Heavy NP shift - University of Massachusetts Amherst

18 Pages·2006·0.55 MB·English

Journal of
Memory and
Language
JournalofMemoryandLanguage54(2006)389–406
www.elsevier.com/locate/jml
Heavy NP shift is the parser’s last resort:
q
Evidence from eye movements
Adrian Staub a,*, Charles Clifton Jr. a, Lyn Frazier b
aDepartmentofPsychology,UniversityofMassachusetts,Amherst,MA01003,USA
bDepartmentofLinguistics,UniversityofMassachusetts,Amherst,MA01003,USA
Received19July2005;revisionreceived2December2005
Availableonline2February2006
Abstract
Twoeyemovementexperimentsexploredtherolesofverbalsubcategorizationpossibilitiesandtransitivitybiasesin
theprocessingofheavyNPshiftsentencesinwhichtheverb’sdirectobjectappearstotherightofapost-verbalphrase.
InExperiment1,participantsreadsentencesinwhichaprepositionalphraseimmediatelyfollowedtheverb,whichwas
either obligatorily transitive or had a high transitivity bias (e.g., Jack praised/watched from the stands his daughter’s
attempt to shoot a basket). Experiment 2 compared unshifted sentences to sentences in which an adverb intervened
betweentheverbanditsobject,andobligatorilytransitiveverbstooptionallytransitiveverbswithwidelyvaryingtran-
sitivitybiases.Inbothexperiments,evidenceofprocessingdifficultyappearedonthematerialthatintervenedbetween
the verb and its object when the verb was obligatorily transitive, and on the shifted direct object when the verb was
optionallytransitive,regardlessoftransitivitybias.WeconcludethattheparseradoptstheheavyNPshiftanalysisonly
when it is forced to by the grammar, which we interpret in terms of a preference for immediate incremental
interpretation.
(cid:1)2005Elsevier Inc. Allrights reserved.
Keywords: Sentencecomprehension;Parsing;HeavyNPshift;Argumentstructure;Eyemovements
Considerthe comprehensionofsentencesinwhich a (1) Lucy ate t with a fork [the extremely delicious,
i
direct object is separated from the verb by intervening bright green broccoli].
i
material, as in (1)below:
This construction is known as heavy NP shift.
According to the classic syntactic account (Ross, 1967;
but cf. Kayne, 1998; Rochemont & Culicover, 1997),
q This research was supported by National Institutes of heavyNPshiftinvolvesamovementoperationthatdis-
Health Grant HD-18708 to the University of Massachusetts places the verb’s direct object from its underlying posi-
and by a University of Massachusetts Graduate School tion adjacent to the verb, where it receives both
Fellowship to the first author. Portions of this research were accusativecaseandathematicrole,toanadjoinedposi-
presented at the Eighteenth Annual CUNY Conference on
tion to the right of other verbal arguments and/or
HumanSentenceProcessing,Tuscon,Arizona,April2005.We
adjuncts. As in other movement operations, a trace of
thankKeithRaynerforhishelpfulcomments.
* Correspondingauthor.Fax:[phone]. the moved constituent remains in the pre-movement
E-mailaddress:[email protected](A.Staub). position. Notably, heavy NP shift is typically able to
0749-596X/$-seefrontmatter (cid:1)2005ElsevierInc.Allrightsreserved.
doi:10.1016/j.jml.2005.12.002
390 A.Staubetal./JournalofMemoryandLanguage54(2006)389–406
apply only when the direct object noun phrase is rela- In sentences that involve wh-movement, the filler is
tivelylongor‘‘heavy’’relativetotheinterveningconstit- usually easily identifiable, but the gap is not signaled
uent, as evidenced by the comparative awkwardness of unambiguously. It is possible for the parser to posit a
(2a)andthe unacceptability of (2b): gap in a position that turns out to be incorrect, or to
overlookthetruegapsite.Considerthefollowingmod-
(2a) Lucy ate t with afork[the broccoli]. ificationof (3a)above:
i i
(2b) Lucy ate t with afork[peas].
i i
(4) [Which boy] did the teacher punish the girls for
i
To our knowledge, there have been no published teasing ti?
experimental studies of the comprehension of heavy
NP shift constructions. (We know of only one experi- As (3a) indicates, the gap corresponding to which boy
mental study of heavy NP shift production: Stallings, could occur after punish; but as (4) indicates, this gap
MacDonald, & O’Seaghdha, 1998.) This is surprising couldinsteadappearlaterinthesentence,afterteasing.
fortworeasons.First,heavyNPshiftisrelativelycom- Areaderorlistenerwhodoespositagapsiteafterpun-
mon. Wasow (1997) found that in a corpus of printed ishwillhavetoreanalyzeuponencounteringthegirlsin
materials, heavy NP shift accounted for between 5 and (4).Ontheotherhand,areaderorlistenerwhodoesnot
10% of all sentences with direct objects, with this rate initially posit a gap site after punish will have to do so
varying by verb type (in a manner to be discussed after encountering for his in (3a), since punish requires
below). In a separate corpus study focusing on certain a direct object.
idiomatic constructions, Wasow found heavy NP shift In an influential paper, Fodor (1978) distinguished
ratesof around50%. three strategies or principles that the processor could
Second, and perhaps more importantly, heavy NP use to identify the gap location. The processor could
shifthasmuchincommonwithwh-movement,thepro- adopt a highly conservative last resort strategy, posit-
cessing of which has been studied in great detail (e.g., ing a gap only when the grammar requires one. On
Aoshima, Phillips, & Weinberg, 2004; Boland, Tanen- the other hand, the processor could adopt the liberal
haus, Garnsey, & Carlson, 1995; Clifton & Frazier, first resort strategy of positing a gap in the first gram-
1989; Crain &Fodor, 1985; Fodor, 1978, 1989; Frazier matically licensed location. Finally, the parser could
& Clifton, 1989; Garnsey, Tanenhaus, & Chapman, posit a gap following a particular verb only after
1989; Pickering & Traxler, 2003; Stowe, 1986; Traxler assessing whether the verb frequently takes an argu-
&Pickering,1996).Wh-movementoccursinbothques- ment of the same type as the filler (e.g., a noun phrase
tions and relative clauses, as in (3a) and (3b), or prepositional phrase), and noting whether an argu-
respectively: ment of this type appears following the verb. If the
verb does tend to take an argument of the relevant
(3a) [Whichboy] didtheteacherpunisht forhisbad type, and one does not immediately appear, then the
i i
behavior? processor posits a gap after the verb. This is the lex-
(3b) The teacher punished the boy [whom] she had ical expectation strategy.
