Professional Documents
Culture Documents
Use of Memory Load Interference in Processing Spoken Chinese Relative Clauses
Use of Memory Load Interference in Processing Spoken Chinese Relative Clauses
Use of Memory Load Interference in Processing Spoken Chinese Relative Clauses
Introduction
The working memory system plays an essential and fundamental role in sentence compre-
hension. An effective strategy to process and understand a sentence requires the listener
or the reader to assign syntactic relations to words or sequences of words by performing
* Tuyuan Cheng
2012tuyuan@gmail.com; Tycheng@ntin.edu.tw
1
General Education Center, National Tainan Junior College of Nursing, Tainan, Taiwan
2
Department of Psychology, National Taiwan University, Taipei, Taiwan
3
Graduate Institute of Linguistics, National Taiwan University, Taipei, Taiwan
13
Author's personal copy
1036 J Psycholinguist Res (2018) 47:1035–1055
computational operations on items retained for short periods of time in memory. Sen-
tences containing a relative clause (RC) component are known to cause difficulty in com-
prehension due to the complexity of syntactic structures and thus have been widely used
to investigate how working memory is involved and constrains sentence processing and
comprehension.
Research on comprehension of relative clauses has also carried implications for theories
of human language parsing. The processing difficulty profile for relative clauses, especially
RCs in head-initial languages, has been a long-standing issue in the sentence processing lit-
erature. The issue that has been raised is how and why two RC structures, namely subject-
gapped relative clauses (SRCs) and object-gapped relative clauses (ORCs) showed asym-
metrical processing and comprehension difficulty. In head-initial languages like English,
current consensus holds that SRCs, as in (1a), are easier to process than ORCs, as in (1b).
(1) a. The senator1 [that (e1/Subject) attacked the reporter] admitted the error.
b. The senator1 [that the reporter attacked (e1/Object)] admitted the error.
These two types of RCs contrast with respect to the location/function of the gap, indexed
as e1, the empty position the head noun ‘senator’ (termed as filler) leaves, and serving as
different grammatical functions in each RC sentence. Comprehending and processing these
postnominal RCs requires one to integrate and resolve the structural dependency between
the filler and the gap.
Over the past decades, researchers began to turn their attention to the processing of RCs
in head-final languages like Japanese, Korean and Chinese, wherein the prenominal posi-
tioning of RC exhibits a gap-filler relationship, in opposition to the filler-gap serial relation
in head-initial languages. Such a distinction between prenominal and postnominal position-
ing of RCs provides a uniquely valuable domain for testing theories of human parsing. The
processing/comprehension difficulty profile for RCs in head-final languages represent a
challenge for theories of processing and comprehension, as varied results have yielded and
SRCs are not consistently reported easier to process than ORCs. For example, using the
traditional self-paced reading task, some reported an SRC preference (Japanese: Ishizuka
et al. 2003; Miyamoto and Nakamura 2003, 2013; Nakamura and Miyamoto 2003; Korean:
Kwon et al. 2004, 2006; Chinese: Lin 2006; Lin and Bever 2006a, b, 2007), while others
observed an ORC preference (Chinese: Chen et al. 2008; Gibson and Wu 2013; Hsiao and
Gibson 2003). These conflicting findings have also been replicated across several different
methodological paradigms including ERP (SRC advantage in Japanese: Ueno and Garnsey
2008; ORC advantage in Chinese: Packard et al. 2011), and eye tracking (SRC advantage
in Korean: Kwon et al. 2010; ORC advantage in Chinese: Sung et al. 2015, 2016).
While a number of competing accounts for why SRCs enjoy a processing advantage
over ORCs in head-initial languages have been advanced and, as a result, many studies
have largely moved their battlefield to the arena of head-final languages to test the compet-
ing theories, relatively few studies have related RC asymmetrical processing preference to
the auditory RC comprehension difficulty profile. Specifically, very few studies have exam-
ined the RC processing preference in light of working memory involvement. Moreover,
few studies have examined RC comprehension difficulty by integrating the results of both
online processing and post-online comprehension, despite the fact that it is the commonly
adopted approach to understand how readers or listeners go from processing to comprehen-
sion, and that the approach is widely employed in the investigation of comprehension defi-
cit (e.g., Kemper and Kliegl 2007).
