Effects of Processing Difficulty on Judgments of
Acceptability∗
Gisbert Fanselow & Stefan Frisch
1 Introduction and Overview
There is a certain tension between the role which acceptability judgements play in linguistics
and the level of their scientific underpinning. Judgements of grammaticality form the
empirical basis of generative syntax, but little is known about the processes underlying their
formation and the nature of factors different from grammar contributing to them.
This paper illuminates the impact of processing difficulty on acceptability. Section 2
reviews evidence showing that parsing problems often reduce acceptability. That processing
difficulty may increase acceptability is less obvious, but this possibility is nevertheless borne
out, as Section 3 shows, which reports the results of several experiments dealing with locally
ambiguous sentences involving discontinuous NPs, NP-coordination, and VP-preposing. The
preferred interpretation of a locally ambiguous construction can have a positive influence on
the global acceptability of a sentence even when this reading is later abandoned. If the initial
analysis of a locally ambiguous construction would, however, render the sentence ungrammatical, the local ambiguity decreases acceptability, as our experiment focusing on long whmovement in Section 4 shows . The global acceptability of a sentence thus turns out to be
influenced by local acceptability perceptions during the parsing process.
∗
We want to thank Caroline Féry, Heiner Drenhaus, Matthias Schlesewsky, and Ralf Vogel for helpful
discussions, and Jutta Boethke, Jörg Didakowski, Ewa Trutkowski, Julia Vogel, Nikolaus Werner, Nora Winter,
and Katrin Wrede for technical support. The research reported here was supported by DFG-grant FOR 375
2 Decreased acceptability caused by processing problems
From its very beginning, generative syntax has subscribed to two fundamental convictions:
the notions of grammaticality and acceptability have to be kept strictly apart, and grammatical
sentences may be unacceptable because of the processing difficulty they involve (Chomsky
1957, Chomsky & Miller 1963). There is a consensus that the latter situation is exemplified
by multiple center embeddings such as (1). Their acceptability decreases with the number of
self embeddings involved, yet they are constructed according to the principles of English
grammar, and should therefore be grammatical. Processing explanations for their low
acceptability seem well-motivated, since they fit into cognitive theories of language
processing (see Lewis 1993).
(1)
the man who the woman who the child who the mosquito bit loves kicked the horse.
Strong garden path sentences such as the horse raced past the barn fell (Bever 1970) illustrate
the same point: sentences not violating any of the constructional principles of English may
have properties that make it close to impossible for human parsing routines to identify their
correct grammatical analysis. This renders them unacceptable.
While it is fairly uncontroversial that effects of strong processing problems should not
be derived from grammatical principles, the interpretation of milder parsing difficulties is less
uniform. Consider, e.g., the fronting of objects in free word order languages like German.
Experimental studies have revealed that object initial sentences such as (2b) are less
acceptable than their subject initial counterparts (2a) (Bader & Meng 1999, Featherston to
appear, Keller 2000).
(2)
a. der
Tiger
the.nom tiger
b. den
Löwen
hat
den
Löwen
gejagt
has
the.acc
lion.acc
chased
hat
der
Tiger gejagt
the.acc
lion
has
the.nom
tiger
chased
Keller (2000) and G.Müller (1999) make syntactic constraints responsible for the lower
acceptability of object initial sentences, but there is a different explanation at hand: Object
initial sentences are more difficult to parse than subject initial ones, and this may render them
less acceptable. Processing difficulties of object-initial structures have been amply
documented since Krems (1984), see also Hemforth (1993), Meng (1998), Schlesewsky,
Fanselow, Kliegl & Krems (2000), among others. Their low acceptability can be explained in
terms of this processing problem if the latter does not merely reflect the grammatical
difficulty. Indeed, the grammar-independent nature of the additional processing load of object
initial sentences can be argued for.
Fanselow, Kliegl & Schlesewsky (1999) report a self paced reading study which
compared the processing of embedded German subject-initial, object-initial, and yes-no
questions. They found an increase in reading times for the object initial condition (3b),
beginning with the wh-phrase and ending at the position of the second noun phrase (=the
subject, in the object-initial condition).
(3)
es ist egal "it does not matter"
a. wer
vermutlich glücklicherweise den
who.nom presumably fortunately
b. wen
who.acc
c. ob
the.acc man
vermutlich glücklicherweise der
presumably fortunately
recognized
Mann erkannte
the.nom man
vermutlich glücklicherweise der Mann
whether presumably fortunately
Mann erkannte
recognized
den Dekan erkannte
the.nom man the.acc dean recognized
Fanselow et al. (1999) interpret this result in terms of memory cost: A fronted object whphrase must be stored in memory up to the point at which it can be integrated into the
structure (which is immediately after the subject in German). This account is in line with
findings from ERP research. King & Kutas (1995) found a sustained anterior negativity for
the processing of English object relative clauses (as compared to subject relative clauses),
which H. Müller, King & Kutas (1997) relate to memory. Felser, Clahsen & Münte (2003),
Fiebach, Schlesewsky & Friederici (2002), and Matzke, Mai, Nager, Rüsseler & Münte
(2002) found a sustained LAN in the processing of German object-initial wh-questions and
declaratives, which is again attributed to the memory load imposed by the preposed object
until it can be integrated into the structure. The claim that object initial structures involve a
processing difficulty is thus well supported. It is natural to make this processing difficulty
responsible for the reduced acceptability of sentences like (2b).
Subjacency violations constitute a second domain in which processing difficulty rather
than grammar reduces acceptability. Current grammatical models such as Minimalism
(Chomsky 1995) or Optimality Theoretic Syntax (Grimshaw 1997) cannot cope with
subjacency violations easily. The descriptive apparatus of Minimalism is too restrictive for
capturing most subjacency induced islands. Optimality Theory faces a different difficulty:
When structures such as (4) violate subjacency, there is normally no grammatical way of
formulating them. Their ungrammaticality is thus hard to reconstruct in OT, which claims that
structures are ungrammatical only if better competitors exist (see Fanselow & Féry 2002, and
the references cited there).