i
asked t tostoprunning around. The last resort strategy predicts that the parser will
i
never posit a gap when an alternative analysis is avail-
LikeheavyNPshift,wh-movementcandisplaceadirect able. Fodor notes that this prediction is inconsistent
object noun phrase (or other argument of a verb) from with clear intuitions about processing difficulty, and
its underlying position, leaving in this position a trace research using on-line measures such as eyetracking
of movement (in psycholinguistic terminology, a gap). and self-paced reading has consistently provided evi-
Incomprehending a sentence in which either heavy NP dencethatreadersdonotfollowthelastresortstrategy
shiftorwh-movementhastakenplace,thereaderorlis- (Crain & Fodor, 1985; Stowe, 1986). In fact, a fair
tener must associate the moved constituent (or filler) amountofevidencenowsupportsthefirstresortstrate-
with the underlying position in which this constituent gy over the lexical expectation strategy (e.g., Frazier &
isassignedathematicrolebyaverb.Animportantdif- Clifton,1989;Pickering&Traxler,2003;Traxler&Pic-
ferencebetweentheseconstructions,however,isthatwh- kering, 1996; but cf. Boland et al., 1995). For example,
movement results in a word order in which the direct Pickering and Traxler (2003) tracked participants’ eye
objectorotherargumenthasmovedclosertothebegin- movements as theyreadsentenceslike (5a and5b):
ningofthesentence,whileheavyNPshiftdisplacesthe
direct object to a position later in the sentence. As a (5a) That’s the plane that the pilot landed behind in
result,theorderinwhichareaderorlistenerencounters the fog atthe airport.
the filler and the corresponding gap is reversed in these (5b) That’s the truck that the pilot landed behind in
twoconstructions. the fog atthe airport.
A.Staubetal./JournalofMemoryandLanguage54(2006)389–406 391
Participants took longer to read the region that NP shift continuations are possible, e.g., Lucy ate with
included the verb landed in (5b), and took longer to a fork rather than a spoon. In this case, the heavy NP
read the next region in (5a). This suggests that readers shiftanalysisisnotforceduntiltheshiftednounphrase
were interpreting the plane or the truck as the direct itselfis encountered.
object of landed, even though the verb is less Thedifferencesbetweenthepredictionsofthelexical
frequently used with a noun phrase complement than expectation and last resort strategies center on the role
with a prepositional phrase complement. In (5a), this of verb subcategorization information. According to
analysis can no longer be maintained when the reader the lexical expectation strategy, the point at which the
reaches in the fog; in (5b), the implausibility of land- parser constructs the heavy NP shift analysis should
ing a truck causes the reader to reanalyze almost vary depending on verb subcategorization bias, with
immediately. high-transitiveverbsincreasingthelikelihoodofanearly
In this article, we present two experiments designed (i.e.,ontheimmediately post-verbalmaterial) adoption
to examine the question of when, in the course of pro- ofthisanalysis.Butaccordingtothelastresortstrategy,
cessing a heavy NP shift sentence, the processor first onlythepossibilityofanintransitiveframematters,not
constructs the heavy NP shift analysis. It is possible to thefrequencyofthisframe.Iftheverboccursonlyina
distinguish first resort, last resort, and lexical expecta- transitiveframe,theparserwillconstructtheheavyNP
tionstrategiesthatareanaloguestoFodor’s(1978)three shift analysis on the immediately post-verbal material;
strategiesfortheprocessingofwh-movement.Aproces- but if the verb allows an intransitive frame, the parser
sor that adopts the first resort strategy will rank the will not construct the heavy NP shift analysis until it
option of heavy NP shift above any other option; i.e., reaches the shifted noun phrase itself. In other words,
in (1) above it will posit a gap after the verb ate even the two strategies make different predictions for verbs
beforecheckingwhetheradirectobjectfollowsimmedi- thataretransitive-biasedbutnotobligatorilytransitive:
ately in the input. Intuitively this strategy seems quite the lexical expectation strategy predicts that with these
unlikely,sinceitsuggeststhatintegratingadirectobject verbs,theparserwillconstructtheheavyNPshiftanal-
after ate (e.g., Lucy ate the broccoli with a fork) should ysisonthematerialthatintervenesbetweentheverband
actually induce difficulty because the direct object posi- direct object, while the last resort strategy predicts that
tionisalreadyfilledbyatraceofmovement.Moreplau- theparserwillnotconstructtheheavyNPshiftanalysis
sibleisalexicalexpectationstrategy,whichpredictsthat untillater.
the parser will posit a trace of a shifted noun phrase if Weassumethatthefirstresortstrategyforprocessing
theverbisrelativelytransitive-biasedandadirectobject wh-movement arises from a general preference to con-
doesnotimmediatelyappear.Ifthisistheparser’sstrat- struct syntactic analyses in a manner that maximizes
egy, then the heavy NP shift analysis will not be con- immediateincrementalinterpretation(Clifton&Frazier,
structed for (1) until the reader or listener begins to 1989;Crocker,1996;Frazier,1987;Pickering,1994).We
processwithafork.Eventhen,theparserwillconstruct thinkthatthesamegeneralpreferencearguesforthelast
the heavy NP shift analysis only if the verb eat is suffi- resortstrategyin the processing ofheavyNP shiftcon-
ciently transitive-biased. If eat is not sufficiently transi- structions. A preference for maximizing incremental
tive-biased, the parser will assume, upon encountering interpretation is evident in the processing of a range of
withafork,thateatisoccurringinanintransitiveframe. structural ambiguities, resulting in, for example, the
It will then be forced to construct the heavy NP shift well-attested preferences for Minimal Attachment and
analysis upon encountering the shifted noun phrase Late Closure (Frazier, 1978; Frazier & Rayner, 1982).