13
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1037
This paper intends to fill in the research gap. In this report we describe an investiga-
tion of on-line auditory RC sentence processing using the incremental moment-by-moment
measurements from the self-paced listening technique (Ferreira et al. 1996) as our depend-
ent variables in explicitly defining the processes of Chinese RC speech comprehension in
particular and RC comprehension in general.
Furthermore, the present study attempts to look into the puzzles as to how and why
some relative clauses are easier to understand. A novel dual-task, memory-load interfer-
ence design was used to examine the auditory processing and comprehension of Chinese
RCs. The rationale is: As proposed by some memory-based models (Carpenter et al. 1994;
Just and Carpenter 1992), there is a general working memory resource pool, with linguistic
and arithmetic processing relying on a shared, limited pool of working memory resources.
If this single-resource memory theory holds, one may investigate how syntactic structures
are involved in processing and comprehension by examining the effects of a concurrent
memory load on sentence processing. Since a concurrent load is expected to interfere with
many aspects of sentence processing, we hypothesize that sentence comprehension diffi-
culty can be defined in terms of the effect of the interference load. This idea is based on
the suggestion that processing syntactically more complex sentences should show signifi-
cantly more deteriorating performance as compared with processing syntactically simpler
sentences under memory load conditions (Caplan and Waters 1999; Gordon et al. 2002;
Waters et al. 2003). We also hypothesize that the online processing advantage in the RC
sentence types under interfering conditions is expected to spill over to the post-online com-
prehension. Namely, the online processing profile in either SRCs or ORCs should have
certain effects on one’s post-online RC understanding.
The objectives of this study were twofold. On the one hand, we seek to re-examine the
issue of Chinese RC processing by focusing on whether and how memory load interferes
with listeners’ RC processing profile and what their comprehension performance is like
under differing memory load conditions. On the other hand, evidence of such an effect
of load on online and post-online syntactic processing offers us an innovative approach
to address the RC processing asymmetry. A fuller picture of how Chinese RCs are under-
stood in both processing and comprehension, with the working memory system involved, is
likely to emerge.
Unlike the postnominal RCs in English shown in (1), Chinese RCs precede their head
noun, forming a head-final construction and exhibiting a Gap-Filler relationship (i.e., the
‘relativized’ or ‘extracted’ element appears before the head noun it associates with). The
prenominal positioning of Chinese RCs is illustrated in (2), where de is a relativizer (REL),
the RC marker.
13
Author's personal copy
1038 J Psycholinguist Res (2018) 47:1035–1055
13
Table 1 Previous studies on Chinese relative clause (RC) processing with experience-based and memory-based models
Method Study Online measurement within sentence Post-sentence
measurement
Pre-DE: embedded N–V/V–N RC marker DE Post-DE: head noun/matrix verb
n.s., no significant difference; m.s., marginally significant; –, not reported; >, longer than or induce greater cost
*Better performance
1039
13
Author's personal copy
1040 J Psycholinguist Res (2018) 47:1035–1055
(Hsiao and Gibson 2003; Chen et al. 2008; Sung et al. 2016). For the DE region, only
Hsiao and Gibson (2003) observed the ORC processing advantage on this critical region.
It is also of some interest to note the dissociation between the online and post-online
comprehension performance as reported in these studies. As can be seen in Table 1, stud-
ies of memory-based accounts showed more of an ORC advantage on the part of post-
online comprehension, but exceptional cases can be found. Sung et al. (2016), for example,
detected marginal SRC preference of native Chinese speakers on their post-online data.
As for studies of experience-based accounts, some observed no difference between SRC
and ORC in post-sentence comprehension performance (e.g., Jäger et al. 2015), while oth-
ers reported neither an SRC nor an ORC advantage (e.g., Yang et al. 2010). In particular,
the study in which an ORC advantage in post-sentence comprehension was observed (e.g.,
subjects have significant lower accuracy in answering questions following SRCs,77%, than
those following ORCs, 87%) did not even consider these comprehension accuracies as a
reliable index of the subjects’ comprehension of RCs (Lin 2014:193). A potential problem
with these results arises: if subjects have solved the processing difficulty during the online
processing, why were convergent results hardly ever found in post-online comprehension?
A tentative conclusion we draw from these previous debates is this. The Chinese RC pro-
cessing profile, as a whole, is affected by multiple factors and hence processing differences
arise. Memory-based models focus on how the parser keeps the parsed elements in work-
ing memory before they fade away immediately, while the experience-based models focus
on how the parser generates expectations for what may come next.