(4)
??what do you wonder who has bought
Înterestingly, processing contributes, however, to the low acceptability of some island types,
among them wh-islands. Kluender & Kutas (1993) argue that syntactic islands arise at “processing bottlenecks” when the processing demands of a long distance dependency at the
clause boundary add up on the processing demands of semantically complex words such as
who or whether. This processing problem is reflected in dramatically reduced acceptability.
Processing accounts of the wh-island condition furthermore allow us to understand satiation
(Snyder 2000) and training effects (Fanselow, Kliegl & Schlesewsky, in prep.) that are
characteristic of wh-island violations: repeated exposure facilitates the processing of
sentences such as (4), and renders them more acceptable since their relative acceptability is
linked to processability.
Processing difficulty reduces acceptability in further areas. S. Müller (2004)
documents the discussion of extraposition from NP. He shows that the low acceptability for
extrapositions from certain attachment sites must be explained in terms of attachment
preferences, but has been mistakenly taken to reflect ungrammaticality in the early generative
literature. Experimental evidence (Sam Featherston, p.c.) suggests that German unambiguous
subject relative clauses get a better acceptability rating than object relative clauses, but subject
relative clauses with a locally ambiguous relative pronoun are less acceptable than their
unambiguous counterparts, which also may be explained in terms of the processing
difficulties coming with locally ambigous relative pronouns (Frisch, Schlesewsky, Saddy &
Alpermann 2001, 2002).
3 Increased acceptability linked to processing problems
3.1 General remarks
Processing difficulty may make a sentence less acceptable than one would expect on the basis
of its grammaticality. In principle, the reverse might also exist: processing difficulty renders a
sentence with low grammaticality fairly acceptable. E.g., this could be the case when the
factor making the structure ungrammatical is difficult to detect. A number of empirical
observations can be interpreted in this direction. Marks (1965:7) shows that the position of a
grammatical violation is relevant for the degree of (un-)acceptability: Sentences such as boy
the hit the ball turn out to be less acceptable than the boy hit ball the. Meng & Bader (1999)
and Schlesewsky, Fanselow & Frisch (2003) (among others) found chance performance
(rather than clear rejection) in speeded acceptability rating tasks for ungrammatical transitive
sentences such as (5) containing illegitimate combinations of two nominative NPs.
Schlesewksky et al. explain such results with the assumption that the case marking of the
second NP tends to be overlooked in nominative-initial sentences.
(5)
*welcher
Gärtner sah der
Jäger
which.nom gardner saw the.nom hunter
The experiments reported here were carried out in order to systematically investigate whether
positive effects of processing difficulties exist. In particular, we wanted to find out whether
local ambiguities can have mitigating effects on ungrammaticality. We expected such a
mitigating influence on the basis of the following reasoning: The fact that parsing problems
may reduce acceptability suggests that the global acceptability of a sentence does not only
reflect properties of the final analysis of the string, but also the “local acceptability” of
intermediate steps. When the intermediate steps have “better” grammatical properties than the
final analysis of the string, an impact of the quality of intermediate steps on global
acceptability should lead one to expect that global acceptability is increased by the relatively
well-formed intermediate parsing steps. In particular, we studied discontinuous noun phrases
(Experiment 1), subject verb agreement (Experiment 2), VP-preposing (Experiment 3) and
long distance movement (Experiment 4).
3.2 Experiment 1: Discontinuous Noun Phrases
NPs can be serialized discontinuously in German, as illustrated in (6c). See Fanselow (1988),
Fanselow & Ćavar (2001), and Riemsdijk (1989) for different analyses of discontinuous NPs
(DNP), and Bader & Frazier (2004) for offline experiments involving DNP.
(6)
a.
er
liest
he
reads
[NP
viele
Bücher]
many books
b.
Viele Bücher liest er
many books reads he
c.
Bücher
liest
er
books
reads he
viele
many
German DNP are subject to two grammatical constraints on number (cf. Fanselow & Ćavar
2001): an agreement constraint, and a ban against singular count nouns appearing in the
construction. Apart from a few exceptional constellations, DNPs are grammatical only if the
two parts agree in number (the AGREEMENT constraint). While such a constraint holds for
DNPs in many languages, German is exceptional in disallowing singular DNPs for count
nouns, as the contrast between (7a) and (7b) illustrates (the SINGULARITY constraint). The
constraint derives from a general requirement that articles may be absent in German in partial
and complete noun phrases only if the NP is headed by a plural or mass noun. Some dialects
repair the ungrammaticality of (7b) by “regenerating” (Riemsdijk 1989) an article in the left
part of the NP as shown in (7c), in other dialects, there is simply no grammatical way of
expressing what (7b) tries to convey. For exceptions to these generalizations, see Fanselow &
Ćavar (2001) and van Hoof (1997).
(7)
a.
alte
Professoren
liebt
sie
keine
old.pl professors.pl loves she
no.pl
“she loves no old professors”
b.
c.
*alten Professor
liebt
sie
old.sg professor.sg
loves she
keinen
no.sg
einen alten
Professor
liebt
an
professor
loves she
old
sie
keinen
no
Many German nouns (such as Koffer “suitcase”) do not express grammatical number
unambiguously: they do not distinguish singular and plural morphologically for nominative
and accusative case. The left periphery of the DNP construction (8) is therefore (locally)
compatible with a plural interpretation, which is excluded when the singular determiner
keinen is processed. Up to this point, however, the phonetic string allows an analysis in which
the SINGULARITY ban is not violated. Introspection suggests that this local ambiguity of
number increases acceptability as compared to other singular DNP: (8) sounds much better
than (7b). Experiment 1 tested the hypothesis that local ambiguities of number increase the
acceptability of singular DNP in German.
