Introduction
Numerous studies have focused on the
processing of structurally ambiguous
sentences in order to shed light on the
nature of the human sentence processing
mechanism. Several studies have shown
that there are cross-linguistic differences
in the resolution of ambiguous relative
clauses (RCs) like Someone shot [the
servant]DP1 of [the actress]DP2 [who was
on the balcony]RC, where the RC can
modify either DP1or DP2 in the
preceding complex DP. A high (DP1)
attachment preference has been reported
in Dutch (Brysbaert & Mitchell, 1996;
Desmet, Brysbaert & De Baecke, 2002),
French (Baccino, De Vincenzi & Job,
2000; Colonna, Pynte, & Mitchell,
2000; Quinn, Abdelghany & Fodor,
2000, Zagar, Pynte, & Rativeau, 1997),
German (Hemforth, Konieczny,
Scheepers & Strube, 1998; Hemforth,
Konieczny & Scheepers, 2000), Persian
(Arabmofrad & Marefat, 2008), and
Spanish (Cuetos & Mitchell, 1988),
whereas a low attachment (DP2)
preference has been found in Brazilian
Portuguese (Finger & Zimmer, 2000;
Miyamoto, 1998), English (Cuetos &
Mitchell, 1988; Frazier & Clifton, 1996;
Deevy, 2000), Norwegian, Swedish, and
Romanian (Ehrlich, Fernandez, Fodor,
Stenshoel & Vinereanu, 1999). A number
of
attempts
have
been
made to account
for these cross-linguistic differences.
The dominant account in the literature for
RC attachment preferences has been
based upon Recency (Gibson,
Pearlmutter, Canseco-Gonzalez, &
Hickok, 1996), Late Closure principle
(Frazier & Fodor, 1978), or Right
Association (Kimball,1973), which
assume that all human languages are
processed in the same way. Frazier
(1978) defined Late Closure as "When
possible, attach incoming material into
the clause or phrase currently being
parsed" (p. 49). This is similar to
Kimball's (1973) Right Association
principle "Terminal symbol is optimally
associated to the lowest nonterminal
node" (p. 24). According
to Recency
principle, constituents such as RC
modifiers are attached to the most
recently processed (or closest) phrase,
regardless of the language being
processed (Fernandez, 2003). The
principle of Late Closure has been
found to apply in a number of
languages, with a wide variety of
constructions.
Cuetos and Mitchell (1988) were among
the first to challenge the ‘universalist’
view; they showed that parsing does not
proceed in the same way in all languages.
In their experiment, they gave
constructions comparable to Someone
shot [the servant]DP1 of [the actress]DP2
[who was on the balcony]RC to English
and Spanish native speakers. They found
that English subjects tended to attach
RCs to DP2. On the other hand, speakers
of Spanish showed
an overall high
attachment (DP1) preference. This led
them to conclude that certain parsing
strategies may not be universal, and that
there exists cross-linguistic variation.
Following
Cuetos
and Mitchell, further
studies were conducted in order to
examine RC
attachment preferences in
other languages. The results have
provided additional support for the view
that the Late Closure principle might not
be generalized cross-linguistically, and
there exist strategies that are more likely
to be operative in other languages.
The data obtained from Gibson et al.
(1996) experiment on sentences with
three potential antecedents for the
ambiguous RCs in English and Spanish
provided evidence for developing the
Predicate Proximity principle, as another
factor that competes with the universal
Recency principle. According to
Predicate Proximity, incoming
constituents are favorably attached to a
verb argument. Gibson and Pearlmutter
(1998) argued that in certain languages
like
German, Russian and Spanish,
Predicate Proximity outranks the
Recency principle and leads to a high
(DP1) attachment preference. By
contrast, in configurational languages
such as English, Norwegian, and
Swedish, ambiguous modifiers attach to
the most recent phrase in harmony with
the locality principle of Recency (Felser,
Roberts, Marinis & Gross, 2003), leading
to low-attachment instead.
Ambiguous RCs in Persian
According to Karimi (2001, p. 31),
“Persian is a null-subject verb final
language with SOV word order in
declarative sentences and subordinate
clauses”. Similar to English, Persian RCs
are post-nominal, and since there is no
relative pronoun in Persian, RCs are
always introduced by complementizer ke
in Persian (Taleghani, 2008, p. 84).