To tease apart the two competing models, one possible approach is to test on the basis of
one model, and to investigate to what extent it can apply to or against the other model. Intu-
itively, working memory system involves in syntactic processing and thus relates to pro-
cessing and comprehension. If one shows more processing difficulty in a target sentence,
interference of memory load will inevitably worsen his or her processing on this target
sentence. In addition, if one shows more processing difficulty in a target sentence, he or she
will inferably have more difficulty in understanding it. Presumably, if there exists asym-
metrical processing difficulty in Chinese RCs, interfering RCs with a concurrent memory
task should make RC processing less inefficient. Processing syntactically more complex
RC sentences should show disproportionately worse pattern, as compared with the process-
ing of syntactically simpler RC sentences.
The present article reports data from such a study based on the assumption. A digital
interference task involved in a self-paced listening paradigm, with the purpose to explore
the effect of memory load interference on online RC processing and to investigate whether
and how it may constrain the RC post-online sentence comprehension.
Exploring the effect of memory load on on-line syntactic processing of auditory RC
sentences tests directly on the memory-based model and offers a potentially sensitive
approach to the questions of whether the RC processing difficulty profile interacts with the
working memory system.
The self-paced listening paradigm, relative to the self-paced reading task, has an advantage
in investigating the incremental and natural processing. For one thing, the classical word-
by-word self-paced reading task, not allowing subjects to look back, fails to reflect read-
ers’ back and forth eye movement for integration and comprehension in natural reading.
13
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1041
Secondly, the stricture that the subjects may not go back to process prior materials is in fact
closer to the natural listening comprehension. Although listening is more closely related
phylogenetically to language processing and language development, very few studies have
been done on the online auditory sentence processing, and still fewer studies, to the best
of our knowledge, have been undertaken to investigate Chinese RC processing in listening
modality.
Natural listening is highly compressive in processing, and is expected to be closely
related to the functioning of working memory. The self-paced listening paradigm, com-
bined with memory-load interference task, will allow us to measure the integration cost
of the incremental segmented input in different processing phases under differing loading
conditions.
Based on the suggestion that processing more syntactically complex sentences will
overload working memory more than simpler sentences (Caplan and Waters 1999), we
hypothesize that the concurrent working memory load conditions (e.g. arithmetic inter-
ference load) manipulated in the course of comprehension will have effects on sentence
processing as well as on comprehension. Processing syntactically more complex sentences
would require disproportionately greater listening times online and show deteriorating per-
formance in the post-online sentence comprehension.
Three major questions were raised and addressed as in what follows:
1. In the on-line measurement, how does the processing profile of SRC versus ORC look
like under varying conditions of memory load interference? Is there memory load effect
on SRC or ORC? Will listening times change as the memory load conditions vary? And
if that is the case, the processing of which RC type is likely to be more affected?
2. In the post-online comprehension, which RC type shows greater disruption with respect
to the interfering load measures on post-sentence comprehension, including compre-
hension-question response times, comprehension-question accuracy and digit-recall
accuracy?
3. How is the working memory system involved in Chinese RC syntactic processing?
How does the working memory constrain the incoming string of words in on-line and
post-online processing? To what extent the experience- and memory-based accounts
are applicable to the current results?
Method
Participants
Sixty-one college students (40 females and 21 males; 20–24 years–old) were recruited from
National Taipei University of Technology and National Tainan Junior College of Nursing
to participate in the study. The participants speak Mandarin Chinese as their native lan-
guage, have no history of neurological or psychiatric disease, and possess normal (or cor-
rected-to-normal) visual acuity and normal hearing by self-report.
A self-paced listening task, the Auditory Moving Window (AMW) task (Ferreira et al.
1996) was conducted. This technique is considered to be able to capture patterns of
13
Author's personal copy
1042 J Psycholinguist Res (2018) 47:1035–1055
resource allocation during on-line linguistic processing (Waters and Caplan 2001, 2002).
Following the AMW paradigm, the participants in the current study listened to sentences
that were digitized and divided into a series of individual word segments. The partici-
pants had to pace themselves through the sentences as quickly as possible by pressing the
spacebar on the keyboard interfaced with the computer’s operating system for the succes-
sive audio presentation of each segment of the sentence. It is presumed that when listen-
ers require extra time to process information in the flow of an incoming sentence, they
will exhibit correspondingly longer pause durations before initiating subsequent segments.
