journal of memory and language · the incrementality hypothesis is appealing because it accounts...

Journal of Memory and Language 46,57–84 (2002)doi:10.1006/jmla.2001.2797, available online at http://www.academicpress.com on

How Incremental Is Language Production? Evidence from the Productionof Utterances Requiring the Computation of Arithmetic Sums

Fernanda Ferreira and Benjamin Swets

Michigan State University

The incremental approach to language production assumes that the production system interleaves planning andarticulation processes. Two experiments examined this assumption. In the first, participants stated the sums of two

the se-ted aslanning the diffi- exper-

latenciesc prem- languageak incre-

A 9)—

andandganvicefor RichCarnallDavMikandcom

AFerBuil488

two-digit numbers in one of three different kinds of utterances, the sum by itself, the sum followed by quence “is the answer,” or the frame “The answer is” followed by the sum. Problem difficulty was manipulawell, so that in some conditions, speakers could (in principle) state the tens component of the sum while pthe ones. Latencies to begin to speak were the same for all three utterance types and were affected byculty of the problem as a whole. Utterance durations were unaffected by problem difficulty. In the secondiment, participants were induced to speak incrementally through the use of a deadline procedure. Both and utterance durations were influenced by the difficulty of the problem. This latter finding supports a basiise of the incremental approach: Speakers sometimes speak and plan simultaneously. Nevertheless, theproduction system appears not to be architecturally incremental; instead, the extent to which people spementally is under strategic control.© 2001 Elsevier Science

Key Words: language production; arithmetic; incrementality; syntax; phonology.

lmost since the modern era of psycholin-words—so-called “lemmas” (Levelt, 198

r6

o

asng to in-m-butm-

a-

ersch

onree

guistics began in the 1960s, researchers havegued for a hierarchical model of language pduction (Fromkin, 1971; Garrett, 1975, 1971980) in which processing proceeds from smantic intention to articulation. According to recent version of this model (Bock & Level1994) activity begins in the Message Compnent when a speaker forms an idea of whatwishes to say. The Grammatical Component erates next, and is itself subdivided into twstages of processing. The “Functional Level” slects the meaning and syntactic features

This study was supported in part by grant SBR-93192

57

ut-r’sose an

-heiousm.irs&bly

BCS-9976584 from the National Science Foundati an All University Research Initiation Grant from Mich State University. We thank Marc Brysbaert for his a and ideas that helped motivate this study, and Wim Fproviding the materials. We also thank Gary Shrocard Falk, Mel Johnson, Robin Revette, Leah Englis

rick Williams, and Janis Stacey for their assistance. y, we are grateful to Karl Bailey, Herb Clark, Douidson, Vic Ferreira, Zenzi Griffin, John Henderson, ane Tanenhaus for helpful discussions about the stud the Stanford Language Users Group (SLUGs) for thments on the work.ddress correspondence and reprint requests to Fernareira, Department of Psychology, Psychology Reseading, Michigan State University, East Lansing, M24-1117. E-mail: [email protected].

0749-596X/01 $35.00

ar-o-,e-at,o-hep-oe-of

and assigns them a grammatical role suchsubject or object. Then, positional processitakes place: The word-forms correspondinglemmas are retrieved, and the serial order ofdividual phrases is put together. The next coponent engages in phonological processing (see F. Ferreira (1993) for evidence that the coputation of the intonational and metrical fetures of an utterance precederetrieval of word-forms), and from there the system sends ordto the articulatory organs to produce speesounds appropriately.

Any adequate model of language productimust incorporate mechanisms to explain thimportant features of normal speech: First,terances generally conform to the speakecommunicative goals. Second, speakers chofrom among various ways they might expressidea. For instance, a speaker may say Ten plusseven make seventeen, or Seventeen is the sumof ten plus seven, and so on. The final choice reflects both the communicative needs of tspeaker and the processing states of the varcomponents of the language production systeFinally, although pauses, repetitions, and repacertainly do occur in normal speech (Clark Wasow, 1998), speakers are usually reasona

72on,i-d-iask,h,Fi-gd

ies,eir

ndarchI

© 2001 Elsevier ScienceAll rights reserved.

A

e

l 9

nel

nci

l

nw

t

ee

llnel

errrae

y

e

hsce

nd-tsthe be(p.-handr-

g ofor-

be-edrlye-x-omingan-o-

hebe-odsd,kee,ing do

nil-

ise-uc-hor-tic ofn- toen-

aer--

58 FERREIRA

fluent (Levelt, 1989). At the extreme, no comptent speaker needs to pause before retrieveach successive word of his utterance.

In many recent models of how people tathese features of speech are explained withsingle assumption that language productionincremental (V. Ferreira, 1996; Levelt, 198Roelofs, 1998; Wheeldon & Lahiri, 1997). Onca piece of information at a level of processibecomes available, it triggers activity at the nlevel down in the production system (Leve1989, p. 23: “Each processing component wbe triggered into activity by a minimal amouof its characteristic input”), and this sequenmay continue all the way to articulation. Th“vertical” aspect of the architecture is coordnated with what we might term its “horizontaactivities: At the same time that a piece of infomation works its way from idea to articulatioother pieces are constructed and make their through the system as well. Thus, incremenmodels assume that various levels can operaparallel. At the same time, this vertical paralelism implies a certain amount of horizontal sriality. That is, because the system is incremtal, information at a particular level ofprocessing is not necessarily handled in paraFor example, a word at the end of an utterawould not be syntactically available in parallwith one at the beginning, because the earword was shunted off for phonological procesing after it was syntactically encoded.

To illustrate how an incremental system opates, consider a speaker who wishes to expthe fact that 17 is the sum of 10 and 7. The pcessing events might take place as follows. Fiassume that the concept TEN becomes activfirst. That activation will trigger retrieval of thcorresponding lemma, which will cause thlemma to take its place at the left edge of a stactic frame. (Simultaneously, other concepbecome activated, but their presence is not rvant at the moment.) These events will causeword-form for TEN to be accessed, so the woten may be articulated. A consequence of tarchitecture, then, is that the speaker could ten before knowing exactly how the utteranwill end. As Wheeldon and Lahiri (1997) stat
“What a processor is doing with a particula
ND SWETS

-ing

k,theis;egxtt,illte

si-”r-,ay

tale inl--n-

el.cel

iers-

r-esso-st,ted

en-tsle-therdisaye,

fragment of an utterance should not be depeent on information available in later fragmenof the utterance. For example, constructing initial prosodic units of a sentence should notdependent on how the sentence will end”361). Having now said ten, the speaker is committed to some sort of a global form in whicthe addends occur earlier in the utterance the sum occurs later. Therefore, a likely utteance given these events is something like Tenand seven make seventeen.

The incrementality hypothesis is appealinbecause it accounts for all three featureshuman speech mentioned above. A concept cresponding to what a speaker intends to saycomes activated based on information containin the message-level representation. The eaactivation of the concept leads to early placment in the utterance. Incrementality, then, eplains at least in part how speakers choose framong different sentence forms for expresstheir ideas. Syntactic variations are taken advtage of by the production system to accommdate the states of activation of information in tproduction system. And speakers are fluent cause they do not need to pause for long periof time in order to plan their utterances. Insteain a highly parallel system, a concept may maits way vertically through the architecture whilsimultaneously, other concepts are becomactivated and encoded. Therefore, speakersnot need to plan whether to say ten and sevenmake seventeenor seventeen is the sum of teand seven. The decision is made by the avaability of concepts in long-term memory.

Furthermore, the incrementality hypotheshas received a fair bit of empirical support in rcent experimental studies of language prodtion. The consequence that has been most toughly examined is the notion that syntacforms are chosen to reflect states of activationthe production system. Bock (1986) demostrated that a semantically primed word tendsoccur early in a sentence. Thus, if an experimtal participant is shown the word worship andthen sees a picture of lightning striking church, she is likely to produce a passive uttance such as the church is being struck by light
rning. The idea is that the semantic prime caused

N

staouhc

the

hee

rti tr

mee-tsr

nt cIL

te

um. nt

skn., tlde tesneregrsrer

re-ifys-e,

en--l

ra-or toic-on

ganre

f ahe ach is(ofere ei-toce.ero-lyer

then

nlya-’scalge

.n-by8),ith

mll

ag it

INCREMENTALITY IN LA

the concept CHURCH to become activated. Aresult (i.e., because the system is incrementhe word churchgrabbed the subject position the sentence, which then committed the prodtion system to an utterance in which the otnominal concept occurred later in the sentenas in a passive, for example. Furthermore, ifsystem operates in this fashion, one would pect that fluent production would be easier wthere is syntactic choice rather than when this only one grammatical way to express a paular idea. V. Ferreira (1996) found evidencesupport this implication of incrementality. Paticipants in his experiments saw the initial fraI GAVE on a computer monitor, and then eiththe word TOYS followed by CHILDREN or thword CHILDREN followed by TOYS. In another condition, instead of I GAVE, participansaw I DONATED. The idea is that if a verb pemits either order of the postverbal argumeproduction will be faster because the systemhandle either sequencing of TOY and CHDREN. In contrast, with a rigid verb such as do-nate, it is more difficult to produce a proper uterance when the concept CHILDREN becomavailable sooner than TOYS. The results sported these predictions. Because the systeincremental, it likes to have syntactic choices

Wheeldon and Lahiri (1997) conductedmore direct test of incrementality. Participareceived a noun phrase such as het water(“thewater” in Dutch) followed by the question Watzoek je(“What do you seek?”), and their tawas to answer the question in a full senteusing the noun phrase they were given (i.eikzoek het water). Participants were encouragedbegin to speak as quickly as possible. Wheeand Lahiri recorded utterance initiation timand found that latencies were shorter whenfirst phonological word of the utterance was lcomplex. The characteristics of the other phological words had no effect. They concludthat the production system is radically incmental: What determines when a speaker beto talk is how long it takes her to get the fiphonological word ready. This conclusion is inforced by results from their other three expements, which demonstrated that latencies are
fected by the complexity of the entire utteranc
GUAGE PRODUCTION 59

al),

fc-ere,ex-nrec-o-er

-s,an-

-s

p- is

as

ce

oonshes

o-d-inst-i-af-

when participants were asked to prepare cafully before beginning their utterances. Thus,speakers are allowed to speak naturally, the stem behaves incrementally. As they argu“when it is possible to do so, speakers prefertially initiate articulation following the phonological encoding of the initial phonologicaword of an utterance” (p. 377).

On the other hand, some findings in the liteture are inconsistent with incrementality. Fexample, Lindsley (1975) asked participantsrespond as quickly as possible to a simple pture showing a transitive action (e.g., one perstouching another). He found that speakers beto phonologically encode their utterances befothey had syntactically encoded the object otransitive action but not before they knew tverb. In another study, Meyer (1996) usedword distractor paradigm to examine how muinformation about the words of an utteranceaccessed prior to articulation. Speakers Dutch) produced simple utterances and wpresented with a spoken distractor that wasther semantically or phonologically related either the subject or object noun of the sentenMeyer found that a semantic distractor for eithnoun increased initiation times, while a phonlogical distractor impaired performance onwhen it was related to the subject noun. Meyconcluded that sentence production requiresretrieval of the semantic/syntactic informatioassociated with most of the utterance, but it orequires the retrieval of phonological informtion for the first word or phrase. Thus, Meyerexperiments provide evidence that grammatiencoding requires simultaneous knowledabout both preverbal and postverbal material, afinding that is inconsistent with incrementality

Further evidence that is inconsistent with icremental production comes from a study Stallings, MacDonald, and O’Seaghdha (199who examined the tendency for sentences wmain verbs such as transferredand introducedto occur in either a canonical or a shifted for(e.g.,Mary introduced the new neighbor to Bivs Mary introduced to Bill the new neighbor, re-spectively). They found that the more oftenverb tended to occur in structures separatin
efrom its complement, the more likely partici-

A

oe

e irv

o

tgnoe

u

ypm

to

i

afla

dmu

pn

doen-

and,m

onsgch.es art,n.mele,

8)o-

m-

iththeedak- forgescedn”,

ndteeo-

forultchs

- ltsss

cann-. Inlisew(p.

o-

60 FERREIRA

pants were to produce shifted structures. Mimportantly, latencies to begin to speak wlonger when participants produced sententhat were incompatible with the main verbshifting history, indicating that the structurcompeted to be the ultimate syntactic framethe sentence. This conclusion in turn implthat language production is not always incmental: Not only were structures planned ofairly large domains, but even more than ostructure became activated simultaneously.