Previous studies have shown that Persian
native speakers have a high attachment
preference (Arabmofrad & Marefat,
2008; Marefat & Meraji, 2005;
Moghadassian, 2008). Forty five Persian-speaking monolinguals participated in an
on-line study conducted by Arabmofrad
and Marefat (2008). Fifteen sets of items
were developed and normed; each set
contained sentences in three conditions:
1) RC was semantically related to DP1
and could be attached only to it and not to
DP2 (sentence 1 below); 2) RC could be
attached only to DP2, based on a
semantic relationship (sentence 2 below);
3) an ambiguous condition in which RC
could be attached to both DP1 and DP2
(sentence 3 below).
(1) an mærd pæræstar-e nozad [ke dašt
ghædæm mizæd] ra did
‘That man saw the nurse of the infant
[who was walking].’
(2) an mærd nozad-e pæræstar [ke dašt
ghædæm mizæd] ra did
‘That man saw the infant of the nurse
[who was walking].’
(3) an mærd dokhtær-e pæræstar ke dašt
ghædæm mizæd ra did
‘That man saw the daughter of the nurse
[who was walking].’
The rationale behind the study was that if
the participants prefer DP1 attachment,
then the reaction time for sentences in
which the RC is, contrary to their
expectation, semantically related only to
DP2 would be slower than that for
sentences in which the RC is semantically
related to DP1. And, conversely, if they
have a tendency to attach RC to DP2,
then, reaction time for sentences in which
the RC could only be attached to DP1
would take longer. The participants,
tested individually, were required to
make grammaticality judgments about
the sentences that were presented in a
non-cumulative way. Decisions as well as
decision times were automatically
recorded. The results for accuracy
responses showed no difference between
the three conditions which implies that
after reading the sentences, participants
accurately made grammaticality
judgments about the sentences. However,
analysis of the reaction times for
grammaticality judgment of the sentences
in the three conditions showed that the
participants took shorter reaction times
for sentences in which due to a semantic
cue, RC had to be attached to DP1, but
longer reaction times to sentences in
which the RC had to be attached to DP2.
Moreover, there was no difference
between ambiguous sentences and those
in which RC referred to DP1. These
results show a high attachment preference
by Persian native speakers.
Factors Affecting Attachment Preferences
within a Language
There are a number of individual-level
factors, such as proficiency and WMC,
which can cause intra-lingual differences
in attachment preferences.
As for proficiency, Miyao and Omaki
(2006) used an off-line and an on-line
self-paced reading task to examine RC
attachment preferences of intermediate to
advanced Korean L2 learners of Japanese
and Japanese native speakers. Results
from the off-line sentence interpretation
task showed that both Korean L2 learners
and Japanese natives preferred high
attachment. However, results from the
on-line self-paced reading task showed
that Japanese natives preferred high
attachment but Korean L2 learners
preferred low attachment. To account for
these results, Miyao and Omaki stated
that there may be three different stages in
development of L2 processing: L1
transfer phase, intermediate phase, and
target-like phase. They defined the L1
transfer phase as the stage in which low-proficiency L2 learners transfer their L1
grammar, including their L1 parsing
preference. In the intermediate phase,
medium-level L2 learners’ grammar and
parsing preferences are still developing
and not yet efficient, i.e., their grammar
is non-target like and includes traces of
their L1. In the target-like phase, high-proficiency L2 learners have target-like
parsing preferences. They stated that
participants in their study were in the
intermediate phase, and as a result they
resorted to a parsing strategy that
minimized their processing burden (i.e.,
locality principle), which resulted in low
attachment.
Although some L2 studies on RC
attachment preferences have found no
effect of WMC, numerous studies have
shown that individuals with high WMC
process syntactically ambiguous
sentences differently from those with low
WMC (Kim & Christianson, 2013;
Mendelsohn & Pearlmutter, 1999; Swets,
Desmet, Hambrick, & Ferreira, 2007;
Vos, Gunter, Schriefers, & Friederici,
2001). Swets et al. (2007) examined the
role of WMC in RC attachment
preferences of English and Dutch native
speakers. Test sentences were ambiguous
structures such as: [The maid]DP1 of [the
princess]DP2 [who scratched herself in
public]RC was terribly embarrassed.
