of speech disfluency, proposing an information theoretic view of disfluencies inspired by the Noisy Channel approach.

Mitchell, J.J. (2011). Composition in distributional models of semantics. Ph.D. thesis, University of Edinburgh, Edinburgh.

If you would like to meet the speaker, please contact Irina Stenger.