Finally, just how fluent is real language prduction? Studies of naturally occurring speehave shown that speakers do tend to pausefore major syntactic constituents. For exampHenderson, Goldman-Eisler, and Skarb(1966) found that periods of fluency alternawith phases during which the speaker engain a great deal of pausing. Also, more demaing speaking tasks (e.g., interpreting a cartversus merely describing it) result in less fluspeech (Goldman-Eisler, 1968). Similarly, Fo(1982) examined spontaneous speech for palonger than 200 ms and found that they pceded about 20% of deep clauses. Holm(1988) asked speakers to talk spontaneouslvarious topics and then asked another grouparticipants to read the utterances the forgroup produced. She found that pauses anditations occurred before complement and retive clauses in spontaneous but not read speThese pauses suggest that speakers disrupency in order to plan their speech (see alsFerreira, 1991, 1993), and the fact that they pin units that are approximately clausal in s(see also Bock & Cutting, 1992) indicates ththe system has nonincremental aspects. FinClark and Wasow (1998) have shown that disencies in normal speech are quite common furthermore, that they are more likely to occthe more complex the upcoming constitueThus, although an “ideal delivery” (Clark &Clark, 1977) might be the speaker’s goal, unmost normally demanding speaking circustances, pauses, hesitations, and other disflcies are fairly frequent.

Where are we so far? On the one hand, eximents in which speakers are asked to begi
speak as soon as they can tend to show that
ND SWETS

rereces’ssforese-erne

-ch be-le,ekededd-onntrdses

re-es on ofer

hes-la-ech. flu- F.

lanzeatlly,u-nd,urnt.

er-en-

er- to

ple plan more than one word at a time, as studies that examine pauses and other disflucies in spontaneous speech. On the other hstudies of syntactic choice in production seeto favor incrementality, as does the Wheeldand Lahiri finding that only the characteristicof the first phonological word of an upcominutterance influenced latencies to initiate speeEvidence for quite radical incrementality comfrom what might appear on the face of it to berather unlikely source, a study by BrysbaeNoel, and Fias (1998) on numerical cognitioThe Brysbaert et al. results motivated to soextent the experiments described in this articso their study will be described in detail.

The participants in Brysbaert et al. (199stated the sum of a one-digit number and a twdigit number as quickly as possible. For exaple, someone might see “21 1 4” and respond“25”. The problems were presented either wthe larger number preceding the smaller or in other order. In addition, Brysbaert et al. testtwo groups of participants, native French speers and native Dutch speakers. The reasonlooking at these two groups is that the languadiffer in whether the ones or tens are pronounfirst for a number such as “twenty-five”. IFrench, one says the equivalent of “twenty-fivebut in Dutch one says the equivalent of “five atwenty”. Thus, if participants are faster to initiaarticulation of the sum when the format of thproblem allows them to prepare the first phonlogical word (“five” or “twenty”) of their utter-ance more easily, then we will have support radical incrementality. This was indeed the resBrysbaert et al. reported: For no-carry items suas 21 1 4, French participants showed a 56-madvantage for the 21 1 4 order, and Dutch participants showed a 51-ms advantage for the 4121 order. Brysbaert et al. interpreted the resuas follows: “the Dutch speakers try to get acceto the unit of the response first, because they start programming the pronunciation of the aswer as soon as the value of the unit is knowncontrast, the French speakers have to capitaon the value of the ten, which they must knobefore response execution can be started”67). This, then, is incrementality of the sort pr

peo-posed by Wheeldon and Lahiri (1997): Speakers

N

fiteerivmreeadeiunot

n

T

otgi hctr

a

r

c

mg

di-u-

b- to

in- to

ula-n

en-hatwaswerateenen-

ain

ineingal.

ntsndWel-

de- to ut-ncect,ati-re- thed al-then-heonof ifedestcre-

ofn)e

ce-


prefer the addends to be ordered so that thephonological word of the sum can be compufirst, allowing the calculations required to detmine the other part of the sum to be done duarticulation. These results were obtained ethough carry and no-carry items were randointermixed in the experiment. Furthermowhen both French and Dutch participants wrequired to type their responses, the langudifference disappeared: Both groups preferrehave the two-digit number followed by the ondigit number. This result is also consistent wincrementality, because in both languages nbers appear in print with the tens before the o

Overall, research in language production dnot provide clear evidence one way or the ofor the incrementality hypothesis. What doseem apparent is that under some circumstaspeakers are able to begin to speak as soothey have formulated the smallest bit of lingutic structure (a phonological word), but thatother situations speakers are more cautiousdo not speak until a reasonably large chunkan upcoming utterance has been planned. observation reveals that the language producsystem is strategically incremental: An imptant component of speech planning is to demine whether the situation calls for “blurtinor for more careful planning. This perspectsuggests an important empirical question forsearchers in language production: Even wspeakers find that it is in their interests to artilate as quickly as possible in order to hold floor, do they still plan more than just the fiphonological word of the utterance anyway?finding that they do would suggest that the lguage production system mandates some pning regardless of the speaker’s goals and stgies.

To examine these issues, we conducted experiments that utilized an arithmetic tamuch like the one used in the Brysbaert et(1998) study. In our first experiment, partipants responded to addition problems by pducing three different types of sentence forjust the answer itself, the answer at the bening of a sentence, and the answer at the ena sentence. This experiment yielded no evide
for incrementality: Participants planned the e
GUAGE PRODUCTION 61

rstd

r-ngenly,rege to-

thm-es.es

tire utterance before speaking in all three contions (and latencies were systematically inflenced by the difficulty of the arithmetic prolem). The second experiment was designedstrongly encourage participants to adopt ancremental strategy. Speakers had to beginspeak before a variable deadline. This maniption drastically reduced utterance initiatiotimes and yielded some evidence for incremtality; yet, clear evidence also emerged tspeakers planned the sum even though it sentence-final. In the General Discussion,argue that these results support only modeincrementality. Moreover, they suggest that evwhen speakers are trying to speak increm

heresces

n asis-inand ofhis

tionr-er-”vere-enu-hest An-lan-ate-

twoskal.i-ro-s,in-d ofnce

tally, the system still engages in a certamount of utterance planning.

EXPERIMENT 1

The goal of this experiment was to examincrementality in language production by seewhether we could take the Brysbaert et (1998) effect one step further. Participaadded together two two-digit numbers astated the sum in the form of an utterance. varied the difficulty of computing the tens coumn and of computing the ones column inpendently. In addition, participants were toldstate their answers in one of three ways: Theterance could consist of just the sum, a sentein which the sum was the grammatical subjeor a sentence in which the sum was a grammcal object. Recall that what Brysbaert et al. ported was that speakers preferred to haveaddends arranged so that the leftmost addenlowed them to compute and articulate just first phonological word of the sum. An extesion of this result would be a finding that, in tsum-alone and sum-first conditions, initiatitimes are influenced only by the difficulty calculating the tens column. Moreover, eventhis quite radical incrementality is not observin the experiment, the design allows us to twhether speakers are at least somewhat inmental. If they are, then problem difficulty (the entire problem, not just the tens columshould influence initiation times but only in thconditions in which the sum occurred utteran

n-initially.

62 FERREIRA AND SWETS

e ecd

enbtio

edmoued

baa

e o:iosoboth; ao

ofasn 5ary,orx-

ex- 1.ur-therein-t re- an car-isi-b-nedme.

r ofher.u-ter

oldthe thene.d assureh aen.

ntedd by

Method

Participants. All participants tested for thexperiment were undergraduate studentsMichigan State University, and all participatin exchange for partial credit in their introdutory psychology courses. In the sum-only contion, 40 participants were tested in order to up with 32 participants whose data could used. To be included in the analyses, a parpant had to produce correct answers to mthan 75% of the problems. This criterion elimnated 5 participants; another 3 people werecluded because they did not speak louenough to trigger the voice-key. For the suframe condition (“SUM is the answer”), 3 peple were not included in the analyses becathey made too many errors, and 2 becausvoice-key problems. For the frame-sum contion (“The answer is SUM”), 4 participants wereexcluded because of high error rates, and 1cause of voice-key errors. Thus, in total, dfrom 96 participants were analyzed, 32 in eof the three between-participants conditions.

Materials. Two different stimulus lists wercreated,1 and a given participant saw only onethose lists. Each list contained 168 problemsof those did not require a carrying operat(no-carry problems), 56 did (carry problemand 56 were filler problems. The no-carry prlems occurred in four different conditions: f14 of them, computation of the tens part of sum and the ones part of the sum was easyanother 14, computation of the tens was eand of the ones was hard; for another 14, c
putation of the tens was hard and of the on
Thetton al-n.

1All the materials were generously provided by Wim Fiaof the Catholic University of Leuven, and Marc Brysbaer

atd-i-deci-re

i-x-ly-

-se ofi-

e-tach

f56n),-

reforsym-es

was easy; and for the final 14, computationboth tens and ones was difficult. A column wconsidered easy if the sum was smaller thaand hard if the sum was greater than 5 (GeBow-Thomas, & Yao, 1992). (Problems fwhich either column summed to 5 were ecluded from the no-carry and carry sets.) An ample of each problem type is given in TableThe carry problems also participated in foconditions, “easy”, “hard”, “harder”, and “hardest”. This continuum is based on the size of resulting sum (Ashcraft, 1992). There wethree types of filler problems, those that cluded an addend smaller than 20 but did noquire a carrying operation, ones that includedaddend less than 20 and that did necessitaterying, and ones in which an addend was divble by 10. The no-carry, carry, and filler prolems were put into a random order, constraiso that two problems of the same type (sacondition or filler type) did not occur in a rowThe lists were identical, except that the ordeaddends was swapped from one list to the ot

Procedure. Participants were tested individally. The session began with the experimenreading the instructions. Participants were tthat they would see an addition problem on computer monitor, and their task was to statesum as quickly as possible into the microphoThey were reminded that they should responsoon as possible, but that they should make they had the correct answer. A trial began witfixation cross located in the center of the screOne second later, the problem was presewith both addends on the same line separatea plus sign, with spaces around the plus. participant responded and then pushed a buto proceed to the next trial. He or she waslowed to correct the response initially givest,

ey-
The experimenter pushed one key on the k
TABLE 1

Examples of the Problems Used in Experiment 1

No-carry problems Carry problems Filler problemsCondition Example Condition Example Type Example

Easy 10s, easy ones 21 1 22 5 43 Easy 23 1 28 5 51 Addend , 20, no-carry 16 1 11 5 27Easy 10s, hard ones 21 1 26 5 47 Hard 23 1 38 5 61 Addend , 20, carry 17 1 36 5 53Hard 10s, easy ones 21 1 62 5 83 Harder 23 1 58 5 81 Addend divisible by 10 12 1 70 5 82Hard 10s, hard ones 21 1 66 5 87 Hardest 23 1 68 5 91

now of Ghent University.

INCREMENTALITY IN LANGUAGE PRODUCTION 63

a

orz

e

p

em

c

was

n-ncehendste,

ntsests

reinm tori-alltis-ok

ens),

vs

ct

FIG. 1. Initiation times (in ms) for no-carry problems, Experiment 1.

board if the answer given was correct, and other if it was incorrect. The sessions were tarecorded to allow those experimenter decisito be checked off-line, and to allow the dutional properties of the utterances to be analyat a later point.

For the sum-only condition, participants weasked to state the sum by itself. For the suframe condition, they were asked to give thanswers “in the format ‘X is the answer’, wheX is the sum.” They were then given an exam(“if given 25 1 25, you would say ‘Fifty is theanswer’”). For the frame-sum condition, thwere asked to state their answers “in the for‘The answer is X’, where X is the sum.” Thewere then given an example (“if given 25 1 25,you would say ‘The answer is fifty’’’).