They reported a negative correlation
between WMC and RC attachment
preference. To account for these findings,
they proposed the ‘chunking’ hypothesis:
low span readers do not have enough
resources to parse a sentence without
stopping at intermediate places, which
causes them to pause before the RC. This
pause in processing makes them chunk
DP1 and DP2 into a single unit, leading
them to attach the following RC to the
head of the complex DP, i.e., DP1 (the
maid).
Fodor (2002) proposed the Implicit
Prosody Hypothesis (IPH) to justify the
intra-lingual variation in attachment
preferences. According to IPH, different
prosodic groupings of a sentence can
result in different interpretations. In case
of ambiguous RCs, a prosodic grouping
of (DP1) (DP2 RC) reflects a low
attachment; while a prosodic grouping of
(DP1 DP2) (RC) demonstrates a high
attachment interpretation. Thus, for
sentence (4a), where the prosodic
grouping is (DP1) (DP2 RC) and
prosodic boundary is put after DP1 (i.e.,
servant), the parser is likely to attach the
ambiguous RC to DP2 (i.e., the actress),
while, in sentence (4b) where the
prosodic grouping is (DP1 DP2) (RC)
and prosodic boundary comes after DP2
(i.e., the actress), the parser interprets the
RC as modifying DP1 (i.e., the servant).
(4) a. Someone shot the servant #of the
actress who was on the balcony.
b. Someone shot the servant of the
actress #who was on the balcony.
More support for IPH has been provided
by Jun (2003) who examined native
speakers of English, French, Greek, Farsi
(Persian), French, Japanese, Korean, and
showed that in each language there is a
correlation between attachment
preferences and the default prosody
assigned upon reading ambiguous
sentences containing RCs.
The Present Study
This study undertakes to see if the
different findings across and within
languages with regard to the resolution of
ambiguous RCs can be accounted for by
the role of semantics. Previous research
has shown that the effect of semantics is
“strong enough to over-ride any phrase
structure based locality principle that
might otherwise favor NP1 attachment”
(Felser, et al., 2003, P. 457). But the type
semantics Felser et al. refer to is limited
to the distinction between thematic
preposition with and the case assigner
preposition of. Since the former
preposition constructs a local thematic
domain, the ambiguous RC is associated
with the DP inside this domain. Thus,
semantics was limited to the lexical
semantic features of prepositions and the
results from many studies (Felser et al.,
2003; Papadopoulou & Clahsen, 2003)
have shown that L2 learners are sensitive
to the bias provided by the preposition
with. But in this study, semantics has a
broader domain. In the two experiments
reported in this study, the sentence
contexts are varied to establish a semantic
relationship between the subject of the
main clause and one of the DPs in a
complex DP (Experiment 1) and the verb
of the main clause and one of the DPs in
the complex DP (Experiment 2). In
sentence (5) below, the subject of the
matrix sentence is related to DP1, and in
sentence (6) it is related to DP2.
(5) The doctor saw [the nurse]DP1 of [the
pupil]DP2 [who was in the yard]RC.
(6) The teacher saw [the nurse]DP1 of [the
pupil]DP2 [who was in the yard]RC.
When the parser starts reading a sentence,
the first piece of information s/he
encounters is most often the subject of the
sentence. When the first DP, the doctor,
in the case of sentence (5), is activated,
based on the Spreading Activation Model
(Collins & Loftus 1975; Dell, 1986),
other words that are semantically related
to it also become activated, and when the
parser gets to DP1 the nurse, the two
semantically associated words reinforce
the activation of each other and remain
more accessible in comparison to DP2 the
pupil. Encountering the RC, who was in
the yard, the parser is expected to attach
it to the more accessible DP which is the
nurse. Moreover, being the subject of the
sentence, the doctor occupies a
syntactically prominent position and thus
its accessibility is enhanced (Bock &
Warren, 1985; Brennan, 1995; Brennan,
Friedman, & Pollard, 1987; Prat-Sala &
Branigan, 2000) and this makes its
semantically related item in the sentence
(i.e., the nurse) more accessible as well.