(8) Koffer
hat
sie keinen
suitcase.ambiguous has she no.singular
Description of Experiment 1
Method
40 students of the University of Potsdam participated. They were paid for participation, or
received course credits. Participants rated 112 sentences in pseudorandomized order for
acceptability on a six point scale (1=very good, 6=very bad). There were 4 items per
condition. Each participant saw 40 experimental items, 74 unrelated and 8 related distractor
items.
Material
Experimental items had the form exemplified in (9). In a sentence with a pronominal subject
preceded by the verb and followed by an adverb, an object noun phrase was split such that the
left part (lp) preceded the verb, while the right part (rp) was clause final. The lp could consist
of a single noun (simple), or of a noun preceded by an adjective (like alten, old) agreeing with
the noun. One third of the items used DNPs constructed with nouns distinguishing number
morphologically (morphmark). The lp and rp appeared in singular(s) or plural(p) form.
(9)
a.
Professor
kennt
sie
leider
keinen
professor.sg
knows
she
unfortunately no.sg
b. Professoren
kennt sie
leider
keine
professor.pl
knows she
unfortunately no.pl
simple_s_s
simple_p_p
c. Alten Professor kennt sie leider keinen
complex_s_s
d. Alte Professoren kennt sie leider keine
complex_p_p
e. Professor kennt sie leider keine
simple_s_p
f. Professoren kennt sie leider keinen
simple_p_s
g. Alten Professor kennt sie leider keine
complex_s_p
h. Alte Professoren kennt sie leider keinen
complex_p_s
The other two thirds of the items followed the paradigm in (10), using nouns not
distinguishing number morphologically (morphamb). For half of these, the accusative form is
identical with the stem (Koffer, suitcase)(stem), the remaining nouns form accusative with an
affix (Pilot-en, pilot)(affix). (10) illustrates the experimental paradigm for morphamb stem
nouns. The items with morphamb affix nouns were constructed accordingly.
Simple lp of morphamb nouns are ambiguous for number. Therefore, the distinction between
items (9a,b) and (9e,f) could not be mirrored in (10). In order to balance the design, items
with complex rp (10g,h) using adjectives such as rot “red” were constructed that did not enter
the analysis.
(10)
a.
b.
Koffer
hatte
er
leider
suitcase.amb had
he
unfortunately no.sg
Koffer
er
leider
he
unfortunately no
hatte
suitcase.amb had
keinen
keine
simple_a_s
simple_a_p
c.
Roten Koffer hatte er leider keinen
complex_s_s
d.
Rote Koffer hatte er leider keine
complex_p_p
e.
Roten Koffer hatte er leider keine
complex_s_p
f.
Rote Koffer hatte er leider keinen
complex_p_s
g.
Koffer hatte er leider keinen roten
distractor
h.
Koffer hatte er leider keine roten
distractor
Results and Discussion
In the other experiments, we used a rating scale different from the one in Exp. 1. In order to
increase comparabilty of results, we will use transformed values for mean ratings in this
results section: The ratings on the “1=best/6=worst” scale are mapped to their equivalent on
the “1=worst/7=best” scale used later (transformed value = 8 – ( real value + (real value 1)/5)).
There was no effect of the form of the marking of accusative case (stem vs. affix
morphamb nouns) (F1(1,39)<1))1. Likewise, there was no effect of the complexity of the left
part of the DNP (F1(1,39)<1).
The number of the left part had a clear effect on acceptability. For items with a complex
left part, number could take two values (singular, plural). Plural items were more acceptable
than singular ones (3.88 vs. 3.04; F1(1,39)=52,7;p<0.001). There was a significant interaction
(F1(1,39)=16.35, p<0.001) between the number marking of the left part and the right part of
the DNP, with the number effect of the left part being larger when it matched the number of
the right NP. Among the matching items, plural DNP received a mean rating of 4.72, and
singular DNP, a mean rating of 3.04.
For items with a simple left part, the number of the left part could be singular, plural or
ambiguous. Again, there was a main effect of number (F1(1,39=30,82), p<0.001).
Unambiguous plural items were more acceptable than unambiguous singular items
(F1(1,39)=31.69, p<0.01). The interaction of the number value of the left and the right part
was again significant (F1(1,39)=13,77, p<0.001): the acceptability difference between singular
and plural left parts was larger when the number of the right part matched. For matching
items, unambiguous left plurals got a mean rating of 4.96, and unambiguous left singulars a
mean rating of 3.04 . Figure 1 represents the number effect for unambiguous DNP in which
the number values of the left and right part match.
Figure 1: Number Effect for DNP
7
6
5
Complex
4
3
Simple
2
1
Plural
Singular
The matching effect (number agreement between the left and right part of the DNP) was also
significant both for the items with a complex (F1(1.39)=40.45, p<0.001) and a simple left part
(F1(1.39)=13.77, p<0.001). Thus, as expected, both grammatical constraints on DNP that were
tested in the experiment exert a clear influence on acceptability.
The items beginning with a locally ambiguous left part of the DNP were of particular
interest for our experiment. As illustrated in Figure 2, the mean acceptability of DNP with a
plural left part 4.12 does not differ significantly from the mean acceptability of DNP with an
ambiguous left part 4.36 (F1(1,39)=1.51, p=0.23). DNP with a singular left part got a worse
rating (3.0, F1(1,39)=51.36, p<0.001).
Figure 2:Ambiguity of left part
7
6
5
4
3
2
1
Ambiguous
Plural
Singular
For the items with a simple left part, there also was a main effect of the number of the right
part (F1(1,39)=24.67, p<0.001) and a significant interaction between the number of the left
and the right part (F1(1,39)=13.77. p<0.001).