Therefore, these pause durations between each spacebar press for the subsequent segment
will reflect the participants’ responsiveness to particular linguistic features in question.
We use this technique to capture patterns of resource allocation during on-line spoken
Chinese RC processing, and hold that, based on these patterns, the processing complexity
of different RCs can be inferred.
Under digit load conditions, the participants were required to remember the digits visually
presented before performing the AMW task. The participants first saw the Chinese word
for “ready” on the screen and pressed the spacebar on the keyboard to trigger the pres-
entation of the digits. The length of the digit presentation was determined using 1500 ms
equally for the three digit-load conditions (0-digit, 3-digit, and 5-digit), which was based
on Waters et al. (2003). In the case of the no-digit condition, the Chinese phrase for “no
digit” was shown on the screen. All the participants participated in the three digit-load con-
ditions over the course of two trial sessions, which were counterbalanced across trial items.
At the end of each auditory sentence presentation in the AMW task, the participants had to
repeat the digits they had seen at the beginning.
Stimuli
13
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1043
relations between the NP in the main clause or relative clause and the main verb. Half of
the answers to the statements were yes and half no. Sample experimental items glossed
in English in Stimuli A and B were presented in Table 3.
Design
The study used a 2 × 3 two-way factorial design. The participants were randomly
assigned to one of the two stimuli sets representing combinations of two within-sub-
ject factors of sentence type (SRC vs. ORC) and memory load (no load, 3-digit- load,
5-digit-load). Twelve partially counterbalanced lists were created such that each RC
experimental items appeared in only one condition in a list, and across lists every RC
item occurred in all conditions. The trial items were grouped into an initial warm-up
block with 10 practice trials, followed by eight experimental blocks each containing
four experimental RC items intervened with 8 filler items, forming 96 experimental tri-
als for each set.
Apparatus and Procedures
To prepare the materials for the AMW task, all the trial sentences were recorded by a
female speaker of native Mandarin Chinese and were digitized at a re-sampling rate of
40 kHz with a 16-bit quantization by Praat 5.3.43 (Boersma and Weenink 2013). The
waveform files were then edited (using Pratt) to place a marker at each word boundary.
Prior to defining the boundaries of the presentation for each segment, every waveform
file was inspected aurally for acceptability. Then the sentences were broken into seg-
ments according to the boundaries. Each of the trial sentence stimuli used in the current
study was divided into six segments based on the word boundary.
The experiment was controlled by a microcomputer, specifically an ASUS K42Jv
laptop. The experimental procedures and data collecting were conducted by the E-Prime
2.0 professional operating system, by which reaction times (RTs), the spacebar press
13
1044
13
Table 3 Sample experimental items
Digit memory- Spoken sentence (aurally presented) Comprehension question sentence (Visually presented) True(+)/
load set False(–)
5-digit/
Filler: The head of mafia instigated his followers to threaten the stores The head of mafia instigated the stores –
J Psycholinguist Res (2018) 47:1035–1055
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1045
Schematic Illustration
Repeat the digit
Comprehension
question
Time Segment 6
Segment 5
Segment 4
Segment 3
Segment 2
Segment 1
between each segment, and the timing of the visual (digit) and auditory (sentence) dis-
plays were all measured to the nearest millisecond. Figure 1 shows the schematic illus-
tration for the procedure.
At the very beginning, the participants saw the 1500 ms visual presentation under
one of the three digit conditions (0-digit, 3-digit, 5-digit). Following the presentation
of the digit, the self-paced listening task began. The RTs for each word segment were
measured. After they finished listening to the trial sentence, the participants saw a ques-
tion mark “?” on the screen, and then a yes/no comprehension question concerning the
information in the preceding sentence appeared. The participants heard a short “beep”
sound after they answered the question. Then a sentence appeared on the screen which
required the participants to repeat the digit they had seen at the beginning before listen-
ing to the sentence. The experimenter recorded the participants’ digit recall response
on a score sheet. In an attempt to reduce individual differences in tradeoffs between
attending to the sentence and the digit recall task, the participants were instructed to
try to keep remembering the digits correctly as their first priority while listening to the
sentence as quickly as possible, but they were reminded to avoid listening too quickly to
catch the meaning of the sentence.