Another kind of semantic relationship
that may affect RC attachment preference
is when the verb of the main clause is
semantically related to one of the DPs in
the complex DP preceding RC. Altmann
and Kamide (1999) reported an eye-tracking study in which participants
listened to sentences such as The boy will
eat the cake and The boy will move the
cake while they viewed a scene
containing a boy who was sitting on the
floor surrounded by various items such as
a toy train set, a ball, a toy car, a balloon,
and a birthday cake. When participants
heard the verb eat, they tended to look at
the cake more often compared to when
they heard the verb move. This happened
because the selectional restrictions of the
verb eat prescribed that only one object in
the visual scene could be relevant (i.e.,
the cake), but the verb move could refer
to all of the movable objects in the scene
(i.e., the toy train, the ball, the toy car, the
balloon, and the cake). This finding
shows that by using the selectional
restrictions of the verb, comprehenders
are able to predict an upcoming direct
object of the verb. In sentence (7) below,
encountering the verb inject, the parser
“not only analyses [it] . . . but also
predicts upcoming unseen elements”
(Arai & Keller, 2013, p. 525). Thus, when
the parser sees the nurse (DP1), a strong
association is developed between the verb
and this DP which is expected to make it
more accessible in comparison to the
lawyer DP2, when the parser is at the
stage of attaching the RC to a preceding
DP. In sentence (8), on the other hand,
there is a strong semantic relationship
between the verb defend and DP2 the
lawyer compared to DP1 the nurse, and is
thus more accessible, and if priming plays
a role, the RC is expected to attach to
DP2.
(7) Someone injected [the nurse]DP1 of
[the lawyer]DP2 [who was on the
balcony]RC.
(8) Someone defended [the nurse]DP1 of
[the lawyer]DP2 [who was on the
balcony]RC.
Thus, in line with previous studies
(Schafer, Carter, Clifton, & Frazier,
1996), we predicted that the primed word
would become more salient and therefore
would attract the ambiguous RC. The
present study builds on the above
findings and aims to delve into the role of
priming in RC attachment preferences of
Persian L2 learners of English. In
particular, this study aims to provide
answers to the following question:
Does priming one of the DPs in the
complex DP through associating it with
the subject/verb of the sentence influence
the RC attachment preferences of Persian
L2 learners of English with different
proficiencies and WMCs?
Experiment 1
The aim of this experiment was to explore
the impact of semantic priming created
by the association between the subject of
the matrix sentence and one of the DPs
preceding RC on participants’
preferences with different proficiency
and WMCs.
Method
Participants
33 Persian-speaking learners of English
(mean age 20, range18-22, 11 females)
majoring in English Language and
Literature at two universities participated
in this experiment as a course
requirement. They were B.A. students
and had not received advanced linguistic
instruction and had no idea about the aim
of the study. Persian was the L1 of all
participants and they had all started
learning English at high school. Two
participants were excluded because they
had not completed the proficiency test
and 2 more ones were excluded because
they could not meet the criterion of 90%
comprehension accuracy of fillers (see
below for details). In this way, data from
29 participants were used for data
analysis.
Instruments
The Operation Span Task (OST)
The vast majority of previous studies
have used the reading span task as the
sole index of WMC while there exist
many different assessments, which tap
different working memory mechanisms
(Conway, Kane, Bunting, Hambrick,
Wilhelm & Engle, 2005). The reading
span task assesses the ability to sustain
and process information through reading
sentences while participants’ attention is
split. In this study, OST was used rather
than the reading span task. In OST,
instead of sentences, mathematical
operations are used. Both tasks (i.e.
reading span task and operation task)
have been shown to predict sentence
comprehension performance (Turner &
Engle, 1989) but O’Rourke (2013) stated
that only operation span predicts
accuracy for syntactically complex
sentences.
The OST was administered using
Microsoft PowerPoint 2010. Participants
were presented with sets of simple
equations ranging from two to five
equations per set. There were three trials
for each set size, resulting in a total of
orty-two (3 × (2 + 3 + 4 + 5) = 42)
equations for the entire test. A sample set
including three items is presented below:
a. 5 + (8 × 2) = 21, ?, Z
b. (2 + 9) – 4 = 5, ?, Y
c. 5 × (7 – 2) = 45, ?, B
???