itself (theextremely delicious, brightgreenbroccoli). In a wh-movement construction, the filler cannot be
Finally,thelastresortstrategypredictsthatthepars- assignedaroleinthesentence’smeaninguntilitisrelat-
erwillpositatraceofashiftednounphraseonlywhen ed to a particular verb (or other phrasal head) as a
thegrammarforcestheheavyNPshiftanalysis.Crucial- theme, goal, instrument, etc.Incremental interpretation
lyforourexperiments,thepointatwhichthegrammar ismaximized by locating the underlying position of the
forcesthisanalysisvariesbasedonthesubcategorization moved constituent as quickly as possible. In a heavy
possibilitiesassociatedwiththeverb.Iftheverbisoblig- NPshiftconstruction,ontheotherhand,theinterpreta-
atorily transitive, the heavy NP shift analysis is forced tion of the verb phrase is speeded if the parser avoids
whenthereaderorlistenerencountersthematerialinter- positing a trace of a moved direct object. In (1) above,
veningbetweentheverbandthedirectobject.Forexam- adopting an intransitive frame upon encountering with
ple,iftheverbdevourweresubstitutedforeatin(1),the a fork allows the reader or listener to interpret this
heavy NP shift analysis would be required upon reach- phraseasmodifyingeat.Bycontrast,adoptingtheheavy
ing the prepositional phrase with a fork, since there is NP shift analysis requires the comprehender to delay
no grammatical continuation of Lucy devoured with a semantic interpretation, because on this analysis with a
fork other than a heavy NP shift construction. But if fork is modifying the eating of something specific, but
the verb is optionally transitive, like eat, non-heavy itisnotyet clear what.
392 A.Staubetal./JournalofMemoryandLanguage54(2006)389–406
Before presenting the experiments in detail, we Experiment 1
note the relevance of the issue we are investigating
to a proposal by Wasow (1997). Wasow performed Experiment1wasdesignedtotestthepredictionsof
several corpus studies to explore the circumstances the last resort strategy for the processing of heavy NP
under which speakers and writers produce sentences shiftagainstthepredictionsofthefirstresortandlexical
with a shifted word order. Specifically, he investigated expectation strategies. We had participants read heavy
whether speakers and writers produce the heavy NP NPshiftsentencesinwhichwevariedthesubcategoriza-
shift word order for their own benefit, or whether tion possibilities associated with the verb, and we
they take into consideration the needs of the listener recorded their eye movements while they read. Eye
or reader. He assumed that the heavy NP shift word movements are arguably the best indicator of moment-
order is especially difficult for the comprehender to-moment processing during reading (Rayner, 1998).
when the verb allows an intransitive frame, since in We presentedsentences like the following:
these circumstances the parser could be temporarily
misled into adopting this frame. If language produc- (6a) Jack praised from the stands his daughter’s
ers take this fact into consideration, heavy NP shift attemptto shoota basket.
constructions should appear more frequently with (6b) Jack watched from the stands his daughter’s
obligatorily transitive verbs than with optionally tran- attemptto shoota basket.
sitive verbs. He found, however, that in both printed
materials and spoken language heavy NP shift is in In(6a),themainverb(praised)requiresadirect object.
fact more common with optionally transitive verbs Consequently,thereadermustadopttheheavyNPshift
than with obligatorily transitive verbs. Wasow analysis to attach the immediately post-verbal preposi-
suggested that by using the heavy NP shift word tional phrase (from the stands). In (6b), on the other
order with optionally transitive verbs, producers are hand, the verb (watched) is strongly transitive-biased,
leaving their options open, since they can decide butallowsanintransitiveframe;thesentencecouldcon-
relatively late in the sentence whether or not to tinuewith,forexample,anadjunctclauseafterthepost-
include a direct object. verbalprepositionalphrase,e.g.,Jackwatchedfromthe
OurhypothesisthatheavyNPshiftistheparser’slast standsashisdaughtertriedtoshootabasket.In(6b),itis
resortisquiteconsistentwithWasow’sassumptionthat theshifteddirectobjectnounphraseitself(hisdaughter’s
listeners can be misled into adopting an intransitive attempttoshootabasket)thatalertsthereaderthatthe
frame.Infact,thelastresortstrategypredictsthatlisten- sentence requiresa heavyNP shiftanalysis.
erswillalwaysadoptanintransitiveframe,whenoneis The last resort strategy predicts that readers will, in
available.Wealsothinkthat,asWasowimplies,revising fact,adopttheheavyNPshiftanalysisatthesetwodif-
this initial analysis in favor of the heavy NP shift anal- ferent points in the two sentences. Consistent with the
ysisis likelyto induce processing difficulty. assumptiondiscussedabovethatconstructingthisanal-
Unlike Wasow, however, we suspected that the ysis is likely to induce processing difficulty, we made
heavy NP shift word order may also induce processing two specific predictions regarding the pattern of data.
difficulty when the verb is obligatorily transitive. First, we expected to see longer reading times and/or
When material other than a direct object arrives after more regressive eye movements on the prepositional
an obligatorily transitive verb (e.g., Lucy devoured with phrase in (6a) than in (6b). Second, we expected to
a fork), the parser has no choice but to construct the see shorter reading times and/or fewer regressions on
heavy NP shift analysis. However, the fact that the the shifted noun phrase in (6a) than in (6b). In short,
verb is obligatorily transitive may cause readers or lis- we predicted a processing tradeoff, with difficulty aris-
teners initially to attempt to attach the post-verbal ingineachsentenceatthepointatwhichtheprocessor
material as the direct object, and to have to reanalyze. hasnooptionbuttoconstructtheheavyNPshiftanal-
This is especially likely to be the case if the parser has ysis. This point is relatively early when the verb is
a predictive component, enabling it to build the struc- obligatorilytransitive,andrelativelylatewhenthe verb
ture corresponding to the direct object upon process- is optionally transitive.