Results
Online RT Results
The current study was primarily concerned with online syntactic processing at each word
segment and end-of-sentence performance by participants under three digit-load conditions
with two types of RC sentences. Filler sentences, with different characteristics from the
target RCs, were used to shield the purpose of the task from the participants, and thus to
possibly steer them away from developing expectations and strategies associated with the
processing of sentences similar to target RC structures. Thus, fillers were only included in
our statistical analysis to get a holistic picture but were excluded from the following report.
13
Author's personal copy
1046 J Psycholinguist Res (2018) 47:1035–1055
SRC/0-digit SRC/3-digit
RT SRC/5-digit ORC/0-digit
ORC/3-digit ORC/5-digit
1800
1700
1600
1500
1400
1300
1200
1100
1000
900
800
T1 T2 T3 T4 T5 T6
Word segment
Fig. 2 Mean RTs (in ms) per word segment (T1–T6) by digit load for SRCs/ORCs
Figure 2 shows the mean RTs under the three digit-load conditions at each word region
(from Segment 1: T1 to Segment 6: T6), with SRCs and ORCs displayed together for com-
parison. Response times (RTs) indicated the listening times (latency) from the onset of the
word to the spacebar press initiating the successive word, namely the duration of each word
segment. All the RTs for each word segment in the AMW task were the dependent measure
of the on-line sentence processing metric in all conditions, which were presumed to index
the RC complexity in question. For each word, RTs less than 200 ms (i.e., those reactions
were so prompt that they were likely to be made without thinking), or beyond 2.5 standard
deviations of the mean RTs of the condition to which the trials belonged and under which
the trials were responded to by each subject, were treated as outliers and excluded from
data analyses. The percentage of outliers was 6.25%. The RTs were analyzed only from
sentences for which the end-of-sentence comprehension task was responded to correctly.
One participant’s data were discarded because of operating system errors, yielding 60 sub-
jects in total.
We followed the mixed-effects models in analyzing the repeated measurement RT data
with subjects and items as crossed random effects provided by Baayen et al. (2008). First,
in order to verify that the simple model is justified, a likelihood ratio test was conducted
across the online regions measured and the post-online tasks conducted, as presented in
Table 4.
Secondly, all the measures obtained in the experiment, including the online RTs from
T1 to T6, the post-sentence comprehension online RT (T7), the post-sentence tasks of
comprehension accuracy (CACC), and the post-sentence digit recall accuracy (DACC),
were analyzed in a linear mixed-effects model using the lme4 package in Baayen et al.
(2008) and the lmerTest package in R (R Development Core Team 2007). The analyses
were conducted with RTs, CACC, and DACC as the dependent measures, subjects and
items entered as cross random factors, and the fixed factors were Sentence Code, Digit
Load, and Sentence Code × Digit Load interaction. All parameter estimates, as well as
13
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1047
p values, associated with the t tests for each effect in online processing, are listed in
Table 5.
Table 5 shows online measures analyzed with mixed-effects model by setting Sentence
code, Digit load and the Interaction as fixed effect. The modeling shows that significance
of sentence code was only observed on T3, the segment DE region, t = − 2.846, p < 0.01,
indicating that participants, compared to ORCs, listened to SRCs with shorter time at
DE region under 0-digit-load condition. Compared to 0-digit-load condition, participants
spent more processing time under 5-digit-load condition across most of the regions (T1
[t = 7.043, p < 0.001], T2 [t = 4.088, p < 0.001], T3 [t = 5.902, p < 0.001], T4 [t = 3.676,
p < 0.001], and T5 [t = 2.674, p < 0.001]), showing that higher digit load demands more
processing efforts. Interaction between sentence code and digit load is only observed on
T3, the DE region, where participants under SRC × 5-digit-load condition, processed sig-
nificantly longer than participants under ORC x 0-digit-load condition (t = 2.336, p < 0.05).
This indicates that higher digit load exerts greater influence on SRC processing.
In sum, Table 5 shows that the online RTs of SRCs and ORCs were affected across the
board by the 5-digit-load condition and that participants did not seem to show process-
ing variation on SRCs or ORCs with the individual digital load conditions (0-digit-load,
3-digit-load, and 5-digit-load). Under the 0-digit-load condition, i.e., free of interference,
participants only show critical processing difference in the region of Chinese relativizer
marker DE, and on this point, SRCs had the processing advantage.
The single-resource memory-based theory (e.g., Just and Carpenter 1992) claimed that
an external memory load and the load related to complexity of syntactic processing shared
the same resource pool. Researchers (Caplan and Waters 1999; Waters et al. 2003) sug-
gested that evidence supporting this position should come from interaction between sen-
tence type and (memory) load and that processing (listening) times at the most demanding
point of the syntactically more complex sentences should display more disproportionate
deterioration compared with syntactically simpler sentences as the memory load increases.