Before starting the task, a 15-item warm-up activity was administered. Then the
42-item OST was presented in a fixed
order for all participants. Each item
appeared on the screen and remained
there for 5 seconds. The length of this
interval was based on the findings of a
pilot study. Then a question mark (‘?’)
appeared on the screen, at this point
participants were instructed to judge the
accuracy of the equations by saying ‘true’
or ‘false’. After the equation judgment,
participants were instructed to press the
space bar. Then a capital letter appeared
on the screen to be read aloud. After 3
seconds this letter disappeared and the
participants proceeded to the next item.
When the participant reached the last
item in a set, three question marks (‘???’)
appeared. The participants were
instructed to stop at this point and recall
the letters in the order in which they had
appeared in the set. The experimenter
recorded the responses on an answer
sheet.
Two scores were reported based on this
test: one for the true/false answers to the
equations (judgment accuracy) and one
for the number of letters recalled
accurately. The correlation coefficient
between these two scores was .80 (p =
.000). Since there was the possibility that
the participants focus on the letters they
were to recall (which was a measure of
their WMC) and take the truth value of
the equations not seriously, their
performance on equations was
considered as a criterion for selection.
In order to be qualified to participate in
the study participants were required to
correctly judge at least 38 equations out
of 42 (i.e. 90 percent).
As for the OST, one point was given for
every letter correctly recalled in the
correct order. The scores ranged between
25 and 42; and the mean and standard
deviation were 32 and 5.31, respectively.
Reliability of this test, based on the KR-21 formula, turned out to be .63.
Proficiency Test
Prior to the main test, the participants’
proficiency in English was assessed
through the Oxford Placement Test
(OPT, Allan, 2000). The scores ranged
between 22 and 54 (out of a maximum
possible score of 60); and the mean and
standard deviation were 39.59 and 7.25,
respectively. KR-21 reliability of this test
turned out to be .76.
Descriptive statistics for the OST as a
measure of WMC and OPT as a measure
of proficiency are documented in Table 1
below.
The Main Test
The main test consisted of 70 sentences
including 10 practice items (three of
which also served as warm-ups across
four versions of the main test), 20 test
sentences, and 40 fillers
Test Sentences
The test sentences were all structurally
ambiguous sentences containing a main
clause and an RC, which could refer to
either of two preceding DPs that were
linked together by genitive of and
functioned as the object of the sentence.
The RC, in all sentences, was introduced
by the relative pronoun who. The subjects
of the sentences were animate and
represented different occupations such as
teacher, lawyer, doctor, etc.
Based on the relationship between the
subject of the main clause and either of
the two DPs in a complex DP, test
sentences were categorized into three
types: DP1-biased subject in which the
subject of the main clause and DP1 were
semantically related; DP2-biased subject
in which the subject of the main clause
and DP2 were related; Unbiased subject
with no specific relationship between the
subject of the main clause and either of
DPs. Examples for each category are
provided below:
DP1-biased subject
(9) The doctor saw the nurse of the pupil
who was in the yard.
DP2-biased subject
(10) The teacher saw the nurse of the
pupil who was in the yard.
Unbiased subject
(11) The lawyer saw the nurse of the pupil
who was in the yard.
Sentences (9), (10) and (11) were
regarded as a set of test sentences.
To assess the relationship set by the
researchers between the subject and
either of the DPs, a norming study was
conducted. Seven applied linguists and
25 participants from the same pool as in
the main study, none of whom
participated in the main experiment, took
part in the norming study. They were
asked to decide whether there was an
occupational semantic relationship
between the underlined words. All the
items were likert-scaled, ranging from 0
to 4, where 0 meant there was not any
semantic relationship between the words
and 4 meant they were strongly related.
Based on the results of the norming, 10
sets (those for which 90% of the subjects
had rated the semantic relationship as 4)
were selected. Each set had 3 test
sentences, so we had a total of 30
sentences to be used in the main test.
These 30 test sentences were divided into
two versions to make the test shorter and
to discourage the participants from
developing any strategies. Each version
included 10 subject-biased sentences
(five DP1 and five DP2 subject-biased
sentences). For each DP biased item,
there was an unbiased item. So there were
10 unbiased-subject sentences in each
version. If the DP1-biased-subject item of
a set appeared in Version 1, it was
replaced with DP2-biased-subject item in
Version 2. The unbiased-subject item of
each set was common in both versions.
Filler Sentences and Comprehension
Questions
Forty filler sentences were developed.