An experimental session lasted about 45 mand participants were free to take a break tween trials whenever they wished. The sessbegan with a practice session consisting ofproblems (with the same characteristics asfillers) and ended with a debriefing in whithey were told the purposes of the study.

Results and Discussion

Latencies and accuracy data, no-carry prolems. A trial was excluded if the response tim
was less than 300 ms or greater than 20 s (2%
n-pe-nsa-ed

rem-ir

rele

yat

y

in,be-ion30theh

b-e

the data were excluded) and if the response incorrect. Latencies were analyzed as a 3 3 2 32 mixed factorial. The first variable is betweeparticipants and represents the three utteratypes (sum-only, sum-frame, frame-sum). Tother two variables are within-participants aconcern the difficulty of the problem: The tencolumn was either easy or harder to calculaand similarly for the ones. (The two differelists produced the same results, so all analycollapse over this variable.) All reported effecare significant at p , .05.

The latencies for the no-carry items ashown in Fig. 1. People took longer to begsaying the answer when it was difficult for theto calculate the tens and when it was difficultcalculate the ones. The effect of the two vaables was additive, and it was identical for three between-participant conditions. The statical analyses are as follows: Participants tolonger to initiate production of the sums whthe tens were more difficult (2397 vs 2194 mF(1,93) 5 24.22, MSe 5 164606, and longerwhen the ones were more difficult (2433 2159 ms),F(1,93) 5 101.86, MSe 5 70999.The two variables did not interact,F(1,93) 52.44, MSe 5 52436. There was no main effeof utterance type,F , 1, and no significant in-
ofteractions between this variable and the others.

A

n

e

hset

st

yiins

da

c

ea

el

ca

c

a

f-yactser;as

re-mentsr-ile

aveted

and

sthethan

n in-s a 3e- ut- isltyelyder,

re by-

on-um,

sy,acy

y,e

64 FERREIRA

The average latency was 2299 ms in the suonly condition, 2307 ms in the sum-frame codition, and 2281 in the frame-sum condition,F, 1.

The latency data place constraints on the tent to which the language production systemincremental, because the difficulty of both ttens andthe ones contributed to initiation timeIn other words, participants began their uttances only once they knew both parts of sum, not just the part corresponding to the fiproduced phonological word. (Recently, Brybaert and Fias have conducted a study idento this one but with Dutch-speaking participanfrom the Catholic University of Leuven. Thehave reported the same results: both the dculty of the tens and the difficulty of the ones fluence time to initiate production of the sumPerhaps more surprisingly, the type of utteranin which the participant had to state the sum not matter at all. Indeed, not only was the ptern the same across conditions, the absoluteaction times were remarkably similar as well.does not appear, then, that participants produtheir utterances in an incremental fashion.they had, then latencies in the frame-sum contion would have been the same for the differproblem types, because latencies would hbeen controlled by the time to produce the saphonological word across all conditions (“Thanswer” or “the answer is”). In addition, the rquirement to produce not just the sum but ato remember to place it in a particular sort frame does not seem to have affected lateneither (and we will see next that it does not fect accuracy).

The accuracy data are given in Table 2. Acracy was unaffected by the between-participamanipulation: People were 95, 94, and 94%

Hard tens 96 94

ND SWETS

m--

x- ise.r-herst-

icalts

ffi--.)ceidt-

re-ItedIfdi-ntve

mee-soofiesf-

u-ntsc-

sum conditions,F , 1. There was no main efect of the difficulty of the tens or the difficultof the ones, but the two variables did intersignificantly, F(1,93) 5 7.04. When the tenwere easy, difficulty of the ones did not mattwhen the tens were difficult, accuracy wworse when the ones were hard as well.

The accuracy results make clear that thequirement to produce the sum in a carrier fradid not impose the sort of burden on participathat would result in lower accuracy. Furthemore, the need to hold the sum in memory whsaying “the answer is” does not seem to hcaused any particular burden either, as indicaby the equivalent accuracy in the sum-frame frame-sum conditions.

Latencies and accuracy data, carry problem.Three percent of trials were eliminated from analyses because of response times less 300 ms or greater than 20 s, or because of acorrect response. Latencies were analyzed a3 4 mixed factorial: The first variable is btween-participants and represents the threeterance-type conditions. The other variablewithin-participants and concerns the difficuof the problem as reflected in the progressivincreasing size of the sum (easy, hard, harhardest).

The latencies for the carry problems ashown in Fig. 2. Latencies were unaffectedutterance type,F , 1. In general, and equivalently across all three between-participant cditions, latencies were longer the larger the sF(3,93) 5 45.97, MSe 5 231628: latencieswere 3145, 3580, 3839, and 3862 in the eahard, harder, and hardest conditions. Accurtoo was unaffected by utterance type (F , 1),but it did vary according to problem difficultF(3,93) 5 7.78, MSe 5 0.0099. People wer

ns,

ones

curate in the sum-only, sum-frame, and frame-87% accurate in the easy and hard conditio

TABLE 2

Percentage Correct for No-Carry Problems, Experiment 1

Utterance typeSum-only Sum-frame Frame-sum

Easy ones Hard ones Easy ones Hard ones Easy ones Hard

Easy tens 94 94 94 94 93 94
97 92 96 93

5

otu

oeu.l a

-

ohte

u3t d

t

edt-L’sto

ndrstd

in-n.arei-atue

en-y-uredose.of

reif-

he

FIG. 2. Initiation times (in ms) for carry problems, Experiment 1.

and 81 and 83% accurate in the harder and hest conditions.

The data from latencies and accuracy for bthe no-carry and carry problems provide litsupport for the notion that the language prodtion system is highly incremental. To explothe incremental hypothesis further, the duratiof the utterances were measured and analyzorder to assess whether participants compthe sum online during utterance production (ias they were saying “the answer is”) as welprior to its initiation. Disfluencies were anlyzed as well.

Utterance durations. Data from the framesum condition were analyzed in order to asswhether speakers might have engaged in splanning, or checking of their answers, as tproduced the frame component of their utances. Unfortunately, not all the recordinmade of the experimental sessions were of form quality, and therefore only 26 of the participants whose data were included in analyses for the frame-sum condition couldincluded in these analyses of duration and fluencies. (Analyses of the latencies and acracy levels of just these 26 participants de
mined that this subgroup showed the sapattern as the larger one.) The utterances w
ard-

thlec-

rensd intede.,as-

essmeeyr-

gsni-2hebeis-cu-er-

digitized at a rate of 20 kHz using ComputerizSpeech Laboratory (CSL, from Kay Elemerics), and durations were measured using CSwaveform editor. The utterance was divided inparts as follows, the sequence the answer is, thesum, the first digit of the sum, and the secodigit of the sum. Any pause time before the fidigit was included as part of its duration, anany pause time before the second digit wascluded as part of the second digit’s duratioOne concern is that the digits themselves different in the different experimental condtions (see Appendix A), and it is possible thsome of the differences in sum durations are dto the digits themselves and not the experimtal manipulations. We will present the data anway because we wish to explore any measthat might potentially reveal incrementality, anbecause the results are not discrepant from thfor the answer is. Results are shown in Table 3

For the no-carry utterances, the duration the answer iswas longer when the tens weeasyto compute than when they were more dficult (475 vs 445 ms),F(1,25) 5 16.84, MSe 51431. Also, the duration was longer when tones were difficult (454 vs 467 ms),F(1,25) 5


meere3.88, MSe 5 1489,p , .10. These two variablesdid not interact,F , 1. Clearly, the pattern here

awa

3

h

em

e-c

e

.e

r

Oftheon-edterre%)r-ses;

ce-

ourhe thekly.lt,u-

en-al-tingen,

ts to

we re-nd tok-

theenak-o-

Sum hardest 485 758 422 336

is not straightforward. It appears that speaktook longer to get through the frame when thnoted that the first part of the sum would be eto compute, and at the same time, there some tendency for the speakers also to tlonger when the ones were difficult.

The duration of the sum was longer when cculation of the tens was difficult vs easy (687650 ms),F(1,25) 5 8.33, MSe 5 2858. The du-ration of the first digit of the sum (the tens) wlonger when the tens were difficult (347 vs 3ms),F(1,25) 5 2.87, MSe 5 1222,p , .10, andsimilarly for the duration of the second digit—the second digit was longer when the tens been difficult (340 vs 315 ms),F(1,25) 5 6.84,MSe 5 2273. It appears that when the tens wmore difficult speakers stretched out the tithey spent saying the sum.

For the carry utterances, the durations of theanswer is, the first digit, and the second digwere unaffected by problem difficulty (see Tab3), all Fs , 1.

Analyses of disfluencies for frame-sum uttances, both problem types. A laboratory assistant unfamiliar with issues in psycholinguistiand cognitive science listened to the utteranand noted any disfluencies. Disfluencies wcategorized into three major types, based ontype of editing term the speaker used (ClarkWasow, 1998): “uh”, “um”, and other (e.gwords such as “geez” and profanities). Thwere five potential locations for a disfluency, bfore the utterance, after “the”, after “answeafter “is”, and after the first digit of the sum. (N
disfluency term occurred inside a word.) Ove
erseysyaske

al-vs

as5

ad

ree

itle

r-

scesrethe &,ree-”,o

all, editing terms of any kind were quite rare. the 3854 utterances that were digitized for duration and disfluency analyses, only 69 ctained a disfluency (and only 2 of those involvthe term “um”), and half of those occurred afthe word “is”. A total of 3530 utterances weassociated with correct responses; 63 (1.78included some disfluency. A total of 324 utteances were associated with incorrect respononly 6 (1.85%) included a disfluency.

In sum, this experiment yielded little evidenfor incrementality. A skeptic could argue, however, that the results are due to the nature oftask. The experimental situation calls for tspeaker to produce error-free utterances, andspeaker has no need to begin to speak quicThe problems themselves were quite difficuparticularly the carry problems. Therefore, calclation might have required all the speakers’ atttion and not permitted the participants to “dutask” such that they were speaking and calculaat the same time. The second experiment, thwas changed so as to encourage participanspeak incrementally as much as possible.

Before turning to the second experiment,will address one final issue concerning thesesults. Recall that our goal was to try to exteBrysbaert et al.’s (1998) result that seemedsupport radical incrementality. If French-speaing participants (for instance) preferred the xx1x order because it allowed them to articulate tens while calculating the ones column, thperhaps with two two-digit addends our speers of English would plan only the tens comp


TABLE 3

Durations of Utterances in the Frame-Sum Condition (in ms), Experiment 1

Duration of Duration of Duration of Duration of the answer is sum digit 1 digit 2

No-carry problemsEasy 10s, easy ones 466 654 333 321Easy 10s, hard ones 485 647 337 310Hard 10s, easy ones 442 685 346 339Hard 10s, hard ones 448 688 348 340

Carry problemsSum easy 504 713 390 323Sum hard 483 754 420 334Sum harder 501 751 413 338

r-nent of the sum before beginning to speak. One

N

oa

ad

di la

o

m

,anir

in-ro-othtaln

thedi-

as-usein-


possible challenge to this logic that can almcertainly be dismissed centers around the lguage difference across the two studies. Tlead is almost certainly unpromising, becauseall relevant respects English and French alike. A more likely possibility focuses arounthe difference associated with adding two twdigit addends versus adding a one-digit antwo-digit addend. The former calculation more difficult, and perhaps is too demandingallow the dual-tasking (articulation and calcution) that incrementality in this task demandThe next experiment provides us with an opptunity to evaluate this suggestion, because included the simpler mixed addend proble
that Brysbaert et al. used and varied the order
rwrba

si, to

e of ofri-iton-d-

ben.asbe-di-

der. ofher.b-

lems had the one-digit problem to the left of the

which the addends occurred.