ing the verb (e.g., Frazier & Fodor, 1978; Kimball, The first resort and lexical expectation strategies do
1973). We hypothesized, in short, that processing not predict differences between (6a) and (6b). The first
heavy NP shift sentences is likely to be difficult resortstrategypredictsthattheheavyNPshiftanalysis
whether the verb is obligatorily or optionally transi- istheparser’sinitialanalysis,regardlessofverbsubcat-
tive; we designed our experiments to test the idea that egorization possibilities. Therefore, it predicts that this
this difficulty would appear on the material that inter- analysis should induce difficulty in neither (6a) nor
venes between verb and direct object when the verb is (6b).As notedabove,weregarded thefirst resortstrat-
obligatorily transitive, and on the shifted noun phrase egy as implausible, and considered the lexical expecta-
when the verb is optionally transitive. tion strategy to be the serious competitor to the last
A.Staubetal./JournalofMemoryandLanguage54(2006)389–406 393
resortstrategy.Thelexicalexpectationstrategypredicts praised, he or she would see the version of the other
thattheparserwillconstructtheheavyNPshiftanalysis praised/watched stimulus set that included the verb
prior to reaching the shifted direct object in both (6a) watched.
and (6b). In (6a), this analysis is required by the gram- The 12optionallytransitiveverbsthatweusedwere
mar. In (6b), the verb (watched) is transitive-biased, alltransitive-biased.These12verbsappearedinboththe
and according to the lexical expectation strategy, this Connine, Ferreira, Jones, Clifton, and Frazier (1984)
willinducetheparsertoadopttheheavyNPshiftanal- production norms and the Gahl, Jurafsky, and Roland
ysisontheprepositionalphrase.Todistinguishthelex- (2004)corpus-basednorms.Wecalculatedthetransitiv-
icalexpectationandlastresortstrategies,wemadesure itybiasofeachverbasthemeanofthescoresobtained
that all of our optionally transitive verbs were in fact from these two sources, using scoring categories that
transitive-biased. were common to both sets of norms. Note, however,
that while all the verbs had high transitivity ratings
Method accordingtobothsetsofnorms,thecorrelationbetween
the scores obtained from these two sources was very
Participants close to zero (r=.01), presumably because of the
Thirty-sixnativespeakersofAmericanEnglish,who restricted range of transitivity ratings for the verbs we
were students at the University of Massachusetts, were used. Table 1 provides the transitivity biases obtained
given course credit or were paid $5 to participate in from each set of norms, as the percentage of the total
the experiment. All had normal or corrected-to-normal occurrences of each verb in which the verb was used in
vision, and all were na¨ıve to the purpose of the a transitive frame, as well as the means of these scores.
experiment. Themean transitivitybiases of the optionally transitive
verbsrangedfrom.66(worship)to.93(visit),withover-
Materials allmeanof .80 andSD=.07.
Weconstructed24setsofsentenceslike(6a)and(6b). Of the 12 obligatorily transitive verbs that we used,
Ineachsentence,aprepositionalphraseappearedimme- threeappearedinneithertheConnineetal.northeGahl
diately after the verb, followed by a direct object noun etal.norms,andtwoothersappearedonlyintheCon-
phrase. In all but one item, the prepositional phrase nine et al. norms. To obtain transitivity bias ratings
was a locative or temporal adjunct of the verb. In the from a common source for all 12 verbs, we conducted
remaining item, the prepositional phrase was an argu- a production norming study with 16 participants. For
ment of the verb in both sentences. The noun phrase each verb, participants were asked to write a sentence
wasmaderelativelylong(or‘‘heavy’’)byincludingarel- containingtheverb,usingthe‘‘firstsentencethatcomes
ativeclause,aprepositionalphrasemodifier,oranargu- to mind.’’ The verbs of interest were intermixed with
mentoftheheadnoun.Theonlydifferencebetweenthe several optionally transitive and intransitive verbs. We
two versions was the choice of verb. In the (a) version, excluded five responses in which the participant used
the verb was obligatorily transitive; we (and several the critical word as a noun, rather than a verb, leaving
independent raters) judged that a direct object had to atotalof187responses.Ofthese,sixresponsesfellinto
appear after the prepositional phrase in order for the categories that were clearly not relevant to the experi-
sentencetobegrammatical.Werefertothisastheoblig- ment.Fourusedtheverbbotherina‘‘negativepolarity’’
atorycondition.Inthe(b)version,theverballowedfora construction (e.g., Don’t bother to clean it up.); this
range of possible continuations. We refer to the this as
theoptionalcondition.Thefullsetofmaterialsisprovid-
edin Appendix A. Table1
The24setsofsentencesused12differentverbpairs, Experiment1:Transitivitybiasesofoptionallytransitiveverbs
with each verb pair appearing in two sets of sentences.
Verb Connineetal.(1984) Gahletal.(2004) Overall
Eachverbpairwasmatchedforlength,andtheobliga-
worship .72 .60 .66
torily and optionally transitive verbs did not differ sig-
govern .79 .75 .77
nificantly in frequency, based on the Francis and
attack .77 .94 .86
Kucˇera (1982) norms. The obligatorily transitive verbs
watch .84 .73 .79
had a mean log frequency of 3.93, SD=1.27, and the
n leave .82 .78 .80
optionally transitive verbs had a mean logn frequency visit .96 .91 .93
of 4.81, SD=1.15. Two lists were constructed so that follow .94 .66 .80
each participant saw one sentence from each pair, and push .94 .64 .79
12 sentences overall in each of the two conditions. The paint .89 .89 .89
lists of stimuli were arranged so that each participant hear .64 .78 .71
saweachverbonlyonce;i.e.,ifaparticipantsawthever- teach .83 .79 .81
clean .80 .89 .84
sion of (6) that included the obligatorily transitive verb
394 A.Staubetal./JournalofMemoryandLanguage54(2006)389–406
constructioncanberuledoutbythereaderofourexper- question.Averageaccuracyforthecomprehensionques-
imentalsentencesbythetimeheorshehasreachedthe tionswasabove85%,withnoparticipantscoringbelow
verb,sincetheverbhasnotbeenprecededbyalicensing 75%. The entire experiment lasted approximately
negativeelement. Intwoothercases,the verbwasused 30min.