Like test sentences, all the fillers included
RCs which were introduced with a
relative pronoun. Thirty-five out of 40
filler sentences were not ambiguous. The
remaining 5 filler sentences were
ambiguous and looked exactly like test
sentences, but it was the unambiguous
part that was questioned.
The filler items were used to ensure that
the participants attended to the content of
the sentences they read on the monitor.
So participants whose score on filler
sentences was less than 90% were
excluded.
All the sentences were followed by a fill-in-the-blank comprehension question; (1)
to find out which DP was recognized by
participants as the host of the RC; (2) to
check whether participants paid attention
to the content of the test. With regard to
fillers, the questions only served the
second purpose because their answers
could be checked for accuracy. The
answers to the ambiguous experimental
items simply indicated a participant's
preference and could not be checked for
their truth value. A sample item of a test
sentence followed by the corresponding
comprehension question is provided
below.
(12) The doctor saw the nurse of the
student who was in the yard.
………………was in the yard.
To ensure that ordering had no effect, the
item presentation order in the two
versions of the main experiment was
reversed. So there were four versions
including 63 items: 3 warm-up sentences,
20 test sentences, and 40 fillers. Warm-up
and filler sentences were the same across
the four versions.
Procedure
The OST and the main test were
administered in two different sessions
with a one week interval. The Rapid
Serial Visual Processing paradigm (in
which parts of a sentence are presented in
a time-controlled manner, this short
period of time between the presentations
of different parts of the sentence prohibit
development of any strategy by the
participants) was adopted and each
stimulus was presented on the screen for
nine seconds in black letters on a white
background. This interval was decided
upon based on the findings of a pilot
study. After that a fill-in-the-blank
comprehension question appeared on the
screen and participants were required to
provide an answer to the question by
typing in the specified space. Participants'
answers were automatically recorded.
Scoring System
We identified whether the participants
referred to DP1 or DP2 in each item. In
case a participant completed the sentence
with the whole phrase “DP1 of DP2”, this
would count as a DP1 choice because it is
the head the whole phrase. Out of the 580
answers (29 participants × 20 items)
provided, 378 were DP1 and 178 were
DP2 and 24 (4.14%) were not included in
the analysis because they referred neither
to DP1 nor to DP2.
In this study, WMC and proficiency
scores were not used to classify
participants into groups because such
classification leads to “loss of statistical
power” (1983, p. 260). Moreover, as
Conway et al. (2005) put it “information
and power are lost, because less
variability is captured by categories than
a continuum” (p. 782).
Results
Before analyzing the data, answers to the
comprehension questions following the
fillers were checked to ensure the
participants had read the sentences
attentively. Those with accuracy scores
lower than 90% (2 participants) were
excluded from analysis.
Table 2 shows the frequency and
percentage of DP1 and DP2 choices
across the three conditions. In the DP1
biased condition, 67.4% of the responses
referred to DP1 while only 32.6% of the
responses referred to DP2. Similarly, in
the DP2 biased condition, a large
percentage of replies, i.e., 72.5% referred
to DP1, but just 27.5% of the replies
referred to DP2. In the unbiased
condition, too, DP1 responses were twice
as many as DP2 replies (66.1% vs.
33.9%, respectively). As can be seen,
irrespective of the semantic
manipulation, the participants rarely
selected DP2 as the antecedent of the RC.
The following table presents the
correlation coefficient between WMC
and proficiency and DP1 attachment
preferences across different conditions.
As the table displays, WMC correlates
with DP1 choices negatively and
significantly in all conditions, meaning
that as WMC increases, preference for
DP1 decreases. As for proficiency, the
correlations are negative in all cases but
do not reach significance.
The Pearson correlation between total
DP1 choices (the total of DP1 choices in
the three conditions) and WMC indicates
that the effect size of the correlation was
medium (r = -.442, N = 29, R2 = .195). On
the other hand, the Pearson correlation
between total DP1 choices and
proficiency was not statistical and the
effect size turned out to be small (r = -.225,
N = 29, R2 = .051).
The following figure shows the scatter
plot for the correlation between WMC
and total DP1 choices. The loess curve,
having smoothed the data, shows that
preference for DP1 across different
conditions decreases as WMC increases.