EXPERIMENT 2

Because of the surprisingly strong nonincmental effects obtained in Experiment 1,made changes to the experimental task in oto maximize the chances that incremental havior could emerge. First, problems were measier. Second, by introducing a timing barthe task, we gave participants an incentivebegin to speak quickly rather than to wait unthey felt entirely ready. Finally, the stimuluproblems were left on the screen when partpants began to speak, unlike in Experiment 1which participants’ voices caused the stimulidisappear. All utterances were produced in form “The answer is SUM” (see the section
pilot work below).
75 1 4 5 79

GUAGE PRODUCTION 67

stn-

his inre

o- asto-

s.r-

wes in

e-e

dere-detototil

ci-in

tohen

Method

Participants. Fifty participants were testedall of whom were undergraduates at MichigState University receiving partial credit in theintroductory psychology courses. To be cluded in the analysis, a participant had to pduce correct answers to more than 75% of bthe practice problems and the experimenproblems. All participants met these criteria. Aadditional 4 participants were tested prior to 50 whose data we are reporting. These 4 inviduals were tested in pilot work designed to sess what would happen if speakers could any utterance form they wished (see “Pilot structions,” below, for more details).

Materials. Two different stimulus lists wercreated, and a given participant saw only onethem. Each list consisted of 112 problems, 56the no-carry experimental items from Expement 1 (all of which consisted of two two-digaddends), and 56 new no-carry problems csisting of a one-digit addend and a two-digit adend. Computation of the ones column couldeasy or hard, and similarly for the tens columA column was considered easy if the sum wbetween 1 and 5, and hard if the sum was tween 6 and 9. Examples illustrating the contions are shown in Table 4.

Problems were presented in a random orThe lists were identical, except that the orderaddends was swapped from one list to the otIn any given list, half of the mixed addend pro

two-digit problem, and half had the addends in

TABLE 4

Examples of the Problems Used in Experiment 2

Mixed addends problems Same addends problemsCondition Example Condition Example

Easy tens, easy ones 2 1 41 5 43 Easy tens, easy ones 21 1 22 5 4341 1 2 5 43 22 1 21 5 43

Easy tens, hard ones 5 1 24 5 29 Easy tens, hard ones 21 1 26 5 4724 1 5 5 29 26 1 21 5 47

Hard tens, easy ones 3 1 71 5 74 Hard tens, easy ones 21 1 62 5 8371 1 3 5 74 62 1 21 5 83

Hard tens, hard ones 4 1 75 5 79 Hard tens, hard ones 21 1 66 5 87
66 1 21 5 87

A

me r

-crinkenet ohh

igoeu ie

ityigr tt

eeueea tlsmoeoseus

uner os

-

ob-y a

thethend.hatd,the

iedere to thatved

arasse it it.theng tri-rge

nts’wasiredthen.

tots

ingmesenre-

acycus-

b-torb-eganor theum

68 FERREIRA

the other order; for the same addend problehalf were arranged so that the smaller addwas to the left of the larger addend, and themaining half of the problems were arrangedthe other order.

Pilot instructions. The goal of this experiment was to allow participants to speak as inmentally as possible. As we were developthis experiment, we reasoned that speamight even choose to use different utteratypes to facilitate this goal. For example, if thknew the sum quickly and wanted to get it outhe way, they might choose to say “SUM is tanswer”; if they needed extra time, they migprefer the other arrangement, and they meven want to include extra “filler” words tstretch out the amount of planning time thwould have before having to articulate the s(e.g., “I think that the answer to that oneSUM”). If this strategy were observed, wwould have more evidence for incrementalas Levelt argued (1989, p. 245), speakers mchoose the form and even the content of theiterance to accommodate the way that an uance is evolving as it is encoded left to right.

Thus, the first four participants we testwere told to use any utterance form they likexcept that we required them to place the sinto a complete sentence. No examples wprovided, to prevent our biasing their responsAfter four participants, it became clear thspeakers did not wish to take advantage oflatitude we had given them; on all 448 tria(112 trials 3 4 participants), the speakers chothe form “The answer is SUM” (the frame-sucondition of Experiment 1). The instructions fthe actual experiment, then, which excludthese four participants (because this latter grreceived different instructions from the subquent participants), required participants to the frame-sum format for their responses.

Procedure. Participants were tested individally. The session began with the participareading a set of written instructions. They wtold that they would see arithmetic problemsthe computer monitor, and that their task wa
give the answer in the format “The answer SUM”. They were then given an example, e.g
ND SWETS

s,nde-in

e-grs

ceyf

etht

yms

;ht

ut-er-

dd,mres.these

rdup-e

-tsento

“if you saw 25 1 25, you would say ‘The answer is fifty’”.

Participants were also warned that the prlem on the screen would be accompanied btiming bar, and that they should answer question before the timing bar counted all way down and produced a loud “beep” souThe timing bar was a horizontal rectangle tgradually shrunk in size until it disappearewhereupon a beep sound was emitted by computer. The duration of the timing bar varrandomly between 2 and 4 s. Participants wtold that they should begin to speak quicklyavoid being beeped. They were also assuredthey could correct their response if they belieit to be incorrect.

The 2- to 4-s value range for the timing bthat we ultimately used for this experiment wthe first range that we attempted, and becauworked so well we saw no need to changeOur goal at the start was to pick a value for lower bound of the deadline that was loenough that it could be beaten on almost allals, because we did not want to eliminate a laproportion of our data based on the participahaving been beeped. The lower bound of 2 s about the average time that participants requin the first experiment to begin to respond to no-carry problems in the frame-sum conditioThe 4-s upper bound was chosen simplyallow there to be trials on which participanwould easily beat the deadline. After analyzthe data from the pilot participants it becaclear that the deadline values we had choworked optimally: Response latencies were duced by more than two-thirds, and accurwas not compromised (see Results and Dission).

Finally, participants were told that the prolem would remain on the computer monithroughout the trial. In Experiment 1, the prolems disappeared as soon as participants bto speak, which may have made it difficult fpeople to time-share the tasks of articulatingframe of the utterance while computing the sof the problem from memory.
is.,
A trial began with a fixation cross located inthe center of the screen. Participants hit a button

aece

oo

t

eintthwni

n

nop

o m

heatersesor-

er-na-

ason-ol- the

daly-areu-anageinme forx-ce

thectsre-ly,cu-Asac- in

la-as


to bring up the addition problem. They were lowed to correct their responses if they wishbut only the first response was considered in egorizing a trial as correct or incorrect. The perimenter then typed in their response and“enter” to proceed to the next trial. The sessiwere tape-recorded to allow utterance duratito be analyzed at a later point.

An experimental session lasted about 35 mThe session began with a practice session coning of 32 problems (with the same characterisand conditions as the experimental stimuli) aended with a debriefing in which participanwere told the purposes of the study. It is also rvant to evaluating the effectiveness of the timbar to note that at the beginning of the practrials, participants sometimes got beeped—is, they allowed the timing bar to elapse. Hoever, by the end of the practice trials, participawere never beeped, nor did they allow the timbar to elapse during the experimental triaClearly, for whatever reason, participants fouthe beep aversive and performed on every expmental trial so as to avoid hearing it.

Results and Discussion

We will first describe the latency and accuradata. We will begin with the problems consistiof two two-digit addends (same addend prlems), because these data are the most comble to the data collected in Experiment 1. Thwe will proceed to the problems consistingone- and two-digit addends (mixed addend prlems). Next, we describe utterance durationsboth types of problems. The final section su

Easy tens 97 96 95Hard tens 99 97 97

NGUAGE PRODUCTION 69

l-d,at-x-hitnsns

in.sist-icsndtsle-g

iceat-ts

ngls.d

eri-

cygb-ara-enofb-

for-

Latencies and Accuracy Data

For latencies, a given trial was excluded if tresponse time was less than 300 ms or grethan 5 s. Trials associated with incorrect responwere also excluded (trials were considered increct if the first response given as part of the uttance was the wrong sum). Latencies were alyzed as a 2 3 2 3 2 factorial. All three variableswere within-participants. The first variable worder of addends. The other two variables ccerned the difficulty of the problem: The tens cumn was either easy or harder to calculate, andsame was true for the ones column.

Same addend problems. No differences baseon order of addends were observed, so all anses collapse over this variable. Latencies shown in Fig. 3. First, latencies to begin articlation were much faster in Experiment 2 ththey were in Experiment 1. Whereas the averlatency for frame-sum, no-carry problems Experiment 1 was 2281 ms (the exact saproblems analyzed here), the average latencythe two two-digit addend problems in this eperiment was 634 ms. This marked differensuggests that the changes implemented intask were extremely effective in getting subjeto respond more quickly—latencies were duced by over 70%. Even more surprisingthis increase in speed did not compromise acracy: People were correct on 96% of trials. comparison of Table 2 and Table 5 reveals,curacy was just as high in this experiment asthe comparable conditions of the first.

Contrary to the incrementality hypothesis,tencies were longer when the ones column w

21

rd 1s

marizes the data on disfluencies. All effects aresignificant at p , .05 unless otherwise stated.

more difficult to calculate,F(1,49) 5 11.84,MSe 5 6052. For easy ones, latencies were 6

TABLE 5

Percentage Correct, All Problems, Experiment 2

Problem typeMixed addends problems

Same addends problems Single-digit Double-digit Smaller addend first Larger addend first addend first addend first

Easy 1s Hard 1s Easy 1s Hard 1s Easy 1s Hard 1s Easy 1s Ha

97 99 96 98 9993 99 96 99 98

h to

e

-ththe ialtenn

nr.ec

eirst,

tothe

be-es,

n-

rdiffi-. Ififfi-erehenesesr-ard-elyvi-ere

eo-ic-re-

FIG. 3. Initiation times (in ms) for all problems, Experiment 2.

ms; for hard ones, 647 ms. The difficulty of ttens had no effect: latencies were 637 ms ineasy tens condition and 631 in the hard tens cdition, F , 1. There was no interaction betwedifficulty of the tens and the ones,F(1,49) 51.13, MSe 5 2387,p . .25.

The pattern of results here is somewhat dferent from the pattern observed in Experime1: Only the difficulty of calculating the ones influenced latencies, rather than that of both tens and the ones. This difference indicates participants planned to a lesser extent in Expment 2 than they did in Experiment 1, whichconsistent with their being more incrementOn the other hand, the effect of ones difficumakes clear that, to some extent, speakplanned their utterances all the way to the ebefore they began to speak. This would certaiseem to represent evidence against the argumthat language production is radically incremetal, even in a situation that encourages incmentality and discourages long-term planning

The accuracy data are given in Table 5. Ovall, accuracy was very high, and performanvaried little over conditions (from 93 to 99%correct). Significant effects of the manipulatvariables were observed, nonetheless. F
there was a main effect of addend order (97.1
ehen-

n

if-nt

eatri-sl.yrsdlyent-

e-

r-e

d

for the order in which the small addend wasthe left of the larger addend, and 95.6% for opposite order),F(1,49) 5 5.00, MSe 5 0.0041.Second, there was a significant interaction tween the difficulty of the tens and the onF(1,49) 5 5.39, MSe 5 0.0046. Third, orderand difficulty of tens interacted,F(1,49) 5 5.02,MSe 5 0.0033. Finally, even the three-way iteraction was significant,F(1,49) 5 5.58, MSe5 0.0037. The one effect that is straightforwato describe is the interaction between the dculty of the tens and the difficulty of the onesthe tens were easy to calculate, then the dculty of the ones did not matter; if the tens wmore difficult, then accuracy was 98% in teasy ones condition and 95% in the harder ocondition. The rest of the accuracy differencare difficult to interpret. Fortunately, the diffeences are small, and our main concerns reging accuracy are that error rates be relativlow (which they were) and that there be no edence of speed–accuracy trade-offs (which thwas not).

Mixed addend problems. The latencies for thproblems consisting of a one-digit and a twdigit addend are shown in Fig. 3. Overall, partipants were faster to begin speaking when p


%sented with these problems than with those

N

3

fwelc

ar

o

8

eutie

othn

cunmton

mb

ne. tnged

ttey

w,

arero-

d-forginast

dealntsnal

ar-nes

),sc-nds

u-nesrd,d-u-

9%

esthertic-he-an-s.for-ec-nsen-eyer-tsn,en-

re


consisting of two two-digit addends (594 vs 6ms), F(1,49) 5 20.61, MSe 5 15425. In addi-tion, main effects were found for difficulty oboth the tens and the ones columns. People faster to begin to speak when the ones were to calculate than when they were hard to calate (583 vs 605 ms),F(1,49) 5 8.54, MSe 55270. The effect for the tens column went inunexpected direction: Participants were fastebegin their utterances when the tens were diffi-cult rather than easy (603 vs 584 ms),F(1,49) 516.88, MSe 5 2223.