in a common idiom (e.g., Carry on.). A reader of our
experimental sentences can rule out such an idiomatic Results anddiscussion
use by the time the post-verbal prepositional phrase is
reached. The verb had a direct object noun phrase in We analyzed three regions in each sentence. The
179ofthe remaining 181responses (99%). prepositionalphraseregionconsistedoftheprepositional
We note that exactly half of the experimental items phrase that appeared immediately following the verb
had a definite description (e.g., the sailors, the math (from the stands in 6a and 6b). As discussed above, we
teacher), rather than a name, in subject position. For predicted elevated reading times and/or an increase in
these 12 items, it was possible in principle for partici- regressionsontheprepositionalphraseintheobligatory
pants to adopt an initial analysis according to which conditioncomparedtotheoptionalcondition.Thenext
the verb marked the beginning of a reduced relative region, the noun phrase region, consisted of the direct
clause (though in many cases this analysis would have object noun phrase up through the head noun (his
been very implausible). We considered it quite unlikely daughter’sattempt).Thisregionwasthreeorfourwords
that participants would actually adopt this analysis, inlength,withtheexceptionofoneiteminwhichitwas
since the reduced relative analysis is strongly dispre- five words. We predicted elevated reading times and/or
ferred,in mostcontexts(e.g., Binder,Duffy,&Rayner, an increase in regressions on the noun phrase in the
2001;Ferreira&Clifton,1986;cf.Spivey&Tanenhaus, optional condition compared to the obligatory condi-
1998).Inanyevent,statisticalanalysesshowednosignif- tion.Finally,weanalyzedaspilloverregionthatconsist-
icant differences between items that did, and did not, edof the remainder ofthe sentence (toshoot abasket).
temporarily allow a reduced relative analysis, so we do Fourreadingtimemeasureswerecomputed:firstfix-
not distinguish between these two groups of items in ationduration,firstpasstime(whichisreferredtoasgaze
reportingourresults. duration when discussing single-word regions), go-past
Theexperimentalsentenceswereintermixedwith118 time, and percent regressions (Rayner, 1998). The first
fillersentencesofvarioustypes,noneofwhichusedthe two measures reflect early stages of processing such as
heavyNPshiftwordorder.Thesentenceswerepresent- lexical access (Reichle, Rayner, & Pollatsek, 2003), but
ed in an individually randomized order to each parsing difficulty has also been shown to affect these
participant. measures (e.g., Frazier & Rayner, 1982). First fixation
duration is simply the duration of the first fixation in a
Procedure region, whether it is the only fixation in the region or
Participantsweretestedindividually.Eyemovements the first of multiple fixations. Though this measure is
were recorded using a Fourward Technologies Dual typically used with single-word regions, we computed
PurkinjeGeneration6eyetracker,whichhasanangular it in this experiment because we anticipated that effects
resolutionoflessthan10minofarc.Theeyetrackerwas of verb type might appear as early as the first fixation
interfaced with an IBM compatible computer. All sen- on the prepositional phrase and/or the shifted noun
tencesinthisexperimentweredisplayedonasingleline, phrase. First fixation duration on the spillover region
withamaximumlengthof80characters.Whileviewing is unlikely to be informative, but we report it in the
wasbinocular,onlytherighteyewasmonitored.Stimuli interest of completeness. First pass time is the sum of
were displayed on a 15-in. NEC MultiSync 4FG moni- all fixations in a region prior to leaving the region for
tor. Participants were seated 61cm from the computer thefirsttime,eithertotheleftortheright.Go-pasttime
screen; at this distance, 3.8 characters subtended 1(cid:2) of istheelapsedtimefromfirstfixatingtheregionuntilthe
visualangle. readerleavestheregiontotheright,includinganytime
Onarrivalatthelaboratory,participantsweregiven spenttotheleftoftheregionafteraregressiveeyemove-
instructions and had a bite bar prepared for them that ment and any time spent re-reading material in the
served to stabilize the head. A calibration routine was region before moving on. (Go-past time is also some-
performed,anditsaccuracywascheckedaftereachsen- times called regression path duration.) Finally, the per-
tence.Participantswereinstructedtoreadthesentences cent regressions measure gives the probability that a
for understanding, and to read at a normal rate. After reader makes a regressive eye movement after fixating
readingeachsentence,theparticipantspressedabutton theregion.Thismeasureincludesonlyregressionsmade
toremovethesentence.Thefirsteighttrialsoftheexper- duringthereader’sfirstpassthroughtheregion;itdoes
imentalsessionwerepracticetrials.Comprehensionwas notincluderegressionsmadeafterre-fixatingtheregion.
checked on approximately 30% of all trials during the Effectsofsyntacticmisanalysisareoftenapparentinthe
experiment by presenting the participant with a yes/no go-past and regression measures. We did not compute
A.Staubetal./JournalofMemoryandLanguage54(2006)389–406 395
late measures such as second pass time and total time these ANOVAs rejected the null hypothesis at the .05
because we were interested primarily in the processing level. These F statistics, as well as MinF0, are reported
costsincurredatthefirstencounterwiththepreposition- inTable 3.We discussthe regionsindividually below.
al phraseandthe subsequentnounphrase.