As is evinced in the figure, there is a sharp
slope indicating that when WMC
increases, preference for DP1 severely
drops, while those with lower WMCs
prefer DP1.
We used a Mixed Effect Model to
evaluate the relation between different
variables with preference. All statistical
analysis were carried out using R (R Core
Team (2014). R: A language and
environment for statistical computing. R
Foundation for Statistical Computing,
Vienna, Austria. URL). Among the 9
models including the simplest model with
only one variable to the most complicated
one with all interactions, we chose the
model with the best AIC. The inclusion of
interactions did not improve the model,
so they were not included in the final
model. This model includes the cross
random coefficient of condition and
random intercept defined in the levels of
subject and item and with main effects of
condition, standardized proficiency and
WMC. The results are demonstrated in
Table 4. These results revealed no
significant effect for condition and
proficiency (ps > .05), but WMC had a
significant effect (P < .05), indicating that
those who selected DP1 had a
significantly lower WMC. In this way,
semantic manipulation was found to play
no role in the participants’ preferences.
Experiment 2
This experiment explored whether
Persian-speaking EFL learners' RC
attachment preference is affected by the
semantic relationship between the verb of
the main clause and one of the DPs in a
complex DP across different proficiency
and WMCs.
Method
Participants
A total number of 33 participants (14
females) were selected from the same
pool as those in Experiment 1. None of
them had participated in Experiment 1.
Their ages ranged between 18-22 years.
Two participants were excluded because
they did not show up for the main test.
And 2 more were excluded since they did
not satisfy the requirement on the fillers
(90% accuracy).
Descriptive statistics for WMC and
Proficiency are documented in Table 5.
Martials
The OST, OPT, and the practice test,
warm-up sentences, filler sentences, and
fill-in-the-blank comprehension
questions were similar to those used in
Experiment 1 in terms of structure and
number. Test sentences that were
different are elaborated below.
Test Sentences
The structure of experimental sentences
was similar to that used in Experiment 1,
but in Experiment 2 a semantic
relationship was established between the
verb of the main clause and one of the
DPs in the complex DP. Each pair of
words, i.e., the DP and the verb, was
chosen on the basis of their semantic
relationship. The following website was
consulted to determine the semantic
relationships, http://semantic-link.com.
Based on the relationship between the
verb of the main clause and either of the
two DPs in the complex DP, test
sentences were categorized into three
types: DP1-biased verb with an
occupational semantic relationship
between the verb of the main clause and
DP1; DP2-biased verb with a relationship
between the verb of the main clause and
DP2; and Unbiased verb with no specific
relationship between the verb of the main
clause and either of the DPs. Examples
for each category are provided below:
DP1-biased verb
(13) Someone cured the doctor of the
teacher who was preparing to go home.
DP2-biased verb
(14) Someone scored the doctor of the
teacher who was preparing to go home.
Unbiased verb
(15) Someone saw the doctor of the
teacher who was preparing to go home.
Sentences (13), (14), and (15) are
regarded as a set of experimental
sentences. Similar to Experiment 1, a
norming study was conducted to establish
the semantic relatedness of the verbs and
DPs.
A sample experimental item and a sample
filler item followed by their
corresponding comprehension questions
are provided below.
(16) Someone cured the doctor of the
teacher who was preparing to go home.
………………was preparing to go home.
(17) Someone knew the police officer to
whom I gave my passport.
I gave him my………………… .
As in Experiment 1, there were four
versions of the main test and each version
included 63 items: 3 warm-up sentences,
20 test sentences, and 40 fillers.
Procedure
Procedure was identical to that of
Experiment 1.
Scoring System
The scoring system was the same as that
of Experiment 1.
Results
Table 6 presents the frequency and
percentage of DP1 and DP2 choices
across the three semantically manipulated
conditions. As in the previous
experiment, the participants’ preference
for DP1 was stronger for DP1 than DP2
irrespective of the condition. In each of
the three conditions, preference for DP1
was twice more than that for DP2. This
finding is comparable to that in
Experiment 1.
Table 7 displays the correlation
coefficient between WMC and
proficiency with DP1 attachment
preferences across the three conditions.
In the same vein as in Experiment 1,
WMC correlated significantly negatively
with DP1 choices in all conditions and
with the total DP1 choices. Proficiency,
too, had a negative correlation with DP1
preferences but failed to reach
significance.
