Contraction use as a measure of speaker choice in spontaneous speech

Overview

One useful source of data for testing psycholinguistic theories of utterance planning is the choices speakers make in spontaneous speech. Speakers will often be able to choose between multiple syntactic structures, lexical items to fill those frames, and pronunciations of those lexical items. The choices speakers make at each of these levels can reveal the strategies they're using and the priorities of the processing system.

This study investigates morpho-syntactic reduction, or contraction, in American English. This includes examples like

• I [am/'m] happy today
• You[have/'ve] been happy lately
• He was[not/n't] happy yesterday

Methods

We collect all instances of a reducible element from the Paraphrase corpus. From the surrounding context, we measure a number of linguistic and psychological factors known to influence language planning processes. We use logistic regression models with bootstrapping for speaker effects, and logit mixed models to test the influence of these factors on contraction use.

Studies

We have complete three studies to date, each representing a different reducible element:

• HAVE	have/'ve	had/'d	has/'s
  BE	am/'m	are/'re	is/'s
  NOT	not/n't

Current findings

We consistently find that speakers' use of contractions is predicted by the UniformInformationDensity theory (UID). Speakers use contractions when the reducible element bears relatively little information, and full forms when it is more surprising.

We have also tested the predictions of another theory of utterance planning, AvailabilityBasedProduction (ABP). This theory predicts that the availability of the upcoming element (operationalize as frequency of the word after the reducible element, and conditional probability of the next word given the reducible element) will determine whether a reduced or full form is used. Full forms should be used before difficult-to-retrieve lexical items or hard-to-construct syntactic forms. Our results regarding ABP are mixed.

Ongoing work

We're currently investigating ways of explaining the intuitions behind availability-based theories in information theoretic terms consistent with UID.

We're planning to run these studies in other corpora of spontaneous speech, including the BNC and the Fisher corpus.

We intend to add a few other similar (though, in some cases, less conventionalized) instances of reduction:

• modal reduction: would/'d
• gotta, wanna, gonna, sorta, kinda