1. 1Department of Experimental Psychology
University of Ghent, Belgium
2University of Nijmegen, The Netherlands
3University of Edinburgh, United Kingdom
Grammatical encoding: a resource-limited process? Robert J. Hartsuiker1, Pashiera N. Barkhuysen2, & Matt Watson3
Introduction
Speaking is clearly constrained by cognitive resources - consider for example how difficult it is to simultaneously hold a conversation and maneuver a car through rush hour traffic. But what is the nature of these resource-limitations? According to Levelt (1989) the Conceptualizer is limited by Working Memory, but the Formulator and Articulator ‘are claimed to be largely automatic’ (p. 21). Jou and Harris (1992) supported this claim by showing dual-task effects on conceptualization (e.g., number of pauses) but not on grammatical encoding (e.g., producing number concord). But Just and Carpenter (1992) proposed a different account: There is a single ‘verbal working memory’ (VWM) capacity that constrains each subcomponent of sentence processing. Individuals vary in the size of this capacity, and capacity is taxed in conditions with extrinsic verbal loads. Although Just and Carpenter’s research addressed language comprehension, they claimed that production is similarly constrained by VWM-resources.
This study tests whether producing number agreement between subject and verb is constrained by VWM resources. We exploited the phenomenon that sentence fragments with a distributive (multiple token, MT) reading (1) elicit more number agreement errors than items with a nondistributive (Single Token, ST) reading (2) in a task where the fragment is repeated and completed (relative to number-matched controls (1a, 2a).
(1) Multiple token (MT): The date on the coins
(1a) MT Control: The date on the coin
(2) Single Token (ST): The cage with the bears
(2a) ST Control: The cage with the bear
Distributivity effects were found in several languages. Vigliocco and Hartsuiker (2002) interpreted this as evidence for a ‘’maximal input” hypothesis: the formulator uses all the information it can, not just the strictly necessary information it needs.
Logic of the experiments
If grammatical encoding is limited by VWM resources, speakers should switch from maximal input to minimal input when resources are scarce, assuming that speakers pay a price for maximal input: having to maintain both the strictly necessary information and the redundant / conflictual information.
Experiment 2: Dutch
64 students at Nijmegen University repeated and completed auditory presented sentence fragments (8/condition) using an adjective. 32 participants performed a dual-task: recalling a three-word memory load presented before each adjective. VWM-span was assessed using Daneman & Green’s (1986) speaking span test. Fig. 3 and Fig. 4 show the proportion of agreement errors by load and span (lowest vs highest quartile).
Thus, the hypothesis that production of agreement is constrained by VWM predicts that distributivity interacts with VWM-span and with verbal memory load: there should be a larger effect of distributivity if more resources are available. In line with that prediction, Hartsuiker et al. (1999) found that agrammatic patients (hypothesized to suffer from pathological VWM-limitations) showed no distributivity effect, whereas healthy controls did.
Experiment 1: English
48 students at Edinburgh University repeated and completed 10 visually presented Multiple Token (1) and 10 Single Token (2) sentence fragments and matched control fragments (1a-2a). Each preamble was preceded by an adjective to be used in the completion. There were two sessions, separated by at least 3 weeks. In one session, a dual-task was performed (recognition of the adjectives). After each session, a memory span task was administered: a series of month names was presented, and participants reported them back in the order of the year (Almor et al., 2001). The percentage of agreement errors for MT and ST items in dual-task and single-task conditions is shown in Figure 1. (There were no errors in the control conditions (1a-2a)).
Fig. 3 Agreement errors by load (%)
Fig. 4 Agreement errors by span (%)
There were significant effects of distributivity, dual-task and speaking span on percentage agreement errors. Distributivity did not interact with load. The (significant) interaction between distributivity and span was in the opposite from predicted direction - there was a smaller distributivity effect for high-span speakers.
There was a significant distributivity effect on percentage agreement errors (cf., Eberhard, 1999). The effect of dual task load was not significant, there was no effect of span (median-split, average of two sessions), and no interaction between any factor.
Fig. 5 shows dual-task performance (percentage words forgotten). More words were forgotten for items with plural local nouns than with singular local nouns, and more words were forgotten with ST and control items than with MT and control items. These factors did not interact.
Fig. 1 Agreement errors by load (%)
Fig. 5 Words forgotten (%)
Conclusions
Manipulations of VWM failed to induce “minimalist processing” (only considering strictly necessary information): Distributivity effects remained, or even increased, in low-capacity conditions.
Nevertheless, these experiments clearly show that VWM does a play a role in the production of agreement: manipulations of load, and comparisons of span groups revealed significantly higher error rates when capacity is low (Experiment 2). Furthermore, number conflicts in the sentence fragments interfered with dual-task performance: both speed of recognition and quality of recall were negatively affected by whether the two nouns in the fragment mismatched in number.
Figure 2 shows dual-task RTs for correct responses. RTs were slower for ‘No’ responses than for ‘Yes’ responses; slower for items with plural local nouns (1-2) than singular local nouns (1a-2a); and slower for MT and MT control items than for ST and ST control items. These factors did not interact.
Fig. 2 dual-task RTs (ms)

