Speakers Reduce Because of Their Own Internal Representations
Jason KahnJennifer Arnold
UNC – Chapel Hill
You really have to watch Federer to understand the beauty of top-flight sports
performance.
Sure, but do you think tennis is as accessible to middle America as
football?
Sure, but do you think tennis is as accessible to middle America as
football?
You really have to watch tennis to understand the beauty of
top-flight sports performance.
Repeated Mentions Get Reduced (e.g. Bard et al., 2000; Fowler & Housum, 1987)
You really have to watch tennis to understand the
beauty of top-flight sports performance.
Linguistically Given Discourse Status
…tennis…
+givenness +predictability
You really have to watch Federer to understand the beauty of top-flight sports
performance.
…tennis…
Linguistically New Discourse Status
-givenness -predictability
General Questions
• What mechanism drives reduction?
• Does it involve audience design?
Audience Design
• Broadly speaking, designing utterance with audience in mind
• When it comes to acoustic reduction…– Joint Discourse Status – represented explicitly,
defined as shared information– Facilitated processing• For the speaker• For the listener
Why shorter duration on second mention?
Joint discourse status
Givón, 1983; Grosz et al., 1995
“Tennis”
Why shorter duration on second mention?
Speaker-internal Activation(The alternative - our proposal)
“Tennis”
Research Questions
• Must we explicitly represent discourse status for the purposes of reduction?– Or can we account for the same data by focusing
on the activation of other necessary representations?
• Must we explicitly represent the listener’s knowledge? – Or is audience design not the primary motivator?
Joint Discourse Status
CONCEPTUALIZATION STAGE
DISCOURSE STATUS(given vs. new)
(what speaker and listener both know)
FORMULATION STAGE
ARTICULATION STAGE
Adapted from Levelt (1989), Schmitt, Meyer & Levelt (1999), and van der Meulen, Meyer, & Levelt (2001)
Joint Discourse Status
CONCEPTUALIZATION STAGE
DISCOURSE STATUS(given vs. new)
(what speaker and listener both know)
“tennis”
Topic continuity tracks givenness information - in other words, givenness also creates predictability information
Fowler & Housum, 1987; Prince 1992
+givenness +predictability
CONCEPTUALIZATION STAGE
FORMULATION STAGE
ARTICULATION STAGE
Activation-based
Adapted from Levelt, 1989; c.f. Balota, Boland & Shields, 1989; Bard et al., 2000; Bell et al., 2009
Activation-based
CONCEPTUALIZATION STAGE
FORMULATION STAGE
“tennis”
Both predictability and givenness should create activation, and thus should be separable
givenness
predictability
Linguistic vs. Non-linguistic Givenness
“The accordion…”
Linguistic vs. Non-linguistic Givenness
“The accordion…”
Bard & Anderson, 1990; Clark & Marshall, 1981; Prince, 1992
CONCEPTUALIZATION STAGE
FORMULATION STAGE
CONCEPTUALIZATION STAGE
DISCOURSE STATUS(given vs. new)
(what speaker and listener both know)
FORMULATION STAGE
Joint Discourse Speaker-internal
Instruction-giving Task
Speaker
Approximately 12 feet
Listener
Experimental Paradigm
Speaker: “The accordion rotates right”
Speaker: “The toothbrush shrinks”
Speaker: “The belt expands”
Experiment 1: Priming Information
“The toothbrush;The belt;
The accordion”
Control Non-linguistic Linguistic
Joint Discourse Status predicts….
Activation-based predicts…
Reduced Duration of the Object Word
Linguistic
Non-linguistic
Control
360 370 380 390 400 410 420
Linguistic < Non-linguistic < Control
*
*
Activation-based Naturally Accounts For These Findings
Non-linguistic information led to reduction
Linguistic information led to more reduction
This task used predictability as a control to investigate givenness…
Predictability & Givenness
• A discourse status account predicts that givenness and topic continuity (i.e. predictability) pattern together, in the same representation
• By contrast, an activation-based model allows either predictability or givenness to lead to reduction.
Fowler & Housum, 1987; Prince 1981
Experiment 2: Target Given1/8 of trials
“The accordion”
Non-linguistic Linguistic
Experiment 2: Target New7/8 of trials
“The toothbrush”
Non-linguistic Linguistic
Reduced Duration of the Object WordTarget Given < Target New
*Linguistic
Non-linguistic
420 425 430 435 440 445 450 455 460
Target New
Target Given
Further Confirmation
• Even in the absence of strong predictability, speakers reduce in response to linguistic givenness information
• Exp.’s 1 & 2 suggest that speakers do not need to model discourse status explicitly for reduction
Speaker-internal Audience Design?
• Traditional views of discourse status say that speakers use it in part to model the listener (Clark & Marshall, 1981; Gundel et al., 1993)
• But if we do away with a representation of discourse status here, we should still ask whether speakers do it because of themselves or because of their listeners
Instruction-giving Task
Speaker
Listener
1) Blocked trials
2) Icon at the top of the screen
3) Headphones
v v
v v
Both
Speaker
Listener
None
405 410 415 420 425 430 435 440 445 450 455
Reduced Object Duration(Both, Speaker) < (Listener, None)
Speaker-internal Activation
• If speakers were tracking discourse status, they should have shown a different pattern of reduction
• Even without discourse status, speakers could have reduced for their listener, but did not
Summary of Results
• Linguistic givenness elicited more reduction than non-linguistic givenness in Experiment 1
• Linguistic givenness elicited reduction even without strong predictability in Experiment 2
• Speakers reduced when, and only when, they had givenness information in Experiment 3 (the listener doesn’t matter here)
CONCEPTUALIZATION
FORMULATION
Non-linguistic givenness created reduction
Linguistic givenness created more reduction
Givenness and predictability have separable effects, suggesting a common
substrate, namely activation.
Discourse Status Matters Elsewhere
• Word order (Arnold, Wasow, et al., 2000; Birner & Ward, 1994)
• Lexical choice (e.g. pronouns vs. more explicit expressions) (Ariel, 2000; Arnold, 1998; Gundel et al., 1993)
• Accenting (Pierrehumbert & Hirschberg, 1990)
Other Potential Models
• These results are still technically consistent with a model that includes an explicit representation of discourse status at the conceptual level.
• But we propose that our model is both more parsimonious and makes additional predictions, which we are currently testing
The Role of Audience Design
• Has effects on word choice, amount of detail, number of words (Arnold, Kahn & Pancani (CUNY Poster Thursday); Bard et al., 2000; Galati & Brennan 2010)
• The effect of audience design on reduction is mediated by the speaker’s internal representations (c.f. Balota, Boland & Shields, 1989; Bard et al., 2000)
The Role of Audience Design
Arnold, Kahn & Pancani, CUNY Poster Thursday
Take Home Message Slide
• Speakers reduce based on the state of their own internal representations– They don’t appear to need an explicit
representation of discourse status– They don’t appear to track the state of their
listener(s)
With Gratitude To…
• The Cognitive and Language groups at UNC for endless discussion, support, and critique
• Kellen Carpenter, Giulia Pancani, Alex Christodoulou, Alyssa Ventimiglia, Jennifer Tate, Sam Handel, and Leighanne McGill for help with these experiments
• And Ellen Bard, Scott Fraundorf, Florian Jaeger, Tuan Lam, Janet Pierrehumbert, and Joseph Tyler for useful discussions