The current results, however, for the two compared pairs of RC sentences, did not
show the expected deteriorating time increase in either SRCs or ORCs. Rather, both the
13
1048
13
Table 5 Mixed-effects analyses on online RT data
Variables T1 Estimate Std. error t value Pr(> |t|) T2 Estimate Std. error t value Pr(> |t|) T3 Estimate Std. error t value Pr(> |t|)
(Intercept) 1225.673 42.478 28.86 <.001*** 1163.45 38.096 30.54 <.001*** 861.546 32.051 26.881 <.001
Sentence Code-SRC 54.108 29.661 1.824 0.068 10.283 24.794 0.415 0.678 − 53.95 18.96 − 2.846 0.004**
Digit load-3 93.777 29.661 3.162 0.002** 41.2 24.794 1.662 0.097 32.775 18.96 1.729 0.084
Digit load-5 208.904 29.661 7.043 <.001*** 101.358 24.794 4.088 <.001*** 111.894 18.96 5.902 <.001***
Sentence Code- − 47.656 41.948 − 1.136 0.256 4.133 35.064 0.118 0.906 26.738 26.813 0.997 0.319
SRC × Digit load-3
Sentence Code- − 10.665 41.948 − 0.254 0.799 52.896 35.064 1.509 0.132 62.64 26.813 2.336 0.020*
SRC × Digit load-5
Variables T4 Estimate Std. error t value Pr(> |t|) T5 Estimate Std. error t value Pr(> |t|) T6 Estimate Std. error t value Pr(> |t|)
(Intercept) 1165.006 46.458 25.08 <.001*** 1249.76 50.883 24.562 <.001 1566.68 69.454 22.557 <.001
Sentence Code-SRC − 2.035 32.268 − 0.063 0.95 − 37.888 34.74 − 1.091 0.276 − 69.215 49.99 − 1.385 0.166
Author's personal copy
Digit load-3 42.492 32.268 1.317 0.188 14.412 34.74 0.415 0.678 − 23.933 49.99 − 0.479 0.632
Digit load-5 118.61 32.268 3.676 <.001*** 92.887 34.74 2.674 0.008** 44.235 49.99 0.885 0.376
Sentence Code- 24.344 45.634 0.533 0.594 62.64 49.129 1.275 0.202 89.946 70.696 1.272 0.203
SRC × Digit load-3
Sentence Code- 69.046 45.634 1.513 0.13 40.842 49.129 0.831 0.406 68.835 70.696 0.974 0.33
SRC × Digit load-5
J Psycholinguist Res (2018) 47:1035–1055
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1049
processing profiles of SRCs and ORCs seemed to show deterioration in parallel with the
load increase.
To ensure that the listeners paid attention to the AMW task and to examine whether and
how the online sentence processing is carried out further into the post-online interpretive
process, we collected post-sentence measures, including participants’ RTs from the post-
sentence comprehension (T7), accuracy of answering questions concerning the preceding
sentences (CACC) and of the digit recall performance at the end of each trial (DACC).
The mean question comprehension accuracy and digit recall accuracy in percent as a
function of each digit load and sentence type condition are presented in Table 6. Fillers
were excluded from the statistical analyses reported here.
Table 7 presented the mixed-effects model analysis conducted on the post-sentence
comprehension RTs (T7), the post-sentence comprehension accuracy (CACC), and the
digit recall accuracy (DACC).
By setting Sentence code, Digit load and their interaction as fixed effect, the modeling
results showed significance on Sentence code-SRC, t = 3.100, p < 0.01, demonstrating that
listeners responded to post-sentence (T7) comprehension questions for SRCs with longer
RTs, compared to ORCs. Significance was observed on CACC for Sentence code-SRC,
t = − 3.073, p < 0.01, indicating that listeners responded to post-sentence (T7) comprehen-
sion questions for SRCs with lower accuracy, compared to ORCs.