When the right part of the DNP has plural marking, plural left parts (4.97) and
ambiguous left parts (4.78,) cannot be distinguished (F1(1,39)=0,68, p=0.42), but there is a
clear difference to DNP with a singular left part (mean acceptability:2.7): F1(1,39)=43.6,
p<.001 (singular vs. ambiguous); F1(1,39)= (singular vs. plural)
Figure 3: Number effect: right plurals
7
6
5
4
3
2
1
Plural
Ambiguous
Singular
When the right part of the DNP has singular marking, the number of the left part has a
significant effect on acceptability (F1(1,39)=6.66, p<0.01). In contrast to DNP with plural
right parts, the DNPs with a singular right part show the following pattern: items with an
ambiguous left part (mean: 3.87) are more acceptable than both those with a singular left part
(mean: 2.98, F1(1,39)=14.38, p<001) and those with a plural left part (mean 3.23,
F1(1,39)=7.61, p<0.01).
Figure 4: Number effect: right singulars
7
6
5
4
3
2
1
Plural
Ambiguous
Singular
As already noted above, the constraints against singular DNPs and DNPs in which the two
parts do not agree for number are visible in our results. Furthermore, a local number
ambiguity of a left DNP-part has a very specific impact on acceptability:
When the right part of the DNP disambiguates it towards a plural interpretation,
sentences beginning with an ambiguous left part are as acceptable as sentences with a plural
left part (Figure 3). This is not surprising: For having a chance of being grammatical, the
morphologically ambiguous left part of the DP must be interpreted with the plural value, since
the constraint against singular DNP cannot be fulfilled otherwise. A right part of the DNP
with a plural marking constitutes no reason for abandoning this plural hypothesis.
Ambiguous left parts are, however, more acceptable than both singular and plural items
(Figure 4) when the right part bears a singular marking. That they are better than DNP with
plural left parts seems to be related to the fact that the ambiguous item can be interpreted as a
singular, so that the number agreement requirement can be taken to be fulfilled. That they are
furthermore better than singular left parts may be related to the fact that the ambiguous item
can be interpreted as a plural, so that the constraint against singular DNP can considered
fulfilled. There are two mitigating effects of local ambiguity, then, but they are based on two
incompatible interpretations of the ambiguous noun.
Experiment 1 has thus confirmed the expectation that the presence of a local ambiguity
can increase global acceptability. In particular, the results represented in Figure 4 are
compatible with the view that intermediate acceptability assessments influence global
acceptability: the option of a plural interpretation for a locally ambiguous noun leads to a
positive local acceptability value, because the ban against singular DNP appears fulfilled.
This positive local assessment contributes to the global acceptability of DNPs even when the
plural interpretation is later abandoned because a singular right part is detected. In contrast to
grammaticality, global acceptability does not only depend on the final structural analysis, but
also on the acceptability of intermediate analysis steps.
This acceptabiliy pattern can also be found with professional linguists. They are not
immune to such “spillover” effects increasing global acceptability. By e-mail, we asked more
than 60 linguists (nearly all syntacticians) with German as their native language for judgments
of 16 split constructions (no distractor items were used in order to increase the likelihood of a
reply). 45 linguists responded by sorting the sentences into the categories *, ?, and wellformed.
Figure 5: Linguists' judgments
50
40
okay
30
?
20
out
10
0
11a
(11)
a.
b.
c.
d.
(12)
11b
11c
11d
Koffer
hat
er
keinen zweiten
suitcase
has
he
no
Roten Koffer hat
er
keinen zweiten
red suitcase
gas
he
no
Professor
kennt sie
genau einen
professor
knows she
exactly one
second
kennt sie
genau einen
Old professor
knows she
exactly one
kennt sie
zwei
professors
knows she
two
ambiguous noun
second
Alten Professor
Professoren
12
unambiguous noun
As Figure 5 indicates, sentence (12) beginning with an unambigous plural items was accepted
by nearly all participants. Two thirds and more of the participants rejected sentences that
began with an unambiguous singular DNP (11b-d), while the reaction to the ambiguous item
(11a) was quite mixed. Linguists’ acceptability judgements thus also seem to be influenced by
the processing factors discussed above.
3.3 Experiment 2: Disjunctive coordination
Experiment 2 was carried out with two goals: First, we wanted to demonstrate a positive
effect of processing difficulty in a domain other than DNP. A second purpose was to test
whether local ambiguities can increase acceptability in syntactic constellations that do not
completely fit into the “classical” preferred reading/reanalysis constellation. We looked at a
construction involving a syntactically unsolvable agreement problem.
Subject-verb-agreement is mandatory in German. Disjunctive coordinations may lead to
expressivity problems (13a,a’), because there are no rules for computing person-number
features for disjoined DPs. Clausal coordination (13b) is a possibility. Sentences involving
clausal coordination (13c) may undergo coordination reduction deleting grammatically
identical material, which yields (13d) (see, e.g., Wilder 1997).
(13)
a.
er
oder
ich
*schlafe
/*schläft
/*schlafen
he
or
I
sleep.1sg
sleep3sg
sleep3pl
a’.
ich oder er *schlafe/*schläft/*schlafen
b.
er schläft oder ich schlafe
c.
er
schläft oder
sie
schläft
he
sleeps or
she
sleeps
er
schläft oder
sie
schläft
d.
While coordination reduction involving 1st and 3rd person singular subjects (13a,a’) sounds
odd, introspection suggests that the construction improves when ambiguous verb forms are
used. Singular modals and past tense verbs do not distinguish 1st and 3rd person overtly. The
examples in (14) sound much better than (13a).