2. Experiment 1: primary and dual task
1Department of Experimental Psychology
University of Ghent, Belgium
2University of Tilburg, Netherlands
3University of Edinburgh, United Kingdom
Primary task: repetition and completion of visually presented sentence fragment, using adjective provided, e.g.: …
Presented: fast The sailboat on the sea
Response: “The sailboat on the sea was fast”
Dual task: 1-back recognition task. Another adjective is presented after the response, and the participant decides (by pressing a key) if that was the adjective used in the sentence before this one:
Presented: fast The sailboat on the sea
Response: “The sailboat on the sea was fast”
Presented: OLD Response: <NO-key>
Presented: yellow The banana for the monkey
Response: “The banana for the monkey was yellow”
Presented: FAST Response: <YES-key>

3. Experiment 2: primary and dual task
1Department of Experimental Psychology
University of Ghent, Belgium
2University of Tilburg, Netherlands
3University of Edinburgh, United Kingdom
Primary task: repetition and completion of sentence fragment, using adjective provided, e.g.:
Presented: fast The sailboat on the sea
Response: “The sailboat on the sea was fast”
Dual task: recall of three words, presented prior to the adjective and fragment, e.g.:
Presented: attitude labor promise
Presented: fast The sailboat on the sea
Response: “The sailboat on the sea was fast”
Response: “attitude, labor, promise”

4. Experiment 1: Month sorting task
1Department of Experimental Psychology
University of Ghent, Belgium
2University of Tilburg, Netherlands
3University of Edinburgh, United Kingdom
December
Series with list lengths 2 - 8
5 lists per series
Score is number of lists sorted correctly (max 35)
M =21; SD = 3.13
Test-retest reliability: r = .59 (*)
0 sec
May
1 sec
October
2 sec
March
3 sec GO
4 sec
“March, May, October, December”

5. Experiment 2: Speaking span task
1Department of Experimental Psychology
University of Ghent, Belgium
2University of Tilburg, Netherlands
3University of Edinburgh, United Kingdom
candles
Series with list lengths 2 - 5
5 lists per series
Score is number of correct(ly ordered) sentences (max 70)
M = 38.5; SD = 5.89
0 sec
joy
1 sec
talked
2 sec
force
3 sec GO
4 sec
“The candles are lit,
Soccer gives me great joy,
We talked for hours,
May the force be with you”

6. References
1Department of Experimental Psychology
University of Ghent, Belgium
2University of Tilburg, Netherlands
3University of Edinburgh, United Kingdom
Almor, A., MacDonald, M. C., Kempler, D., Andersen, E. S., & Tyler, L. K. (2001). Comprehension of long distance number agreement in probable Alzheimer’s disease. Language and Cognitive Processes, 16,
Daneman, M., & Green, I. (1986). Individual differences in comprehending and producing words in context. Journal of Memory and Language, 25, 1-18.
Eberhard, K. M. (1999). The accessibility of conceptual number to processes of subject-verb agreement in English. Journal of Memory and Language, 41,
Hartsuiker, R. J., Kolk, H. H. J., & Huinck, W. J. (1999). Subject-verb agreement construction in agrammatic aphasia: The role of conceptual number. Brain and Language, 69,
Jou, J., & Harris, R. J. (1992). The effect of divided attention on speech production. Bulletin of the Psychonomic Society, 30,
Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99,
Levelt, W. J. M. (1989). Speaking: From Intention to Articulation. Cambridge, MA: MIT Press.
Vigliocco, G., & Hartsuiker, R. J. (2002). The interplay of meaning, sound, and syntax in sentence production. Psychological Bulletin, 128,