There is a suggestion in the data that peresponded faster overall when given the twdigit addend before the one-digit addend (5vs 600 ms for the opposite order),F(1,49) 53.56, MSe 5 4386,p , .07. Further evidencfor this possibility comes both from the accracy data and from comments made by parpants to the experimenter during the experimtal session. First, participants were maccurate when the two-digit number was leftmost addend (98.4 vs 97.3% when the odigit addend was on the left),F(1,49) 5 5.61,MSe 5 0.0023 (although a significant interation between addend order and ones difficqualifies this main effect; see below). Secoafter the experiment was over, participants comented that when a one-digit number was first addend, its position to the left of the twdigit number made it difficult for them to coceptualize the one-digit number as belongingthe ones column. They reported that they sotimes incorrectly considered adding that numto the tens place of the second addend.

This result concerning order of the addereplicates the Brysbaert et al. (1998) effect: Pple prefer the xx 1 x order over its oppositeThey have a tendency with the other order toto add the single addend to the tens columthe second addend because they are tryindeal with the columns in the order in which thwill be spoken. Of course, the fact that this fining emerges in the latency data—at a point wbefore the speaker actually has to articulatesum—suggests that this preference is in planning and not in the articulation. In othwords, when speakers plan their utterances, the
prefer that material unfolds in the ultimate, to
GUAGE PRODUCTION 71

4

ereasyu-

n to

pleo-8

-ci-n-

reee-

-ltyd,

-he--toe-er

dso-

ry of toy-

ellheher

be-articulated order. As will be described belothe preference for the xx 1 x order is evident inthe duration data as well, and these resultsmore compatible with actual incremental pduction.

Overall, the main results for the mixed-adend problem types clearly show that even these easier problems, participants do not beto speak until they have taken account of at lesome aspects of the sum. Even with a great of incentive to speak incrementally, participashowed signs of planning all the way to the fiword of the utterance.

The accuracy data are given in Table 5. Pticipants were more accurate when the owere easier to calculate (98.7 vs 97.0%F(1,49) 5 9.04, MSe 5 0.0033, and there wano effect of tens difficulty. A significant interation was found between the order of addeand the difficulty of the ones,F(1,49) 5 5.57,MSe 5 0.0023. Participants were equally accrate for both orders of addends when the owere easy (98.7%), but if the ones were haparticipants preferred to have the two-digit adend to the left of the one-digit addend. Accracy was 98.1% for the former order and 95.for the latter.

To summarize, latency data from both typof problems revealed that, despite all of changes made to the procedure to induce paipants to speak incrementally, people nonetless persisted in showing nonincremental, plning effects from the difficulty of the problemFirst, people took longer to begin to speak the more difficult two two-digit addend problems than for the mixed addend problems. Sond, difficulty of the ones affected initiatiotimes for both types of problems. While thedata are difficult to reconcile with a radically icremental view of language production, thmay be consistent with a more moderate vsion. In order to test whether participanplanned and calculated during articulatiothereby demonstrating some incremental tdencies, we analyzed utterance durations.

Utterance Durations

Utterances from the first 24 participants we
-digitized at a rate of 20 kHz, and durations were

A

n

fin

iaio

ndn

o

ns

t

735

29tici- ofob-

’sns

nr-

ionere

idec-in-ffi-

late,as

the thewasentns,ter-

Mixed addend problems. Participants took

72 FERREIRA

measured using a waveform editor. (We alyzed only 24 of the 50 participants’ data bcause waveform measurements are extremtime-consuming to obtain, and based on the experiment we estimated that 24 participawould suffice to provide stable estimates these duration means.) Each utterance wasvided into three parts, the sequence the answeris, the first digit of the sum, and the second dof the sum. (As with Experiment 1, it is probbly prudent to put more weight on the duratdata for the frame than for the sum, becausedigits of the sum differ across conditions, atherefore intrinsic durations are not controlleIncorrect trials and trials on which correctioto previously given incorrect partial and/or fuanswers were made were excluded from data. Results are given in Table 6.

Overall, the duration of the answer iswaslonger for the more difficult two two-digit addend problems than for the mixed addend prlems (750 vs 610 ms),F(1,23) 5 58.33, MSe 532334. The same pattern held for the duratiothe first digit of the sum (479 vs 417 mF(1,23) 5 18.18, MSe 5 20534, and the duration of the second digit of the sum (435 vs 3ms),F(1,23) 5 21.93, MSe 5 7851. These dif-ferences are the first bit of evidence that peoare planning as they speak, and that calculaand articulation may go on in parallel.

Same addend problems. The duration of the

Hard tens, hard ones 654 491

ND SWETS

a-e-elyrstts

of di-

git-n

thed.)sllthe

-b-

of),-93

pleion

cult rather than easy to calculate (765 vs ms), F(1,23) 5 5.17, MSe 5 8211, and whenthe ones were difficult to calculate (771 vs 7ms),F(1,23) 5 16.20, MSe 5 5211. This resulsuggests that while uttering the frame, partpants were planning or computing both digitsthe sum. No other significant effects were served for the duration of the answer is. Thedata were similar for the duration of the sumfirst digit. Its duration was longer when the tewere more difficult (506 vs 453 ms),F(1,23) 55.99, MSe 5 22044—that is, its own duratiowas affected by how difficult it was for the paticipant to come up with its value. The duratof the tens was longer also when the ones wmore difficult (524 vs 435 ms),F(1,23) 515.57, MSe 5 24063. These two variables dnot interact. For the duration of the sum’s sond digit, only one effect was significant: the teraction between difficulty of the tens and diculty of the ones columns,F(1,23) 5 9.03, MSe5 6313. When the tens were easy to calcuthe duration of the second digit of the sum wlonger in the easy ones condition than in hard ones condition (459 vs 400 ms). Whentens were harder to calculate, the opposite true (435 vs 446 ms). Again, because differdigits are spoken in the experimental conditiothe duration data for the problems must be inpreted with some caution.

answer iswas longer when the tens were diffi-longer to say The answer iswhen the 1-digit ad-

TABLE 6

Durations of Utterances (in ms), All Problems, Experiment 2

Same addends problemsDuration of frame Duration of digit 1 Duration of digit 2

Easy tens, easy ones 703 414 459Easy tens, hard ones 767 492 400Hard tens, easy ones 755 457 435Hard tens, hard ones 775 555 445

Mixed addends problemsSingle-digit addend first Double-digit addend first

Duration Duration Duration Duration Duration Duration of frame of digit 1 of digit 2 of frame of digit 1 of digit 2

Easy tens, easy ones 581 385 421 579 341 399Easy tens, hard ones 721 529 389 670 434 360Hard tens, easy ones 647 355 420 521 346 413

388 607 456 351

A

(

ta

e

d

o

i

ns

h

b

nikci

y

s

e

ndnsut-”,u-c-

ialas

1,88s est isva- the ofe-n-fterte

ca--

re,re-he

d 2

ltsessenust

-ntx-e-e-heryeyic theAsrter.

entnse ex-

INCREMENTALITY IN L

dend was presented first rather than second vs 594 ms),F(1,23) 5 20.92, MSe 5 2263. Thisfinding is similar to the results we reported in latency data and it replicates those of Brysbet al. (1998). Participants also took longer toticulate the frame when the tens column weasyrather than hard (638 vs 582 ms),F(1,23) 538.52, MSe 5 3904, which replicates the odd efect we observed in the latency data. We afound a large effect of ones difficulty: Particpants spoke more slowly when the ones wdifficult to calculate rather than when they weasy (663 vs 557 ms),F(1,23) 5 59.68, MSe 59087. There was also a significant interactiontween addend order and difficulty of the oncolumn, F(1,23) 5 6.70, MSe 5 2106. Diffi-culty of the ones had a larger effect when thedends were presented in the more difficult orone-digit addend before two-digit addend.

Analysis of the duration of the first digit the sum revealed only one main effect: Whthe ones were difficult, participants tendedtake much longer to utter the preceding d(477 vs 356 ms),F(1,23) 5 15.31, MSe 545801. The duration of the second digit wlonger for the xx 1 x order than for the x 1 xxorder (405 vs 381 ms),F(1,23) 5 7.00, MSe 53922, and longer when calculation of the o(i.e., itself) was more difficult (413 vs 372 mF(1,23) 5 13.10, MSe 5 6259.

Clearly, the duration data overall reveal teven though speakers planned some aspeceven the very final word of their utterances fore they began to speak, they did not prepwell enough to enable them to speak confidewithout engaging in more planning during artulation. This pattern is critical, because it maclear that the production system has both inmental and nonincremental aspects. In additeven when people perform in a situation thatmost forces them to speak as incrementallypossible, they still engage in some planning.nally, the data make clear that it is possiblepeople to speak and calculate simultaneousl

Analyses of Disfluencies, Both Problem Type

Disfluencies were categorized into thrmajor types, based on the type of editing te
the speaker used (Clark & Wasow, 1998): “uh

626

heertar-as

f-lsoi-erere

be-es

ad-er,

fentogit

as

es),

atts ofe-aretly

c-esre-on,al- asFi-for.

erm

“um”, and other (e.g., words such as “geez” aprofanities). There were five potential locatioin which a disfluency might occur, before the terance, after “the”, after “answer”, after “isand after the first digit of the sum. (No disflency occurred inside a word.) If a disfluency ocurred during an incorrect trial or during a trin which a correction was made, that trial wnot counted.

First, in this experiment as in Experiment disfluencies overall were quite rare. Of the 26utterances that were analyzed (24 participant3112 trials per participant), only 21 utteranccontained any type of disfluency at all. Thus, iimportant to keep in mind that these obsertions are based on a very small subset of alldata. The first point to make is that all but twothe 21 disfluencies occurred after “is” (right bfore the articulation of the first digit of the aswer). The other two disfluencies occurred athe articulation of the first digit. Second, we nothat of the 19 disfluencies in the postframe lotion, all but two occurred with problems involving a difficult calculation of the ones. Therefothe disfluency data are compatible with the sults observed in Experiment 1 and with tmain findings of this second experiment.

Comparison of Results from Experiments 1 an

By comparing the latency and duration resufrom the two experiments, it is possible to assin what ways people speak differently whthey can plan carefully versus when they mgrab the floor quickly. Figure 4 allows the comparisons to be made easily for the differeproblem types in both experiments. For the Eperiment 1 results, all data are from the framsum condition only. The top set of four bars dpicts the data for the carry problems from tfirst experiment. Clearly, speakers wait a velong time before beginning to speak when thare confronted with these difficult arithmetproblems. The next set of four bars representno-carry problems from the first experiment. can be seen, latencies to speak are much shoAnd, as described in the Results for Experim1, utterance durations do not change in respoto problem type, because speakers in the first
”,periment dealt with problem difficulty before

a

a

ethac

ea

ofg

ula-he, it

ig-

ut-re

earndlat-hehatt 2,ar-omee

FIG. 4. Time speakers spent preparing and producing frame-sum utterances, both experiments.

they began to speak, not while they were speing. Notice too how similar the striped and grparts of the bars (the duration of the answer isand the sum) are for the top four bars and forfour bars immediately below. This visual impression is simply the result we reported in Eperiment 1: Utterance durations were essentithe same in the carry and no-carry conditions

The middle set of four bars shows the datathe no-carry problems from the second expment. It is interesting to compare them with four bars immediately above. The problems the same (so the utterances are identical in tent); what is different between the two expements is that in the second, speakers were psured to begin to speak quickly. Clearly, latencto speak are much longer in the first experimwhere speakers had the luxury of planning. In
dition, utterance durations are longer in the s
ak-y

the-x-lly.forri-ereon-ri-res-iesnt,d-

ond experiment as well, indicating that some the work speakers did not do before initiatinspeech ended up being carried out during artiction. In Experiment 1, the average duration of tentire utterance was 1232 ms; in Experiment 2was 1771 ms. This difference is statistically snificant,F(1,48) 5 51.89, MSe 5 68795. Yet, al-though speakers extended the duration of theterance by about 500 ms when they wepressured to begin to speak quickly, it is also clfrom Fig. 4 that, overall, speakers in the secoexperiment spent less time planning and articuing their utterances overall than they did in tfirst experiment—about 1 s less. Recall also tspeakers were no less accurate in Experimennor were they any less fluent. Thus, it can be gued that the deadline causes speakers to becmore efficient, in that they can perform the sam


ec-amount of work in less time. Interestingly, even

t

odi

n

f

t

e

a

oigiti

e

r

th

dh

a

et

t

thattherob-, thed 7;d 4.the

igitshus,twoult

hemstic-heynes

enseas-ionsheo-e auliver,thece

durations were longer when the ones were more

einin-ibera,e-nd inillheto-t et

-er-


though this appears to be true, it is also clear left to their own devices, speakers do not choto speak in this maximally efficient manner.