(14)
a.
b.
er
oder
ich
darf
schlafen
he
or
I
may
sleep
ich
oder
er
rief
spät
an
I
or
he
called late
up
This intuition fits well into the hypothesis formulated above concerning global acceptability:
In order to determine the grammaticaliy of (13a,a’) and (14), three computations related to
agreement are necessary. The parser must check whether the verb matches the grammatical
features of the first and of the second noun phrase, and it must be determined whether positive
results for both NPs have been obtained on the basis of the same featural analysis of the verb.
In the case of an ambiguous verb as in (14), the two “local” agreement computations are
successful, and we expect this positive local assessment of acceptability to increase global
acceptability even though the third computation integrating the two local analyses fails.
We conducted a questionnaire study in order to test whether the use of personambiguous verbs increases acceptability.
Description of Experiment 2
Method
48 students of the University of Potsdam participated. They were paid for participation, or
received course credits. Participants rated 120 sentences for acceptability, on a seven point
scale (1=very bad , 7 = very good) As mentioned above, we had changed the ranking scale in
order to avoid similarity with school grades. There were 16 experimental items (4
items/condition) in a within subject design, and 104 items not related to the experiment.
Material
The experimental items had the structures exemplified in (13a.a’) and (14). They began with a
disjunction of the pronouns ich “I” and er “he” in either order. (ER-initial and ICH-initial
conditions). The verb form was always 3rd person singular, but in the ambiguous condition
(AMB), the verb also allowed a 1st person singular interpretation.
Results and Discussion
Figure 6 represents mean judgments of acceptability in Experiment 2. The mean acceptability
of structures beginning with er (13a, 14a) was 4.23, and is thus indistinguishable from the
4.24 mean acceptability of sentences beginning with ich (13a’,14b) (F1<1, F2<1).
Figure 6: Disjoined subjects
7
6
5
UNA
4
AMB
3
2
1
ER
ICH
However, there was a significant effect of ambiguity: the ambiguous structures (14) (AMB)
were rated better (4.5) than unambiguous ones (13a, a’) (UNA) (4.0), F1(1,47)=6,79, p<.05),
F2(1,15)=50.74,p<.001). There was no interaction between both factors
(F1(1,47)=1.1,p=.30,F2<1).
This ambiguity effect is in line with our expectations. The grammaticality of (13a,a’) or
(14) depends on whether the verb agrees with the two subject pronouns, i.e., two independent
computations of an agreement relation are necessary. In (13a,a’), the verb visibly disagrees
with one of the two pronouns. In (14), however, the ambiguous verb form in principle fulfils
the requirements of both pronouns, which makes a local perception of acceptabilty possible.
This local assessment influence the global acceptability of (14) in a positive way.
The results of Experiment 2 cannot be explained in terms of the assumption that
phonological identity is syntactically sufficient for conjunction reduction (so that (14) would
not violate any grammatical principle at all). The 2nd person plural verb form is often identical
with the 3rd person singular form, yet this ambiguity does not render coordinations more
acceptable (Matthias Schlesewsky, p.c.).
(15)
*Ihr
oder
you.pl or
er
kennt
die
he
know3sg/2pl the
Antwort
answer
The difference between (14) and (15) may derive from the fact that the verb forms involve an
additional difference of number in (15).
3.4 Experiments 3a and 3b: Fronted Verb Phrases
In the preceding experiments, local ambiguities involving number (Experiment 1) and person
(Experiment 2) increased global acceptability. Experiment 3 investigated whether case
ambiguities can have the same effect. Such a test was also called for because Featherston
(p.c.) reports that local ambiguities of grammatical function of relative pronouns lead to a
reduction of acceptability. In Experiment 3, we used a construction that allowed us to
maintain the overall design of our study: a sentence violates a certain grammatical principle,
but involves a local ambiguity that makes the structure appear well-formed in initial/local
computations.
German VPs can appear in the Spec,CP position immediately preceding the finite verb
in main clauses, as (16a) illustrates. Such structures are grammatical when the NP in the
fronted VP is an (underlying) object of the verb. The inclusion of an (underlying) subject (as
in 16b) is taken to be much less acceptable.
(16) a.
[VP einen Jungen geküsst] hat
a.acc
boy.acc kissed
has
sie
nicht
she
not
“she has not kissed a boy”
b.
??ein Junge
geküsst
a.nom boy.nom kissed
hat
sie
nicht .
has
her
not
“a boy has not kissed her”
Feminine and neuter nouns do not distinguish morphologically between nominative and
accusative. Unlike (16a-b), (16c-d) with feminine Frau involve a local ambiguity of the
sentence initial NP.
(16)
c.
d.
[VP
[VP
eine
Frau
geküsst]
hat
er
nicht
a.amb woman
kissed
has
he.nom
not
eine.
geküsst]
hat
ihn
nicht
kissed
has
him.acc
not
Frau
a.amb woman
The grammatical restriction against the inclusion of subjects in preposed VPs implies a
parsing preference for initially analyzing eine Frau as the object of the verb geküsst in (16cd). This analysis can be maintained in structures like (16c) in which the second NP is
nominative, but it must be abandoned in (16d) when the pronominal NP is parsed because it
bears accusative case, which identifies it as the object. (16d) should be less inacceptable than
(16b) if global acceptability is influenced by temporary acceptability values: unlike (16b),
(16d) initially appears to respect the ban against the inclusion of subjects in fronted VPs. We
tested this prediction in a questionnaire and a speeded acceptability rating experiment.
Description of Experiment 3a
Subjects
48 students of the University of Potsdam participated. They were paid for participation, or
received course credits. The 16 experimental items (4 per condition) were among the
distractors of Experiment 2.
Material
The experimental items had the structure illustrated in (17)
(17)
Fronted VP = ambiguous subject + verb
a.
Ein
schlaues
a.amb clever.amb
Mädchen
geküsst
hat
ihn
noch
nie
girl
kissed
has
him
not
yet
“a clever girl has not yet kissed him”
Fronted VP = ambiguous object + verb
b.