Priortoallanalyses,sentenceswithtracklosseswere Prepositionalphraseregion
excluded (less than 2% of trials). In addition, fixations The first fixation on this region was longer in the
less than 80ms in duration, and within one character obligatory condition than in the optional condition
ofthepreviousorsubsequentfixation,wereincorporat- (280ms vs. 269ms), but this difference was significant
ed into this neighboring fixation. The same procedure onlyintheparticipantsanalysis.Firstpasstime,howev-
was used to incorporate fixations less than 40ms in er, was significantly longer in the obligatory condition
duration and within three characters of the previous or by both participants and items (646ms vs. 608ms).
subsequent fixation. Remaining fixations of less than Go-pasttimewasalsosignificantlylongerintheobliga-
80ms were deleted, as were fixations of longer than tory condition (767ms vs. 668ms). Finally, there were
800ms. It is thought that readers do not extract useful more regressions in the obligatory condition (14.3% vs.
informationfromfixationsshorterthan80ms(seeRay- 7.5%). In sum, differences on the first pass, go-past,
ner & Pollatsek, 1989), and that fixations longer than andpercentregressionsmeasuresallsupportedourpre-
about800msarelikelytoreflecttracklosses.Lessthan diction of greater processing difficulty on the preposi-
2% of allfixationswere eliminated. tional phrase when the verb was obligatorily transitive
Table2presentstheparticipantmeansoneachmea- thanwhenitwasmerely transitive-biased.
sureforeachoftheregionsweanalyzed,aswellas95%
confidence intervals for the differences between condi- Nounphraseregion
tionmeans(Loftus&Masson,1994;Masson&Loftus, Therewasnohintofafirstfixationdifferenceonthe
2003). For each test of differences between means, we nounphraseregion.Firstpasstime,however,wassignif-
performedanalysesofvariance(ANOVAs)withpartic- icantly longer in the optional condition than in the
ipants (F ) and items (F ) as random effects variables, obligatorycondition(735msvs.693ms).Onthego-past
1 2
and with counterbalancing group as a between-partici- measure,thenumericaltrendwentintheoppositedirec-
pants or between-items factor (Pollatsek & Well, tion, but the difference between conditions was not sig-
1995). We report a difference as significant if both of nificant (ps>.15). Finally, there was a significant
Table2
Experiment1:ParticipantMeanReadingTimes,inMilliseconds,andPercentRegressions
Measure Prepositionalphrase Nounphrase Spillover
(fromthestands) (hisdaughter’sattempt) (toshootabasket)
Firstfixationduration
Obligatory 280 269 259
Optional 269 270 258
95%CIofdifference 1.0to21.6 (cid:1)13.4to13.0 (cid:1)10.0to11.4
Firstpasstime
Obligatory 646 693 795
Optional 608 735 794
95%CIofdifference 13.6to61.4* (cid:1)75.5to(cid:1)9.7* (cid:1)38.8to40.2
Go-pasttime
Obligatory 767 845 921
Optional 668 809 936
95%CIofdifference 61.6to131.6* (cid:1)16.6to89.2 (cid:1)87.3to55.7
Percentregressions
Obligatory 14.3 11.4 9.2
Optional 7.5 6.5 8.2
95%CIofdifference 2.6to11.0* 1.2to8.6* (cid:1)3.2to5.2
Obligatory=Jack praised from the stands his daughter’s attempt to shoot a basket. Optional=Jack watched from the stands his
daughter’sattempttoshootabasket.
Note.Reportedherearethe95%confidenceintervalsforthedifferencesbetweenconditions,calculatedusingtheprocedureforwithin-
participantconfidenceintervalsdescribedbyLoftusandMasson(1994;MassonandLoftus,2003).Differencesthatweresignificantby
bothparticipantsanditems(p<.05)aremarkedwithanasterisk(*).
396 A.Staubetal./JournalofMemoryandLanguage54(2006)389–406
Table3
Experiment1:Analysisofvarianceresultsbyregionforeffectofcondition
Measure Prepositionalphrase Nounphrase Spillover
Firstfixationduration F =4.86,p<.05 Fs<1.5 Fs<1.5
1
F =2.17,p=.16
2
MinF0(1,40)=1.50,p=.23
Firstpasstime F =10.14,p<.01 F =6.93,p<.02 Fs<1.5
1 1
F =4.21,p=.05 F =4.99,p<.05
2 2
MinF0(1,39)=2.97,p=.09 MinF0(1,48)=2.90,p=.10
Go-pasttime F =32.80,p<.01 F =1.94,p=.17 Fs<1.5
1 1
F =12.33,p<.01 F =2.18,p=.15
2 2
MinF0(1,38)=8.96,p<.01 MinF0(1,54)=1.03,p=.32
Percentregressions F =10.54,p<.01 F =6.97,p<.02 Fs<1.5
1 1
F =7.74,p<.02 F =7.37,p<.02
2 2
MinF0(1,49)=4.46p<.05 MinF0(1,54)=3.58,p=.06
Note.Forallanalysesbyparticipants(F ),df=(1,34);forallanalysesbyitems(F),df=(1,22).
1 2
differencebetweenconditionsonthepercentregressions passresultssupportedourpredictionofprocessingdiffi-
measure,andthisresultedfrommoreregressionsinthe cultyontheshiftednounphraseitselfwhentheverbwas
obligatory condition than in the optional condition optionallytransitivecomparedtowhenitwasobligato-
(11.4%vs. 6.5%). rilytransitive.Toperformadirecttestofourprediction
The first pass results supported our hypothesis that of a crossover pattern, with the heavy NP shift word
theshiftednounphrasewouldinducerelativeprocessing order causing difficulty on the prepositional phrase in
difficultywhentheverbwasoptionallytransitive.How- the obligatory condition and difficulty on the noun
ever,thefactthatthereweremoreregressionsfromthe phrase in the optional condition, we performed a 2·2
noun phrase region in the obligatory condition than in ANOVA on first pass time, with verb type and region
theoptionalcondition(leadingtoanumericallygreater asfactors.Themaineffectofverbtypedidnotapproach
go-past time in the obligatory condition) was not pre- significance,buttheinteractionofverbtypeandregion
dicted. We think that the overall results support the was highly significant: F (1,34)=17.81, p<.01;
1
notionthatthenounphraseregionwaseasiertoprocess F (1,22)=8.41, p<.01; MinF0 (1,41)=5.71, p<.05.