Finally, the bottom two sets of four bars shthe data for the mixed addend problems usethe second experiment. The level of resolutpermitted by this graph does not allow the dferences in performance to be seen easily. Wthe figure does reveal is that the differences to problem difficulty in Experiment 2 are of much smaller magnitude than the difference tween carry and no-carry problems in Expement 1 (and also the difference in performadue to the deadline).

The most important conclusion that can drawn from these comparisons is that the difence in performance for the two experimentsnot quite a pay-now/pay-later tradeoff: PeopleExperiment 2 did not need to extend their utances by the same amount of time as they requfor initiation in Experiment 1. People pay laterthey choose not to plan, but their debt is small

All in all, Experiment 2 yielded remarkablinformative data about how incremental speers are or can be. The one result that is lessoperative is the odd effect of tens difficulty fthe mixed addend problems. (For the two-daddend problems, this odd effect of tens dculty was not observed: In the first experimelatencies and durations were longer when thtens were difficult to calculate; in the second periment, latencies were unaffected but dutions were longer for difficult tens.)

This unexpected result is likely due to an afact of the problems used in the mixed-addecondition. The easiest way to see this properto examine Appendix B, which includes all tstimuli from Experiment 2. Consider only thmixed addend problems—the ones that yielthis unexpected result. Keep in mind as well tfor these mixed addend problems, no calcution is required for the tens column. In the etens, easy ones condition, the digits in the dends for the tens place were 2, 3, 4, and 5;digits in the addends for the ones place wer2, 3, and 4. In the easy tens, hard ones condithe digits in the addends for the tens place w2, 3, 4, and 5; the digits in the addends for
ones place were 3, 4, 5, 6, and 7. Thus, in b

hat,ose

w in

onif-hat

dueabe-ri-ce

beer- is iner-ired

ifr.

yk-

co-r

of the easy tens conditions, the numerals had to be added together were similar for tens and ones. Now look at the hard tens plems. In the hard tens, easy ones conditiondigits used for the tens were exclusively 6 anthe digits used for the ones were 1, 2, 3, anAnd in the hard tens, hard ones condition,digits used for the tens were 6 and 7; the dused for the ones were 2, 3, 4, 5, 6, and 7. Tthere was much less overlap between the columns; indeed, the tens column in the diffictens condition virtually “popped out”. Recall tother data for the mixed addend probledemonstrating that one of the difficulties paripants had in trying to answer them is that thad some trouble parsing the tens and ocolumns properly. In the cases in which the twere always 6 or 7, this task becomes much ier, and thus latencies and utterance duratwere shorter. Therefore, this finding for tmixed addend problems is likely of little theretical significance; instead, it appears to bresult of the unique properties of the stimused for these particular problems. Moreothis property of the tens items did not dilute effect of ones difficulty: Latencies and utteran

ffi-nt,ex-

ra-

ti-ndy iseeedatla-sy

ad- the 1,

ion,erehe

difficult to calculate, as would be expected.

GENERAL DISCUSSION

This section will be organized into thremajor parts. First, we will summarize the maresults and discuss their implications for the cremental hypothesis. Second, we will descra model of language production (F. Ferrei2000) that can account for why radical incrmentality of the sort proposed by Wheeldon aLahiri (1997), for example, was not observedthe experiments described here. Finally, we wbriefly consider issues relating especially to tarithmetic task we required our participants perform, focusing particularly on how our results compare to those obtained by Brysbaeral. (1998).

Incrementality in Language Production

According to an incremental model of language production, people do not plan their utt

othances completely before they begin to talk. In-

A

inhnletatehnco

ketacysthheeeeuinrm c

a

en

noum t bda

gl Wcea.

eref t

ent,flu-butffi-n-b-er

ut-tederi-ms

iti-end

id

plecre-tedri-icu-ver-uldplen anotle,

a-s to se-

- the to

tionlemm- the

ei-ciesri-nd.

s tondi-ted

nts

76 FERREIRA

stead, they “time-share” the tasks of plannand articulating, and therefore speak once thave prepared the earlier part of their utterawithout necessarily having formulated later ements. According to a radically incremenmodel, speakers are incremental to the exthat they initiate articulation once they know tfirst phonological word of their utterances. Planing of the next phonological word takes pladuring that first one’s articulation, and so through to the end of the utterance.

The two experiments described here, tatogether, support a limited type of incremenproduction. Incrementality is not an architetural property of the language production stem; instead, it is a parameter of production is under speaker control. In addition, even wa speaker has every incentive to initiate spequickly, production latencies reveal influencof utterance-final material (in the simple, onclause sentences we had our speakers prodThus, language production is not radically cremental—that is, speakers do not initiate pduction before having formulated at least soaspects of the utterance that go beyond therent phonological word.

The results that support these conclusionsthe following.

(1) The task used in Experiment 1 providlittle incentive for people to speak incremetally. Results showed that participants did initiate speech until they had planned the scarefully. The extent of planning was the saregardless of whether the sum occurred atbeginning of the utterance, at the end, or byself. Utterance durations were not influencedproblem difficulty, suggesting that speakers not engage in arithmetic calculations while ticulating.

(2) The task used in Experiment 2 stronencouraged people to speak incrementally.found that utterance durations were influenby problem difficulty. Therefore, it appears thspeakers were calculating as they articulatedthe same time, latencies were also affectedproblem difficulty, suggesting that speakplanned at least some aspects of the sum bspeaking, even though the sum occurred at
end of the utterance. Planning effects were le
ND SWETS

geyce-lnte-en

nl--

atnchs-ce).-o-eur-

re

d-t

meheit-y

idr-

yedtAtbysorehe

extensive than they were in the first experimhowever. In Experiment 2, latencies were inenced by the difficulty of the ones column not that of the tens; in Experiment 1, the diculty of both column calculations affected latecies. For both experiments, more difficult prolems overall were associated with longutterance initiation times: In Experiment 1,terances for no-carry problems were initiasooner than those for carry problems; in Expment 2, utterances for mixed addend proble(a one-digit and a two-digit addend) were inated sooner than those for two two-digit addproblems.

(3) It might be argued that Experiment 2 dencourage people to speak more incrementallythan they did in Experiment 1, but that peomight be capable of speaking even more inmentally still. This argument cannot be rebutdefinitively, but given that latencies in Expement 2 were reduced by over 70%, and partlarly given that they were about 650 ms on aage, it is hard to imagine that people coinitiate speech any faster. After all, when peoare asked to simply name a single word ocomputer monitor, their latencies are often much lower than 600 ms or so (for exampDuffy, Henderson, & Morris (1989) reported ltencies of 608 ms on average for participantname the word that occurred at the end of amantically neutral sentence, e.g.,the womansaw the MOUSTACHE). Thus, the average latency in Experiment 2 is not much more thantime it would have taken participants simplyread and say out loud the phrase the answer. Yet,even with such short latencies, the reactimes managed to display influences of probdifficulty. Furthermore, accuracy was not copromised: Participants were as accurate insecond experiment as they were in the first. Nther were speakers any less fluent: Disfluenoccurred on 1.78% of trials in the first expement, but on less than 1% of trials in the seco

(4) The phrasing the answer is SUMdoes notappear to be an unnatural way for speakerstate the sum of an arithmetic problem, as icated by the results of the pilot study conducin association with Experiment 2. Participa
sswere free to use any utterance form they wished

N

e -adrnbtrye onily o

thyethimeutd

eeri-

r

ttaaae ia

o

alex-ia-c-ksseo-er-oog ale

ionally, in-niti-gi-ndr-en-u- a

ies.nd asr-

s antheen the

edwillally9)chfor

d-i,&

dam-

le-


in which to state the sum, as long as they usfull sentence. Participants spontaneously hitthe form the answer is SUM, and they never wavered from it. This result should be viewed tentative, because the experiment was not signed to test directly whether utterance foand content vary in response to online demaof utterance formulation. The result may viewed as suggestive, however. It hints at possibility that the following general stomight be right: Speakers might sometimchange the form and content of an utteranceline as they become aware that the utterathey are producing is not likely to end happFor example, a speaker might change fromNoone has to give money to a group that they dlike what they’re doingto something like No onehas to give money to groups whose actions don’t approve ofin order to avoid a subjacencviolation (Chomsky, 1977). But they might bless likely to change the form and content of utterance in order to give themselves more tmerely to formulate some later part of an uttance; instead, the results of Experiment 2 sgest that they would choose instead to streout the utterance to give themselves the neetime.

Clearly, the language production systemnot structured such that “processing at all levoccurs in an incremental fashion with a procsor being triggered by any piece of charactetic input from the processors that feed into (Wheeldon & Lahiri, 1997, p. 361). In this “triggering” view, which is what we have terme“architectural incrementality” (and which pemits radical incrementality), a module of thlanguage production system that encountersformation stated in its vocabulary is called inaction automatically, so it performs its computions obligatorily (Fodor, 1983). Architecturincrementality is ruled out in favor of whmight be termed strategic incrementality: A dcision that every speaker must make is howstrike the appropriate balance between plannand initiating speech quickly. The finding thspeakers are in principle capable of speakingcrementally without any cost in accuracy sugests that, in fact, speakers prefer to plan m
carefully than they absolutely need to. Indee
GUAGE PRODUCTION 77

d aon

se-

mdsehe

sn-ce.

n’t

ey

ee

r-g-ched

islss-is-t”

d-ein-o-

lt-tongtin-g-re

recall from Fig. 4 that utterances with identiccontent were overall shorter in the second periment than in the first (adding together inittion plus articulation time). Thus, people are atually more efficient when they speaincrementally: They accomplish more in letime. For whatever reason, however, most pple are not inclined to speak in this manner. Phaps speaking this incrementally is simply ttaxing (just as one may be capable of runnin7-min mile but may prefer a more comfortab9-min pace).

Moreover, although the language operatcan (but does not have to) operate incrementit does not appear that the system is radicallycremental, in the sense that speech may be iated once the content of just the first phonolocal phrase is known. Even in the secoexperiment, in which participants initiated utteances in about 600 ms, people clearly were gaging in arithmetic calculations as they articlated, and in general participants spoke inmaximally efficient manner; even so, latencwere still influenced by problem difficultyThese results are in line with what Griffin aher colleagues (1999, 2000) have observedwell: When participants had to produce utteances consisting of a stock carrier phrase pluobject name, latencies were influenced by frequency of the pictured object’s name evwhen the object name occurred at the end ofsentence (i.e.,They saw the OBJECT).

TAG-Based Model of Language Production

The model of language production proposby F. Ferreira (2000) predicts that speakers be unable to produce utterances in the radicincremental manner proposed by Levelt (198and Wheeldon and Lahiri (1997). This approaassumes that the representational format syntactic information is a version of a Tree-Ajoining Grammar (TAG; Frank, 1992; Josh1985; Joshi, Levy, & Takahashi, 1975; Kroch Joshi, 1985). The basic unit of a TAG is the ele-mentary tree, which consists of a lexical heaand the arguments the head licenses. For exple, access of the word readwould result in acti-vation of not just the word but its associated e
d,mentary tree as well, as shown below.

A

hiz

s

-j

a

yie

r

ical beati- cany tovail-alsed

n.d-

ro-an-erbmati-u-

hist 2entalionte-

red in-

uc-ud-der

eechce’sw-celta-ndhtnds-rbsak- by

sbs,pre-d inndausehac-t-

e

eo

78 FERREIRA

This structure assumes the analysis of claupresented in Chomsky (1986), according which a clause is an Inflectional Phrase (abbviated as IP), and a full clause including tnode for a complementizer is a ComplementPhrase (CP). The abbreviation DP standsDeterminer Phrase (Abney, 1988). Other abbviations are fairly standard, NP for noun phraPP for prepositional phrase, and VP for vephrase.