Ein
schlaues
a.amb clever.amb
Mädchen
geküsst
hat
er
noch
nie
girl
kissed
has
he
not
yet
“he has not yet kissed a clever girl”
Fronted VP = unambiguous subject + verb
c.
Ein
junger Mann besucht
a.nom young man
visited
hatte
ihn
erst
gestern
has
him
only
yesterday
hatte
er
erst
he
only
yesterday
“a young man visited him only yesterday”
Fronted VP = unambigous object + verb
d.
einen jungen Mann besucht
a.acc young man
visited
has
“he visited a young man only yesterday”
gestern
All experimental items involved a preposed VP. The NP in this VP could either bear overt
case morphology (unambiguous condition, 17c-d) or be unmarked for case (locally
ambiguous condition, 17a-b). The second NP in the sentence was a pronoun bearing the case
not realized by the first NP (unambiguous condition), or which disambiguated the initial NP
in the ambiguous condition towards an object or subject reading.
Results and Discussion
As figure 7 shows, fronted verb phrases that include a direct object are more acceptable than
Figure 7: VP-preposing (quest.)
7
6
5
AMB
4
UNA
3
2
1
SUB
OBJ
those that include a fronted subject (mean subject: 3.8 mean object: 4.95, F1(1,47)=34.74,
p<.001,F2(1,15)=37.26,p<.001).
Contrary to our expectation, the was no main effect of ambiguity (F1<1,F2<1) and no
interaction between both factors (F1(1,47)=2.69,p=.11),F2<1). We used the same material in a
speeded acceptability rating experiment.
Description of Experiment 3b
Method
26 students of the University of Potsdam received course credits for participation. The 64
experimental items (16 per condition) and 160 distractor items were presented word by word
on a computer screen. Presentation time was 400 ms/word followed by 100ms blank screen.
By pressing a yes or no button, the participants had to indicate whether they found the
sentence acceptable or unacceptable.
Material
The experimental items were the ones used in Exp. 3a.
Results and Discussion
Figure 8: VP-Preposing (speeded rating)
% acceptable
100
80
60
AMB
40
UNA
20
0
VP with SU
VP with OB
The results of Experiment 3b are represented in Figure 8. There was a main effect of the
grammatical function of the NP in the fronted VP: Fronted VPs that include an object were
rated acceptable in 89.5% of the case, contrasting with 54.6% for VPs including a subject
(F1(1,25)=85.65, p<.0001, F2(1,15)=127.7,p<.001). Items in which the NP in VP was caseambiguous were rated as acceptable more often (75.11%) than items in which there was no
ambiguity (69.0%) (F1(1,25)=6.98, p<.05,F2(1,15)=14.87,p<.05.) The interaction between the
grammatical function of the NP in VP and its ambiguity was also significant (F1(1,25)=10.25,
p<.01,F2(1,15)=29.14,p<.001): for VPs including objects, there was no ambiguity related
difference (ambiguous: 88.9%; unambigous 90.1%). VP-fronting that pied-pipes the subject
was considered acceptable in 61.4.% of the trials when the subject was not overtly case
marked, but only 47.7 of the trials were rated as acceptable when the subject-NP bore an
unambiguous Case marking.
The results of Experiments 3a and 3b demonstrate the syntactic restriction against
underlying subjects in fronted VPs. They show that local case ambiguities do not reduce
acceptability. Furthermore, Experiment 3b confirms our expectation that local ambiguities
may increase acceptability: the temporary fulfilment of the anti-subject constraint in structures
such as (17a) seems to render such structures more acceptable than examples such as (17c), in
which the violation of the anti-subject restriction becomes obvious very soon. Experiment 3b
is thus in line with what we found before in Experiments 1 and 2.
Unlike what we saw in the other studies, locally ambiguous and unambiguous structures
were equally acceptable in Experiment 3a, however. We can only offer some speculations
about the reasons for this difference. The resolution of the local ambiguity in Exp. 3 affects
the assignment of grammatical functions and the interpretation of the sentence, while the
number and person ambiguities had no such effect in Experiments 1 and 2. The need for
revising an initial interpretation may have negative consequences for acceptability that
override positive effects of local ambiguity (Experiment 3a). If this reduction of acceptability
takes place in a time window later than the one used in Experiment 3b, we would understand
why the speeded acceptability rating task shows a positive impact of local ambiguity on
acceptability.
4 Preferred readings decrease acceptability: short vs. long
movement
In Experiments 1-3, the positive results of early computations increased the global
acceptability of a sentence even when the outcome of these early computations had to be
revised later. In the final Experiment 4, we investigated the opposite constellation: what
happens when its initial segments make a grammatical sentence appear ungrammatical?
Experiment 4 focused on structures that Kvam (1983) had used in his informal studies
on long wh-movement constructions in German. He observed that the acceptability of such
sentences depends on whether the grammatical features of the phrase having undergone long
movement would also match requirements imposed by the matrix verb on its arguments. (18)
illustrates two structures in which the subject of a complement clause has been moved into the
matrix clause. (18a) is more acceptable than (18b). When (18a) is parsed, was “what” locally
allows an analysis as an object of the matrix clause, while wer “who” could neither function
as the subject nor as the object of the matrix clause.
(18)
a.
b.
?was denken Sie,
dass
die
Entwicklung beeinflusst hat
what
think you
that
the
development influences has
denken
Sie,
dass
die Entwicklung
beeinflusst hat
who.sg think.pl
you
that
the development
influenced has
??wer
“who/what do you think influences the development”
The same logic underlies the contrast between the relative clauses in (19), in which the
relative pronoun is extracted from an infinitival complement clause. In the more acceptable
(19a), the relative pronoun die also fits the accusative case requirement of the predicate
embedding the infinitive. In the less acceptable (19b), the dative case of the relative pronoun
clashes with the case requirement of the embedding predicate.