The Pearson correlation between total
DP1 choices and WMC was -.508 and the
effect size was medium (R2 = .258). On
the other hand, the Pearson correlation
between total DP1 choices and
proficiency was not statistical and the
effect size was small (r = -.252, N = 29,
R2 = .063).
The figure below shows the scatter plot
for the correlations between WMC and
total DP1 choices. As WMC increases,
preference for DP1 decreases.
Figure 2. Scatter plot for the correlations
between WMC and total DP1 choices in
Experiment 2
To analyze the data, we followed the
same procedure as in Experiment 1, by
evaluating different models including the
simplest one with only one variable to the
most complicated one, with all
interactions, and ending up in a model
with the lowest AIC. The best fitting
model includes the crossed random
coefficient of condition and random
intercept defined in the levels of subject
and item and with main effects of
condition, standardized proficiency and
standardized WMC. Table 8 reports a
summary of all coefficients for the choice
of DP. The main effect obtained belonged
to WMC (p = .008), with more choices
for DP1 by those having lower WMC.
Proficiency and condition had no main
effects (ps > .05).
Discussion and Conclusion
This study investigated whether
ambiguity resolution by Persian-speaking
learners of English as an L2 is sensitive
to priming one of the DPs in the complex
DP by creating a semantic relationship
between the subject/ verb of the main
clause and one of the DPs. The impact of
proficiency and WMC as individual
properties of the participants was also
examined.
The findings showed no priming effect.
In DP1 related condition, as in The doctor
saw the nurse of the pupil who was in the
yard, the first DP, the nurse was expected
to be more accessible, following the
Spreading Activation Model (Collins &
Loftus 1975; Dell, 1986), and hence
selected as the host of the ambiguous RC,
but the second DP the pupil was not. As
far as L2 learners with low WMC and
proficiency are concerned, this prediction
was borne out, but this preference was not
due to DP1’s being primed through its
association with the subject of the main
clause, because even in the DP2 related
condition, as in The teacher saw the nurse
of the pupil who was in the yard, where
DP2 the pupil was expected to be more
accessible, DP1 the nurse was selected as
the antecedent of the RC. The same
results were obtained in the unbiased
condition as well. As for the verb related
data, the findings were exactly the same.
In the three semantically different
conditions, the low WMC L2 learners’
preference was determined by the
Predicate Proximity principle, which
favors attachment to DP1. Earlier results
from different studies (Arabmofrad &
Marefat, 2008; Marefat & Meraji, 2005)
have shown that in Persian, the native
language of the participants of this study,
in which adjuncts can occur between the
verbs and their complements, Predicate
Proximity is operative rather than Late
Closure. Thus, semantic manipulation
doesn’t seem to influence the L2 learners’
preference for the antecedent of
ambiguous RCs. The theoretical account
for these findings may be viewed as
consistent with theories in which a
structural analysis of a newly
encountered word is constructed (Frazier
& Fodor, 1978; Frazier & Rayner, 1982).
But L2 learners with higher WMCs
favored DP2; their preference was like
that of the English native speakers.
English RC attachment preferences of the
L2 learners were not associated with their
proficiency. This result is inconsistent
with Miyao and Omaki’s (2006)
developmental stages of L2 processing.
The factor shown to play a role in RC
attachment preferences of the L2 learners
in this study was WMC. This finding
provides evidence in support of the
‘chunking’ hypothesis, suggested by
Swets et al. (2007). Based on this
hypothesis, low span participants, not
having adequate resources, pause at the
boundary between the complex DP and
the RC and, in this way, chunk DP1 and
DP2 into a single unit, producing, as a
result, a DP1 attachment. In this study,
too, participants with high WMC may
have taken in longer chunks, without any
break at the boundary between the
complex DP and the RC and thus have
attached low. Since the material in this
study was not presented chunk by chunk,
participants had the opportunity to chunk
it themselves; thus, participants with
different WMCs could chunk the
sentences differently leading to different
attachment preferences.
This finding is also consistent with the
predictions of the Implicit Prosody
Hypothesis (Fodor, 2002) which predicts
that the prosodic grouping of the complex
DP and RC in a way that a pause is
inserted after the second DP -which is
proposed to be the case for participants
with low WMC- leads to a DP1
attachment preference.