The verb read is the lexical head and it licenses two arguments—a subject and an obTAG assumes that a verbal head projects only its own VP node but all the clausal projetions as well (IP and CP). Thus, elementtrees are prototypically clause-like. Indeed, thare often described as corresponding roughla simple clause (Kroch, 1987), and as besimilar to Chomsky’s (1955) original kernsentences (Frank, 1992). Now let us examthe amount of syntactic structure that is trieved when a noun such as answerbecomesactivated.

CP

C IP

DP 1’

I VP

V DP

read

DP

D NP

the N’

N

answer

The determiner the brings along its DP nodand licenses an NP argument. The noun answercomes with its NP structure, and by a procknown as substitution(Frank, 1992), plugs intthe NP slot provided by the DP. The critic

cen-point for our purposes here is that neither thenor answercan project any further; in particula

ND SWETS

sestore-eerforre-e,rb

ect.notc-ryey tonglinee-

no clausal nodes are licensed by these lexitems. The implication is that this DP cannotmore than a DP—it cannot acquire a grammcal role such as subject or object, because itreceive no clausal assignment. The only waget clausal nodes is for a verb to become aable. Thus, once is is formulated, then a claustree like the one shown previously is accesalong with the lexical item, and then the DP theanswercan be plugged into its subject positioIt is only at that point that phonological encoing and articulation may begin.

This TAG-based approach to sentence pduction, then, predicts that incrementality cnot be more radical than the subject plus vsequence, because a DP cannot get a gramcal role in the sentence until a verb is formlated. Radical incrementality is ruled out on tmodel. Of course, the results of Experimensuggest that the system is even less incremthan the TAG model allows, because initiattimes showed influences of postverbal marial—i.e., the sum. This much planning occureven though the task gave speakers strongcentives to speak incrementally. In the Introdtion and in F. Ferreira (2000), a number of sties are discussed demonstrating that, unsome circumstances, speakers initiate sponce they have formulated just the sentensubject and verb (e.g., Lindsley, 1975). Hoever, Meyer’s (1996) study provided evidenthat grammatical encoding requires the simuneous formulation of both preverbal apostverbal material. A critical difference migbe the type of verb that heads the clause: Liley’s study employed regular agent-theme vesuch as touch, whereas the sentences that speers produced in Meyer’s study were headedthe copula to be. It is worth noting that TAGtreat copular verbs differently from other versuch that both “arguments” are in essence verbal. Thus, the results reported here anMeyer indicating that speakers plan beyoeven the verb might have been observed becspeakers used to be as the main verb in botstudies. Of course, this hypothesis will not count for Griffin’s (1999, 2000) results (the uterances speakers produced in her study

ss

al

tered around verbs such as saw), so clearly thisr,

N

ode

ar

a

h pvewg

c athioaucet

lsaw

u

blfoarIt

to

ad-le ind.ed are isV.atveey

ure.abces-

.uple.atheal

lts ut-d of

s-ce

iffi-onseirtion,la-as

sizellypro-onin-1;re-ithar-dic-u-n

cu-e

en,n in


issue needs further investigation. What is imptant for our purposes here is that the TAG moputs a lower bound on incrementality: Speakcannot initiate a sentence until they have formlated its subject and verb. Under some circustances—perhaps when the verb is a copulperhaps when a demanding task such as ametic is required—speakers must formulate encode even postverbal material.

Arithmetic and Language Production

One of the most important inspirations for texperiments reported here was the studyBrysbaert et al. (1998), which seems to supradical incrementality. Brysbaert et al. obserthat speakers are faster to produce the ansto arithmetic problems consisting of a two-diaddend and a one-digit addend when the dends are arranged so that the first columnculated was also the first phonological wordticulated. Thus, speakers of Dutch prefer single digit addend to precede the double-daddend, because numbers in Dutch are spso that the ones come before the tens. Speof French prefer the opposite order, becaFrench works just like English: A number suas 44 is spoken so that the tens column precthe ones column. This result seems consiswith radical incrementality.

The results of Experiment 2 are entirely cosistent with the data (but not quite the interpretion) found in Brysbaert et al. (1998). We afound that, for mixed addend problems, speers preferred the arrangement in which the tdigit number came first. Of course, we wougive our result a different explanation, particlarly because the preference for the xx1 x orderemerged in the latency data, many syllablesfore speakers had to articulate the sum itsemight be, then, that the Brysbaert effect ccerns planning: Speakers want to formulate encode their utterances roughly in the ordewhich the constituents will be articulated. other words, even if people planned the enutterance The answer is SUMbefore beginningto speak (to some extent), there is a quesabout the order in which that planning toplace: Did they plan the sum and then the an-
swer is, or did they plan in the other order? Th
GUAGE PRODUCTION 79

r-elrsu-m-, orith-nd

ebyortdersitad-al-r-e

gitkenkerssehdesent

n-ta-ok-o-

ld-

e-. Itn-nd innire

ionk

finding that people preferred to receive the dends in the xx 1 x order suggests that peopmight have planned the utterances roughlythe order in which they would be articulateThe Brysbaert et al. finding, then, can be viewas demonstrating that, in general, speakersinclined to plan in this manner. Indeed, thisthe version of incrementality argued for by Ferreira (1996) in his work demonstrating thpeople speak more efficiently when they hachoices regarding syntactic form than when thare constrained to just one syntactic structOn this view of incrementality, active items grearlier syntactic positions, and as a result, acsibility of concepts influences syntactic formThus, syntactic plans are preferentially built in the order in which words become availabThis type of incrementality does not imply thphonological encoding will take place on tsmallest unit possible (viz. the phonologicword), and it is compatible with our resushowing that speakers plan the sum even interances in which the sum occurs at the enthe utterance.

Furthermore, it is critical to note that Brybaert et al. (1998) did not provide any evidenthat speakers were not influenced by the dculty of both the tens and the ones calculatifor the sum before beginning to speak. Thstudy was not designed to address this quesso the difficulty of the tens and the ones calcutions were not independently manipulated they were here. Indeed, we should emphathat the Brysbaert et al. study was not originadesigned or reported as a study of language duction; instead, it is an important contributito the literature on numerical cognition and lguistic relativity hypothesis (Sapir, 194Whorf, 1956). Their finding that speakers pferred the articulation order is consistent wthe idea of radical incrementality but, as we gued above, is not mandated by it. The pretion that speakers of English would be inflenced by just the difficulty of the tens columwhen producing utterances requiring the callation of two-digit sums was an inference wdrew from their reported work. As we have seit was not supported, probably because eve
ethe original Brysbaert et al. study, speakers

A A

toich

tudane

coorsedoean o

o asinndr

on. O tontnothein,hae

inoth

were- any

teder

an- re-theIn-s atde-theave

tillondre,on-ialt- tocre-ntslan-

heyly

80 FERREIR

planned the entire sum before beginningspeak—they simply did so in the order in whthey would articulate.

Another quite different possibility is thaspeakers in the Brysbaert et al. (1998) stwere in fact speaking incrementally, rather thjust planning in a left-to-right order. It might bthat when speakers produce utterances thatsist simply of nominals (i.e., just the sum) other types of fragments, then the TAG-baclausal constraint that we described above dnot operate. The constraint would be irrelevbecause the utterance fragment consistingonly the sum does not need to be put intclausal structure. In this situation, perhaps a gle phonological word can be formulated aimmediately produced, because the speakenot producing a clause, so grammatical relatisuch as subject and object are undefinedcourse, this line of reasoning must cometerms with the finding from our first experimethat speakers in the sum-only condition did show evidence of being influenced by just tens column for the no-carry problems. Agawe can only speculate, but one possibility is tspeakers were unable to behave as incremtally as they did in Brysbaert et al. becauseour experiment participants had to add b
columns, and for some problems they had to engage in carry operations as well. Perhaps thicombination is critical: Many problems requiredcarrying, making it difficult for the participantsto be confident that they could deal with just thetens independently of the ones; and, all theproblems required participants to add the tencolumn as well as the ones column. These twcharacteristics together might have led participants to believe that their most efficient strategywas to deal with the problems as a whole befor
ND SWETS

y

n-

stf

-

issf

t

tn-

beginning to speak. Let us emphasize thatare simply speculating at this point: Further search must be done on these topics beforedefinitive arguments can be made.

Conclusions

The two experiments that we have reporshed light on the important question of whethlanguage production is incremental. The swer that is most consistent with the resultsported here and in previous work is that system is not architecturally incremental. stead, the extent to which planning occurs ileast partly under speakers’ control, and it pends on the intentions that motivate speech. Moreover, even when speakers hincentives to initiate speech quickly, they sappear to engage in planning that goes beythe immediate phonological word. Therefothe system seems to be architecturally cstrained to require planning beyond the initphonological word, particularly for clausal uterances. At the same time, it is importantstress that language production can be inmental: The results of these experimedemonstrate that speakers are capable of pning upcoming portions of an utterance as tare articulating. This finding is especial
-s
so-

e

striking given that the planning they engagedin was arithmetic calculation, because it mighthave been supposed that addition of two-digitnumbers could not be carried out concurrentlywith utterance articulation. Apparently, the twocan go on in parallel, at least for problems thatdo not require carrying operations. Thus, a fun-damental premise of the incremental view issupported: The language production system iscapable of interleaving planning processes andarticulation.

APPENDIX A

TABLE A1

Problems Used in Experiment 1

No-carry problems Carry problems Filler problems

Prob Sum Condition Prob Sum Condition Prob Sum

21 1 22 43 EE 231 28 51 easy 301 11 4121 1 26 47 EH 231 38 61 hard 601 15 7521 1 62 83 HE 231 58 81 harder 131 79 9221 1 66 87 HH 231 68 91 hardest 701 17 8723 1 21 44 EE 291 23 52 easy 751 13 8827 1 21 48 EH 391 23 62 hard 121 15 2763 1 21 84 HE 591 23 82 harder 661 30 9667 1 21 88 HH 691 23 92 hardest 541 11 6521 1 24 45 EE 241 27 51 easy 181 36 5421 1 28 49 EH 241 37 61 hard 121 58 7021 1 64 85 HE 241 57 81 harder 181 72 9021 1 68 89 HH 241 67 91 hardest 121 61 7331 1 21 52 EE 28 1 24 52 easy 43 1 13 5635 1 21 56 EH 38 1 24 62 hard 73 1 12 8571 1 21 92 HE 58 1 24 82 harder 77 1 10 8775 1 21 96 HH 68 1 24 92 hardest 77 1 19 9621 1 32 53 EE 24 1 29 53 easy 19 1 77 9621 1 36 57 EH 24 1 39 63 hard 17 1 49 6621 1 72 93 HE 24 1 59 83 harder 10 1 78 8821 1 76 97 HH 24 1 69 93 hardest 18 1 46 6433 1 21 54 EE 27 1 25 52 easy 68 1 10 7837 1 21 58 EH 37 1 25 62 hard 62 1 11 7373 1 21 94 HE 57 1 25 82 harder 36 1 17 5377 1 21 98 HH 67 1 25 92 hardest 36 1 18 5421 1 34 55 EE 25 1 28 53 easy 12 1 51 6321 1 38 59 EH 25 1 38 63 hard 40 1 43 8321 1 74 95 HE 25 1 58 83 harder 60 1 12 7221 1 78 99 HH 25 1 68 93 hardest 30 1 42 7231 1 22 53 EE 29 1 25 54 easy 31 1 60 9135 1 22 57 EH 39 1 25 64 hard 76 1 10 8671 1 22 93 HE 59 1 25 84 harder 77 1 17 9475 1 22 97 HH 69 1 25 94 hardest 51 1 30 8122 1 32 54 EE 26 1 27 53 easy 13 1 55 6822 1 36 58 EH 26 1 37 63 hard 18 1 13 3122 1 72 94 HE 26 1 57 83 harder 30 1 33 6322 1 76 98 HH 26 1 67 93 hardest 17 1 73 9033 1 22 55 EE 28 1 26 54 easy 17 1 14 3137 1 22 59 EH 38 1 26 64 hard 62 1 14 7673 1 22 95 HE 58 1 26 84 harder 19 1 16 3577 1 22 99 HH 68 1 26 94 hardest 16 1 11 2723 1 22 45 EE 26 1 29 55 easy 18 1 17 3523 1 26 49 EH 26 1 39 65 hard 50 1 16 6623 1 62 85 HE 26 1 59 85 harder 11 1 61 7223 1 66 89 HH 26 1 69 95 hardest 10 1 40 5031 1 23 54 EE 28 1 27 55 easy 66 1 18 8435 1 23 58 EH 38 1 27 65 hard 39 1 15 5471 1 23 94 HE 58 1 27 85 harder 12 1 70 82

INCREMENTALISM IN LANGUAGE PRODUCTION 81

75 1 23 98 HH 68 1 27 95 hardest 551 14 69

. The terms


TABLE A1—Continued

Problems Used in Experiment 1.