(19) a.
Eine Kerze,
die
a
which he for
candle
er für
gut
hielt, dem
Ludwig zu weihen
good
held
Ludwig to dedicate
the.dat
„a candle which he considered good to dedicate to Ludwig”
b.
Eine Frau,
a
der
er für
woman who.dat he for
angemessen hielt, ein Geschenk zu
geben
appropriate held a
give
present
to
Such contrasts in acceptability are predicted by the hypothesis we pursue in this paper, since
the human parser initially prefers a short movement analysis of wh-phrases or relative
pronouns. Whenever the grammatical features of a preposed element clash with the
requirements of the matrix clause, a perception of ungrammaticality will be triggered. We
expect this perception to decrease global acceptability even if the parser later finds a longmovement interpretation for the preposed phrase under which the sentence is fully
grammatical. Experiment 4 tested whether this expectation is borne out.
Description of Experiment 4
Subjects
48 students of the University of Potsdam participated. They were paid for participation, or
received course credits. The 16 experimental items were among the distractor items of
Experiment 2.
Material
Experiment 4 consists of two subexperiments, one for wh-questions, the other for relative
clauses. In the question subexperiment, the eight experimental items (4 per condition) had the
structure illustrated in (18). The subject of a dass “that”- complement clause is moved into
the matrix clause, consisting of a matrix verb and a pronominal subject.
In the unambiguous wh-condition (18b), the subject extracted from the complement
clause was nominative wer “who”. Because of its case, wer allows no intermediate analysis as
the object of the matrix clause, because of the disagreement with the plural verb, wer can also
not be analyzed as the matrix subject. In the ambiguous wh-condition (18a), the moved whpronoun was “what” is case ambiguous. In its accusative interpretation, it could figure as the
object of the matrix clause, in its (eventually mandatory) nominative interpretation, it is the
subject of the complement clause.
The eight items of the relative clause subexperiment were constructed as illustrated in
(19). A relative pronoun was extracted from an infinitival complement clause. In the
unambiguous relative condition (19b), the dative case of the relative pronoun did not match
the case requirements of the predicate embedding the infinitive. In the ambiguous relative
condition (19a), the accusative relative pronoun was compatible with the case requirements of
the embedding predicate.
Results and Discussion
Figure 9 graphically represents the mean acceptability of locally ambiguous and unambiguous
wh-questions. The acceptability of the locally ambiguous question (4.7) is much higher than
the one of the unambiguous construction (3.15) (F1(1,47)=30.05, p<0.001,F2 (1,7)=62.50,p<001).
Figure 9: wh-questions
7
6
5
4
3
2
1
Unambigous
Ambiguous
The feature incompatibility of the preposed complement clause wh-subject and the
grammatical requirements of the matrix verb in the unambiguous condition (18b) renders the
short movement interpretation for the wh-phrase ungrammatical, and this has negative
consequences for the global acceptability of the structure even though a long movement
analysis is possible.
We obtained a similar result for relative clauses (although the effect size was smaller): when
the relative pronoun locally allows a short movement interpretation, the structure is more
acceptable (4.76) than when the case of the relative pronoun clashes with the requirements of
the matrix clause (4.17) (F1(1,47)=8.28,p<0.01),F2(1,7)=3.73,p=.10)
Figure 10: Relative Clauses
7
6
5
4
3
2
1
Unambiguous
Ambiguous
Both subexperiments show the expected influence of the local ambiguity on global
acceptability: the failure of an initial short movement interpretation of the preposed phrase in
the unambiguous condition decreases acceptability. The effect size was much smaller in the
relative clause subexperiment: the unambiguous wh-condition is less acceptable than the
unambiguous relative condition (4.17 vs. 3.15). Several factors could be responsible for this:
Different case contrasts (nom-acc vs. acc-dat) were used in the two subexperiments, the
question subexperiment involved a additional number agreement clash, and the whexperiment involved movement out of a finite complement clause.
5 Conclusions
Our experiments reveal that local ambiguities influence the acceptability of a sentence. If our
interpretation of Experiments 1, 2, 3b and 4 is correct, there is a spillover from the
acceptability of the initial analysis of a locally ambiguous structure to the global acceptability
of the full construction. Structures violating constraints may appear more acceptable if their
parsing involves an intermediate analysis in which the crucial constraint seems fulfilled.
Grammatical structures may appear less acceptable when their processing temporarily
involves a stage in which a crucial constraint seems violated.
Processing difficulty (understood here as including the need to revise an initial analysis) can
thus have both positive and negative influences on acceptability. The details of the processing
profiles of sentences seem to be reflected in subtle differences in acceptability. Grammatical
models should not reflect such differences.
6 References
Bader, Markus & Lyn Frazier. Interpretation of leftward-moved constituents: Processing
topicalizations in German. Linguistics. To appear.
Bader, Markus & Michael Meng, M. 1999. Subject-object ambiguities in German embedded
clauses: An across-the-board comparison. Journal of Psycholinguistic Research, 28(2), 121143.
Bever, Tom. G. 1970. The cognitive basis for linguistic structures. In Hayes, J. R., editor,
Cognition and the Development of Language. Wiley, New York.
Chomsky, Noam. 1957. Syntactic Structures. The Hague: Mouton.
Chomsky, N. 1995. The minimalist program. Cambridge, Mass.: MIT-Press.
Chomsky, Noam. & Miller, George. A. 1963. Introduction to the formal analysis of natural
languages. In Luce, D. R., R., B. R., & Galanter, E., editors, Handbook of Mathematical
Psychology, volume II. John Wiley, New York.
Fanselow, Gisbert. 1988. Aufspaltung von NP und das Problem der ‘freien’ Wortstellung.