For digit recall accuracy (DACC), listeners for SRCs showed a lower percentage of
correctness numerically compared with ORCs (see Table 6). Mixed- effects model anal-
yses observed Digit load effect on 3-digit-load (t = − 7.580, p < 0.001), and 5-digit-load
(t = − 6.398, p < 0.001), showing that listeners have lower accuracy in recalling 3 digits or
5 digits, compared to 0 digit. The interaction of sentence code x digit load was significant
for SRC code x 5-Digit-load, t = − 2.506, p < 0.05, indicating that participants have lower
digit recall accuracy while listening to SRCs under 5-digit-load condition, compared to
ORC × 0-digit-load condition.
13
1050
13
Table 7 Mixed-effects analyses on post-online RT and post-online task data (CACC and DACC)
Variables T7 Estimate Std. error t value Pr(> |t|) CACC Estimate Std. error t value Pr(> |t|) DACC Estimate Std. error t value Pr(> |t|)
(Intercept) 2661.044 144.43 18.42 <.001 0.908 0.024 38.48 <.001 0.983 0.017 56.734 <.001
Sentence Code-SRC 315.535 101.77 3.1 0.002** − 0.067 0.022 − 3.073 0.002** − 0.015 0.021 − 0.689 0.491
Digit load-3 0.467 101.77 0.005 0.996 − 0.015 0.022 − 0.672 0.502 − 0.16 0.021 − 7.58 <.001***
Digit load-5 − 116.458 101.77 − 1.144 0.253 − 0.04 0.022 − 1.824 0.068 − 0.135 0.021 − 6.398 <.001***
Sentence Code- − 97.571 143.93 − 0.678 0.498 − 0.015 0.031 − 0.475 0.635 − 0.019 0.03 − 0.626 0.531
SRC × Digit
load-3
Sentence Code- 171.173 143.93 1.189 0.234 0.015 0.031 0.475 0.635 − 0.075 0.03 − 2.506 0.012*
SRC × Digit
Author's personal copy
load-5
Discussion
Our AMW self-paced listening paradigm, with the concurrent memory load interference
task, is designed to serve a two-fold purpose: (1) to compare the online listening times
and post-online task performance of two types of Chinese RCs, subject-gapped relative
clauses (SRCs) and object-gapped relative clauses (ORCs) under the concurrent interfer-
ence of digit-load, and (2) on the basis of these performance data, to deduce the processing
difficulty of SRCs vis-à-vis ORCs, and thereby to examine how Chinese RC sentences are
understood from processing to comprehension in listening modality.
We adopted the suggestion of Caplan and Waters (1999), Waters and Caplan (2001,
2002), also Waters et al. (2003) that the pattern of pause durations observed when listen-
ers pace themselves through recorded sentences is a valid test of resource allocation in
online sentence processing. According to the single-resource view (Carpenter et al. 1994;
Just and Carpenter 1992), if the generalized limited pool of memory resources is depleted,
as may be the current case in which allocating resources for digital and linguistic pro-
cessing requires greater effort, then language processing and comprehension will suffer.
Listeners’ patterns of pause durations are expected to display comparable responsiveness
to putative variations in RC syntactic structures. However, contrary to expectations, our
mixed-effects model analyses of on-line listening times showed that SRC and ORC pairs
displayed no significant difference across the three digit-load conditions. There was no sig-
nificant interaction between sentence code and 3-digit-load, nor did we observe the sen-
tence code × 5-digit-load interaction. Despite the overall no-difference, digit-load effect
was observed. Participants’ listening to RCs with 5-digit-load interference is significantly
longer than listening to RCs with 0-digit-load or 3-digit-load interference. Specifically,
listeners paused differentially longer at SRCs under 5-digit-load interference than ORCs
under 0-digit-load at the point of DE marker, resulting in the issue of critical DE region.
Consistent with the prediction of experience-based account, and thus in contrast to that
of memory-based account, we observed that listeners, under 0-digit-load, showed SRC
advantage at the DE-region. Nevertheless, the reversed ORC advantage emerged when the
load increased to 5-digit. This indicated that the extrinsic memory load interfered with RC
syntactic processing and that SRCs seemed to be affected more by greater digit load at
DE-region (T3). However, the current one-site interaction result on T3 that SRCs suffered
more from working memory resource reduction due to the extrinsic digit-load increase do
not suffice to lead to the inference that SRCs pose greater processing difficulty and thus to
some extent are syntactically more complex, as participants generally showed no differen-
tial patterns of AMW pause times in processing auditory SRCs and ORCs under various
concurrent memory-load interferences. In sum, Chinese RC processing asymmetry did not
clearly materialize in our online task.