2
whentheverbwasobligatorilytransitive, sincetheper- We reserve further discussion of these results until we
cent regressions data reflect a very small proportion of have presentedthe results ofExperiment 2.
trials,andthefirstpassdata,bycontrast,reflectprocess-
ing on every trial. (For discussion of the interpretation
of the first pass and regression measures, see Altmann, Experiment 2
1994; Altmann, Garnham, & Dennis, 1992; Rayner &
Sereno, 1994a, 1994b.) Nevertheless, one of our goals Experiment2hadseveralpurposes.First,wewanted
for Experiment 2 was to clarify the pattern of results todeterminewhethertheresultsofExperiment1would
on the noun phrase region. To anticipate the results generalizetoheavyNPshiftsentencesinwhichdifferent
from that experiment, it appears that the difference in material intervened between the verb and the shifted
firstpasstimeisreplicable,butthedifferenceinpercent directobject.InExperiment2,weusedanadverbrather
regressionsisnot. thanaprepositionalphraseasthisinterveningmaterial.
Second, we wanted to compare the processing of
Spillover region heavyNPshiftsentenceswithobligatorilyandoptional-
There were no differences that approached signifi- ly transitive verbs with the processing of unshifted sen-
cance on this region, on any measure. Evidently, all tences with the same verbs. We have interpreted the
effectsoftheexperimentalmanipulationwerecomplete- evidence from Experiment 1 of relative processing diffi-
lyresolvedbythetimereaders’eyesmovedpastthehead cultyontheprepositionalphrasewhentheverbisoblig-
nounofthe shifted nounphrase. atorily transitive, and relative processing difficulty on
In sum, the results of Experiment 1 indicated quite theshiftednounphrasewhentheverbisoptionallytran-
clearly that a post-verbal prepositional phrase causes sitive, as confirming our proposal that the parser con-
processingdifficultywhentheverbisobligatorilytransi- structs the heavy NP shift analysis at one of the these
tive,comparedtowhentheverbistransitive-biasedbut points, depending on verb type. Our account predicts,
canoccurinanintransitiveframe.Inaddition,the first therefore,thatwithanunshiftedwordorder(i.e., when
A.Staubetal./JournalofMemoryandLanguage54(2006)389–406 397
thedirectobjectappearsimmediatelyaftertheverb),the processing the adverb should also be more difficult in
effectswe haveobserved should disappear. (7c) than in (7a), in which the adverb comes before the
Third, we suspected that the percent regressions dif- verb, and there should be no difference between these
ference on the noun phrase region may not have been sentencesonthenounphraseregion.Ontheotherhand,
a reliable effect, primarily because it contradicted the processingthedirectobjectnounphraseshouldbemore
firstpassresults.WethoughtthatareplicationofExper- difficultin(7d)thanin(7b),withnodifference between
iment 1 might confirm the first pass effect on the noun theseconditionson theadverb.
phraseregionwithoutreplicatingthepercentregressions
effect. Method
Finally, we were interested in whether transitivity
preferences play a major role in determining when the Participants
parser adopts the heavy NP shift analysis. In Experi- Twenty-four native speakers of American English,
ment1,alloftheoptionallytransitiveverbsthatweused who were students at the University of Massachusetts,
had a high transitive bias. In Experiment 2, we used were given course credit or were paid $5 to participate
optionally transitive verbs whose biases ranged more in the experiment. All had normal or corrected-to-nor-
widely,andhadonlyamoderatetransitivebiasoverall. mal vision, and all were na¨ıve to the purpose of the
This would enable us to investigate, in a different way, experiment.
the prediction of the lexical expectation strategy that
thelikelihoodoftheparseradoptingtheheavyNPshift Materials
analysis before reaching the shifted noun phrase is cor- Weconstructed28setsofsentencessimilarto(7a–d)
related withthe verb’s transitivity bias. above, in which verb type and adverb position were
In Experiment 2, we tested sentences like the manipulated.Inthe(a)and(c)versionsofeachsentence,
following: we(andseveralotherraters)judgedadirectobjectnoun
phrasetoberequiredafterthemainverbinorderforthe
(7a) The teacher immediately corrected the unusual sentencetobegrammatical.Inthe(b)and(d)versions,
answer thestudent hadgiven. ontheotherhand,themainverballowedthesentenceto
(7b) The teacher immediately applauded the unusual continue either with or without a direct object. For
answer thestudent hadgiven. example, it is possible to construct a version of (7d) in
(7c) The teacher corrected immediately the unusual which the adverb is followed by an adjunct clause: The
answer thestudent hadgiven. teacher applauded immediately when the student gave
(7d) The teacher applauded immediately the unusual herunusualanswer.Thefullsetofmaterialsisprovided
answer thestudent hadgiven. inAppendix B.
The28setsofsentencesused14differentverbpairs,
In(7a)and(7c),theverbisobligatorilytransitive,while with each verb pair appearing in two sets of sentences.
in(7b)and(7d),theverbisoptionallytransitive.In(7a) All but one of the verb pairs was matched for length,
and (7b), the adverb (immediately) appears before the with the verbs in the remaining pair differing in length
verb, while in (7c) and (7d), the adverb appears after byonecharacter.Theobligatorilyandoptionallytransi-
theverb,withtheverb’sdirectobject(theunusualanswer tiveverbsdidnotdifferinfrequency,basedontheFran-
the student had given) shifted to the right. Accordingly, cisandKucˇera(1982)norms.Theobligatorilytransitive
we refer to (7a–d) respectively, as the adverb first oblig- verbshadameanlog frequencyof3.84,SD=1.50,and
n
atorily transitive condition, the adverb first optionally theoptionallytransitiveverbshadameanlog frequen-
n
transitivecondition,theadverbsecondobligatorilytransi- cy of 4.13, SD=1.65. Four lists were constructed so
tiveconditionandtheadverbsecondoptionallytransitive that each participant saw one sentence from each set,
condition. and seven sentences overall in each of the four condi-
Sentences (7c) and (7d) are analogous to the obliga- tions.Inaddition,asinExperiment1thelistsofstimuli
tory and optional conditions in Experiment 1, and we were arranged so that each participant saw each verb
predicted the same pattern of results: relative difficulty onlyonce;i.e.,ifaparticipantsawaversionof(7)that
on the material intervening between verb and noun included the verb corrected, he or she would see a ver-
phrasein(7c),andrelativedifficultyontheshiftednoun sion of the other corrected/applauded stimulus set that
phrasein(7d).Wedidnotexpectthepercentregressions includedthe verb applauded.