No-carry problems Carry problems Filler problems

Prob Sum Condition Prob Sum Condition Prob Sum23 1 32 55 EE 27 1 29 56 easy 18 1 54 7223 1 36 59 EH 27 1 39 66 hard 19 1 69 8823 1 72 95 HE 27 1 59 86 harder 40 1 43 8323 1 76 99 HH 27 1 69 96 hardest 12 1 42 5431 1 24 55 EE 29 1 28 57 easy 14 1 17 3135 1 24 59 EH 39 1 28 67 hard 45 1 30 7571 1 24 95 HE 59 1 28 87 harder 71 1 10 8175 1 24 99 HH 69 1 28 97 hardest 65 1 30 95

Note. EE, easy tens, easy ones; EH, easy tens, hard ones; HE, hard tens, easy ones; HH, hard tens, hard ones
“easy”, “hard”, “harder”, and “hardest” refer to the size of the sum of the carry problems.
APPENDIX B

TABLE A2


Mixed addend problems Two-digit addend problems

Prob Sum Condition Prob Sum Condition2 1 21 23 EE 21 1 22 43 EE22 1 2 24 EE 23 1 21 44 EE2 1 31 33 EE 21 1 24 45 EE41 1 2 43 EE 31 1 21 52 EE2 1 51 53 EE 21 1 32 53 EE21 1 3 24 EE 33 1 21 54 EE3 1 31 34 EE 21 1 34 55 EE41 1 3 44 EE 31 1 22 53 EE3 1 22 25 EE 22 1 32 54 EE51 1 3 54 EE 33 1 22 55 EE4 1 21 25 EE 23 1 22 45 EE31 1 4 35 EE 31 1 23 54 EE4 1 41 45 EE 23 1 32 55 EE51 1 4 55 EE 31 1 24 55 EE5 1 24 29 EH 21 1 26 47 EH33 1 5 38 EH 27 1 21 48 EH6 1 22 28 EH 21 1 28 49 EH23 1 6 29 EH 35 1 21 56 EH6 1 32 38 EH 21 1 36 57 EH33 1 6 39 EH 37 1 21 58 EH6 1 42 48 EH 21 1 38 59 EH43 1 6 49 EH 35 1 22 57 EH6 1 52 58 EH 22 1 36 58 EH53 1 6 59 EH 37 1 22 59 EH7 1 22 29 EH 23 1 26 49 EH32 1 7 39 EH 35 1 23 58 EH7 1 42 49 EH 23 1 36 59 EH
52 1 7 59 EH 35 1 24 59 EH


TABLE A2—Continued


Mixed addend problems Two-digit addend problems

Prob Sum Condition Prob Sum Condition1 1 62 63 HE 21 1 62 83 HE72 1 1 73 HE 63 1 21 84 HE2 1 61 63 HE 21 1 64 85 HE62 1 2 64 HE 71 1 21 92 HE2 1 71 73 HE 21 1 72 93 HE72 1 2 74 HE 73 1 21 94 HE2 1 63 65 HE 21 1 74 95 HE73 1 2 75 HE 71 1 22 93 HE3 1 61 64 HE 22 1 72 94 HE71 1 3 74 HE 73 1 22 95 HE3 1 62 65 HE 23 1 62 85 HE72 1 3 75 HE 71 1 23 94 HE4 1 61 65 HE 23 1 72 95 HE71 1 4 75 HE 71 1 24 95 HE4 1 74 78 HH 21 1 66 87 HH75 1 4 79 HH 67 1 21 88 HH5 1 62 67 HH 21 1 68 89 HH63 1 5 68 HH 75 1 21 96 HH5 1 64 69 HH 21 1 76 97 HH72 1 5 77 HH 77 1 21 98 HH5 1 73 78 HH 21 1 78 99 HH74 1 5 79 HH 75 1 22 97 HH6 1 62 68 HH 22 1 76 98 HH63 1 6 69 HH 77 1 22 99 HH6 1 72 78 HH 23 1 66 89 HH73 1 6 79 HH 75 1 23 98 HH7 1 62 69 HH 23 1 76 99 HH72 1 7 79 HH 75 1 24 99 HH

o .

inn

o

i-

d

t

:

n-

-ce:

ex-

Note. EE, easy tens, easy ones; EH, easy tens, hard

REFERENCESAbney, S. (1988). A computational model of human pars

Paper presented at the First Annual CUNY Confereon Human Sentence Processing.

Ashcraft, M. J. (1992). Cognitive arithmetic: A review data and theory. Cognition, 44, 75–106.

Bock, J. K. (1986). Meaning, sound, and syntax: Lexical pring in sentence production. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 575–586.

Bock, J. K., & Cutting, J. C. (1992). Regulating mental eergy: Performance units in language production. Jour-nal of Memory and Language, 31, 99–127.

Bock, J. K., & Levelt, W. J. M. (1994). Language prodution: Grammatical encoding. In M. Gernsbacher (EHandbook of psycholinguistics(pp. 945–984). SanDiego: Academic Press.

Brysbaert, M., Fias, W., & Noel, M.-P. (1998). The Whofian hypothesis and numerical cognition: Is “twen
four” processed in the same way as “four-and-twentyCognition, 66, 51–77.
nes; HE, hard tens, easy ones; HH, hard tens, hard ones

g.ce

f

m-

n-

c-.),

r-y-

Chomsky, N. (1955). The logical structure of linguistic theory.Bloomington, IN: Indiana University Linguistics Club.

Chomsky, N. (1977). Essays on form and interpretation.New York: North-Holland.

Chomsky, N. (1986). Knowledge of language. New York:Praeger.

Clark, H. H., & Clark, E. (1977). Psychology and languageAn introduction to psycholinguistics. New York: Har-court Brace.

Clark, H. H., & Wasow, T. (1998). Repeating words in spotaneous speech. Cognitive Psychology, 37, 201–242.

Duffy, S. A., Henderson, J. M., & Morris, R. K. (1989). Semantic facilitation of lexical access during sentenprocessing. Journal of Experimental PsychologyLearning, Memory, and Cognition, 15, 791–801.

Ferreira, F. (1991). Effects of length and syntactic complity on initiation times for prepared utterances. Journalof Memory and Language, 30, 210–233.

”?Ferreira, F. (1993). Creation of prosody during sentence pro-duction. Psychological Review, 100, 233–253.


Ferreira, F. (2000). Syntax in language production: An ap-proach using tree-adjoining grammars. In L. Wheeldon

y

fo

me

m

.p

p

-

gi-

s

iod

in

6oy

Holmes, V. M. (1988). Hesitations and sentence planning.Language and Cognitive Processes, 3, 323–361.

e-ns:n,-c-iv.

ctci-

encyer

of5-

ci-

-

far

pro-ri-

ngar--

pir

.uc-

NP

ch

d

(Ed.), Aspects of language production(pp. 291–330).Cambridge, MA: MIT Press.

Ferreira, V. (1996). Is it better to give than to donate? Stactic flexibility in language production. Journal ofMemory and Language, 35, 724–755.

Fodor, J. A. (1983). The modularity of mind. Cambridge,MA: MIT Press.

Ford, M. (1982). Sentence planning units: Implications the speaker’s representation of meaningful relatiunderlying sentences. In J. Bresnan (Ed.),The mentalrepresentation of grammatical relations(pp. 797–827).Cambridge, MA: MIT Press.

Frank, R. (1992). Syntactic locality and tree adjoining gramar: Grammatical, acquisition, and processing pspectives. Ph.D. thesis, University of Pennsylvania.

Fromkin, V. A. (1971). The non-anomalous nature of anoalous utterances. Language, 47, 27–52.

Garrett, M. F. (1975). The analysis of sentence productionG. H. Bower (Ed.),The psychology of learning and motivation, 9, (pp. 133–177). New York: Academic Press

Garrett, M. F. (1976). Syntactic processes in sentence duction. In R. J. Wales & E. Walker (Eds.),New ap-proaches to language mechanisms(pp. 231–256). Am-sterdam: North-Holland.

Garrett, M. F. (1980). Levels of processing in sentence duction. In B. Butterworth (Ed.),Language production.Vol.1: Speech and talk(pp. 177–220). New York: Academic Press.

Geary, D., Bow-Thomas, C., & Yao, Y. (1992). Countinknowledge and skill in cognitive addition: A comparson of normal and mathematically disabled childreJournal of Experimental Child Psychology, 54,372–391.

Goldman-Eisler, F. (1968). Psycholinguistics: Experimentin spontaneous speech. New York: Academic Press.

Griffin, Z. M. (1999). The trade-off between lexical preparatand (dis)fluency for simple utterances. Talk presentethe USC Language Production Conference, Nov.

Griffin, Z. M., & Spieler, D. H. (2000). Speech planning younger and older adults. Poster presented at thenual Meeting of the Psychonomic Society, New Oleans, LA, Nov.

Henderson, A., Goldman-Eisler, F., & Skarbek. A. (196The common value of pausing time in spontanespeech. Quarterly Journal of Experimental Psycholog,17, 343–345.

n-

orns

-r-

-

. In-

ro-

ro-

n.

n at

An-r-

).us

Joshi, A. K. (1985). How much context-sensitivity is rquired to provide reasonable structural descriptioTree adjoining grammars. In D. Dowty, L. Kartunneand A. Zwicky (Eds.),Natural language parsing: Psychological, computational, and theoretical perspetives. (pp. 206–250). Cambridge: Cambridge UnPress.

Joshi, A. K., Levy, L., & Takahashi, M. (1975). Tree adjungrammars. Journal of the Computer and System Sences, 10, 136–163.

Kroch, A. (1987). Unbounded dependencies and subjacin a tree adjoining grammar. In A. Manaster-Ram(Ed.), Mathematics of Language(pp. 143–172) Ams-terdam: John Benjamins.

Kroch, A., & Joshi, A. K. (1985). The linguistic relevance tree adjoining grammar. Technical Report MS-CS-816, Department of Computer and Information Sences, University of Pennsylvania.

Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press.

Lindsley, J. R. (1975). Producing simple utterances: Howahead do we plan? Cognitive Psychology, 7, 1–19.

Meyer,A. S. (1996). Lexical access in phrase and sentenceduction: Results from picture-word interference expements. Journal of Memory and Language, 35, 477–496.

Roelofs, A. (1998). Rightward incrementality in encodisimple phrasal forms in speech production: Verb–pticiple combinations. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24,904–921.

Sapir, E. (1941). In L. Spier (Ed.),Language, culture, andpersonality: Essays in memory of Edward SaMenasha, WI: Sapir Memorial Publication Fund.

Stallings, L. M., MacDonald, M. C., & O’Seaghdha, P. G(1998). Phrasal order constraints in sentence prodtion: Phrase length and verb disposition in heavy-shift. Journal of Memory and Language, 39, 392–417.

Wheeldon, L., & Lahiri, A. (1997). Prosodic units in speeproduction. Journal of Memory and Language, 37,356–381.

Whorf, B. (1956). Language, thought, and reality: Selectewritings of Benjamin Lee Whorf. New York: Wiley.

(Received July 31, 2000)(Revision received January 22, 2001)(Published online October 3, 2001)

journal of memory and language · the incrementality hypothesis is appealing because it accounts...

Documents