Linguistische Berichte 114: 91-113
Fanselow, Gisbert & Damit Cavar, 2002. Distributed Deletion. In: Artemis Alexiadou (Hg.),
Theoretical Approaches to Universals, Amsterdam, Benjamins. 65– 07.
Fanselow, Gisbert & Caroline Féry 2002. Ineffability in Grammar. In: G. Fanselow & C. Féry
(eds).. Hamburg: Buske. 265-397.
Fanselow, Gisbert, Reinhold Kliegl & Matthias Schlesewksy. 1999. Processing Difficulty and
Principles of Grammar. In: Susan. Kemper & Reinhold. Kliegl (eds). Constraints on
Language. Aging, Grammar, and Memory. Kluwer: Boston 171-201.
Fanselow, Gisbert, Reinhold Kliegl & Matthias Schlesewksy. In preparation. Long Movement in
Northern German. A Training Study. Ms., University of Potsdam.
Featherston, Sam. To appear. Universals and Grammaticality: WH.constraints in English and
German. Linguistics,
Felser, Claudia, Harald Clahsen & T. Münthe. 2003. Storage and Integration in the Processing of
Filler-Gap Dependencies: An ERP Study of Topicalization and Wh-Movement in German
Brain and Language 87 (2003), 345-354.
Fiebach, Christian, Matthias Schlesewsky, & Angela Friederici. 2002. Separating syntactic
memory costs and syntactic integration costs during parsing: The processing of German
WHquestions. Journal of Memory and Language.47: 250-272.
Frisch, Stefan, Matthias Schlesewsky, Doug Saddy, & A. Alpermann. 2001. Why syntactic
ambiguity is costly after all. reading time and ERP evidence. AMLAP Saarbrücken 2001.
Frisch, Stefan, Matthias Schlesewsky, Doug Saddy, & A. Alpermann. 2002. The P 600 as as an
indicator of syntactic amboguity. Cognition 85 B 83-B92.
Grimshaw, Jane. 1997. Projection, heads, and optimality. Linguistic Inquiry 28: 373-422
Hemforth, B., 1993. Kognitives Parsing: Repräsentation und Verarbeitung sprachlichen
Wissens. Sankt Augustin: Infix.
Van Hoof, Hanneke. 1997. On Split Topicalization and Ellipsis. Tech. Rep. 112, Arbeitspapiere
des Sonderforschungsbereichs 340, Tübingen.
Keller, Frank. 2000. Evaluating Competition-based Models of Word Order In: Lila R. Gleitman
and Aravid K. Joshi, eds., Proceedings of the 22nd Annual Conference of the Cognitive
Science Society, 747–752. Mahawah, NJ: Lawrence Erlbaum, 2000.
King, J. W., & Kutas, M. 1995. Who did what and when? Using word- and clause-level ERPs to
monitor working memory usage in reading. Journal of Cognitive Neuroscience, 7, 376-395.
Kluender, R., & Kutas, M. (1993). Bridging the gap: Evidence from ERPs on the processing of
unbounded dependencies. Journal of Cognitive Neuroscience, 5, 196-214.
Krems, J., 1984. Erwartungsgeleitete Sprachverarbeitung. Frankfurt/Main: Lang.
Kvam, Sigmund. 1983. Linksverschachtelung im Deutschen und Norwegischen. Tübingen: Niemeyer.
Lewis, Richard. 1993. An Architecturally-based Theory of Human Sentence Comprehension.
Doctoral Dissertation, Carnegie Mellon University.
Marks, Lawrence E. 1967. Judgments of grammaticalness of some English sentences and
semi-sentences. American Journal of Psychology 20: 196–204.
Matzke, M., Mai, H., Nager, W., Rüsseler, J., & Münte, T. F. 2002. The cost of freedom: An
ERP-study of non-canonical sentences. Clinical Neurophysiology. 113: 844-852
Meng, Michael. 1998. Kognitive Sprachverarbeitung. Rekonstruktion syntaktischer Strukturen
beim Lesen. Wiesbaden: Deutscher Universitätsverlag
Meng, Michael & Markus Bader 1999. Ungrammaticality detection and garden path strength:
evidence for serial parsing. Ms..
Müller, Gereon. 1999. Optionality, markedness, and word order in German. Linguistics, 37(5),
777–818.
Müller, Horst. M., King, J. W., & Kutas, M. (1997). Event-related potentials elicited by spoken
relative clauses. Cognitive Brain Research, 5, 193-203.
Müller Stefan. 2004. Complex NPs, subjacency, and extraposition. Snippets - Issue 8 .
Riemsdijk, Henk van 1989. "Movement and regeneration". In: Dialectal variation and the theory
of grammar, P. Benincà (ed.) Dordrecht: Foris, 105-136.
Schlesewsky, Matthias, Gisbert Fanselow, & Stefan Frisch. 2003- Case as a trigger for reanalysis
- Some arguments from the processing of double case ungrammaticalities in German. Ms.
Potsdam.
Schlesewsky, Matthias, Gisberg Fanselow, Reinhold Kliegl & Josef Krems. 2000. The Subject
Preference in the processing of locally ambiguous wh.questions in German. Hemforth &
Konieczny. German sentence processing 65 – 93. Boston: Kluwer.
Snyder, William, 2000. An experimental investigation of syntactic satitation effects. LI 31: 575582.
Wilder, C. (1997) "Some Properties of Ellipsis in Coordination". In: A. Alexiadou & T. Hall
(eds.) Studies in Universal Grammar and Typological Variation. John Benjamins, Amsterdam
[=Linguistik Aktuell 13], 59-107.
1
Given that the distinction between (9a,b vs. e,f) could not be constructed for morphamb nouns, item analyses
could not be computed in a meaningful way for Exp. 1