effectonthenounphraseregionthatappearedinExper- The14optionallytransitiveverbsthatweusedvaried
iment 1 to replicate in this experiment; essentially, we widelyintheirtransitivitybiases.AsinExperiment1,we
suspected that this finding was simply a Type I error. obtainedtransitivitypreferencesfortheseverbsfromthe
Ifitisinfactthecasethattheappearanceoftheadverb Connine et al. (1984) and Gahl et al. (2004) norms.
after the verb in (7c) causes disruption, and causes the Twelve of the verbs were included in the Connine
parser to construct the heavy NP shift analysis, then et al. norms, and 13 were included in the Gahl et al.
398 A.Staubetal./JournalofMemoryandLanguage54(2006)389–406
norms.Wecalculatedthetransitivitybiasofthe12verbs Of the 14 obligatorily transitive verbs used in this
thatappearedinbothsourcesasthemeanofthescores experiment, seven were repeated from Experiment 1.
from these two sources; the correlation between the Because four of the remaining seven verbs appeared in
scores from the two sources was r=.44. For report we neither the Connine et al. nor the Gahl et al. norms,
used only the Gahl et al. norms, since this verb did we had 16 participants construct sentences using these
not appear in the Connine et al. norms. For applaud, sevenverbs,whichwereintermixedwithoptionallytran-
whichappearedinneithersetofnorms(andwhichhap- sitiveandintransitiveverbs.Oneresponsewasexcluded
pens to be the optionally transitive verb in our sample because it used the critical word as a noun, leaving a
setofsentencesabove),weconductedaproductionnor- total of 111 responses. Of these, three responses were
ming study with 10 participants, and determined the clearlyidiomatic(e.g.,Buylow,sellhigh),andfivemade
transitivity bias of this verb to be .60. Table 4 provides use of a particle verb construction (e.g., Joe blocked off
the transitivity biases obtained from each set of norms, the driveway). The latter are clearly not relevant to the
aswellasthemeansofthesescores.Overall,thetransi- present experiment, since the particle in such construc-
tivity biases of the optionally transitive verbs ranged tions can only be separated from the verb by a direct
from .04 (try) to .82 (watch), with a mean of .50 and object noun phrase. Of the remaining 103 responses,
SD=.22. 102 included a noun phrase direct object for the verb
Theoptionallytransitiveverbsthatweusedvarywith (99%).
respect to whether they frequently occur with an argu- The experimental sentences were intermixed with 80
ment other than a noun phrase. Some of these verbs fillersentencesofvarioustypes,noneofwhichinvolved
(e.g.,cheat,walk)frequentlyoccuraspureintransitives, theheavyNPshiftwordorder.Thesentenceswerepre-
i.e., with no internal argument at all. Others, however, sented in an individually randomized order to each
dotendtotakeanargumentotherthananounphrase. participant.
Forexample,trytendstooccurwithanon-finiteclausal
complement(e.g.,trytofixthecar),andreporttendsto Procedure
occur witha finiteclausal complement (e.g., report that The procedure was identical to Experiment 1. Aver-
the battle was over). These differences are glossed over ageaccuracyonthecomprehensionquestionswasabove
in computing overall transitivity preference, both in 90%,withnoparticipantscoringbelow80%.Theentire
ourownexperimentand,frequently,inothersinwhich experiment lastedapproximately 30min.
verbtransitivitybiasisstudied(seeGahletal.,2004,for
discussionofrelatedissues).Weacknowledgethatthese Results anddiscussion
differences may be relevant to the comprehension of
heavy NP shift; Stallings et al. (1998) present evidence Weanalyzedthreeregionsineachsentencethatcor-
thattheyplayaroleintheproductionofheavyNPshift. responded to the three regions we analyzed in Experi-
Webrieflyrevisit thisissue in the Generaldiscussion. ment 1. The adverb region consisted only of the adverb
(immediately), which appeared before the verb in the
adverbfirstconditions,andaftertheverbintheadverb
Table4 second conditions. The next region, the noun phrase
Experiment2:Transitivitybiasesofoptionallytransitiveverbs
region, consisted of the first three words of the direct
Verb Connineetal.(1984) Gahletal.(2004) Overall object noun phrase, which included a determiner, an
worship .72 .60 .66 adjective,andtheheadnoun(theunusualanswer).Final-
perform .32 .76 .54 ly, we analyzed a spillover region that consisted of the
report — .34 .34 nexttwo wordsof the sentence (the student).
follow .94 .66 .80 We computed the same four reading time measures
pull .16 .57 .37 asinExperiment1:firstfixationduration,firstpasstime,
watch .90 .73 .82 go-pasttime,andpercentregressions.Wepointoutthat
try .03 .05 .04
asinExperiment1,thefirstfixationdurationmeasureis
applaud — — —
meaningful only for the first two of the three regions,
study .32 .86 .59
where the beginning of the region marked a phrasal
cheat .22 .55 .39
boundary. Less than 2% of all trials were excluded due
lecture .28 .38 .33
protest .62 .25 .44 totracklosses.WeusedthesameprocedureasinExper-
help .80 .80 .80 iment1tocombineveryshortfixationswithneighboring
debate .20 .49 .35 fixations. Less than 1.5% of remaining fixations were
eliminated due tofalling outsidethe 80–800ms range.
Note. The transitivitybias of applaud, whichappeared in nei-
ther set of norms, was assessed by conducting a single-item Foreachmeasureoneachregion,wefirstperformed
productionnormingsessionwithtenparticipants.Weobtained ANOVAs by participants and by items, with condition
atransitivitybiasof.60. assinglewithin-participantsorwithin-itemsfactorwith