A possible explanation for the indistinctiveness of SRC versus ORC result is the com-
pressive and competitive nature of extrinsic digit load, which may force participants to lis-
ten/process as soon as possible across all the digit load conditions, thus reducing the emer-
gence of potential processing asymmetries in RCs. Another favorable explanation put forth
for these results, and also serving as our effort in solving the debates on the Chinese RC
processing advantage in the literature, is that there could be no such distinctive and inher-
ent RC processing asymmetry existent in Chinese as in head-initial languages like English
that no significant processing difference between Chinese SRCs and ORCs can be detected
under memory load interference. If Chinese SRCs or ORCs do have differential processing
13
Author's personal copy
1052 J Psycholinguist Res (2018) 47:1035–1055
difficulties, the relative degrees of the complexity should be reflected and enhanced with
the increased memory load, as evidenced by studies from memory-based theories.
Further support for this speculation can be drawn on the dissociation results between
on-line and post-online data in our study. Previous studies on Chinese RCs did show the
dissociation between online and post-online performance, irrespective of which account
they are based on. Very few studies examined with sufficient care why there are differ-
ences between participants’ on-line and post-online behaviors. Our mixed-effects model
analyses demonstrated that while there were no significantly differential processing pat-
terns between SRCs and ORCs in online processing, ORCs did have the post-online com-
prehension advantage in general.
With respect to the end-of-sentence (T7) RTs and comprehension performance (CACC),
participants responded to the post-sentence comprehension questions for SRCs with sig-
nificantly longer pause times and significantly lower accuracy, as compared to ORCs. As
shown in Tables 6 and 7, participants, at T7, responded with longer pause times for SRCs
than ORCs, and showed significantly lower comprehension accuracy for SRCs across all
the three digit-load conditions.
Regarding post-online digit recall performance (DACC), Tables 6 and 7 showed that
participants are better with ORCs than SRCs in retrieving the digits across the three digit-
load conditions. A significant effect of sentence code was found. Mixed-effects model
analyses showed that SRCs had a significantly lower accuracy on digit recall under the
5-digit-load condition, suggesting that processing of SRCs is more affected by digit load
interference at the post-online phase.
These results, though not in line with the prediction of single-resource memory the-
ory, conform to the view of separate-sentence interpretation resource theory (Caplan and
Waters 1999) which posits that syntactic processing is so highly domain-specific that it
does not rely on the generalized resources and thus will be relatively unaffected by differ-
ences from overall working memory (Waters and Caplan 2001, 2002; Waters et al. 2003).
According to this view, comprehension difficulty would not arise directly from online syn-
tactic parsing, as it is highly practiced and specialized, but rather from task demands of the
post-online, or “post-interpretive” processes. Following this position, the post-online meas-
ures, showing ORC advantage, could be the post-interpretive processing carried out after
the online, or “interpretive”, processing for the meaning construction has been established.
On this post-online phase, listeners were demanded to organize and maintain the linguistic
materials to answer the comprehension questions and recall the digits. Yet, comprehension
difficulty at this stage does not represent the online processing difficulty.
For either single-resource or separate-resource theory, the critical hypothesis is the
same: processing syntactically more complex sentences should show worse performance
as compared with the processing of simpler sentences. The difference lies in the predic-
tion that the single-resource theory predicts a significant deteriorating pattern for the com-
plex structures, while the separate-resource theory does not predict the deterioration. If the
online parsing is conducted with a separate-sentence-interpretation resource mechanism,
the processing patterns of presumed resource allocation as a function of syntactic complex-
ity would remain stable and comparable for processing RC sentences of different levels of
complexity. Our online results accord with the proposal of separate-resource theory.
13
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1053
Conclusion
This study is an innovative and initial attempt to understand sentence processing and com-
prehension of spoken Chinese relative clauses, by examining the self-paced responses in
the flow of auditory RC sentences, simultaneously interfered by concurrent digit load. The
results sharpen our understanding of and insights into how working memory functions are
involved in the online processing and post-online comprehension. The basic idea adopted
from the memory-based account is that processing more complex sentences should show
more disruption than processing simpler structures when it is interfered with by memory
load, which conversely allowed us to detect the Chinese RC processing difficulty by manip-
ulating the memory load interference in online RC processing.
As working memory has been repeatedly demonstrated to be a key factor in online sen-
tence processing, our attempt in finding reliable effects of concurrent digit load to index
the processing metrics of RC sentence complexity turns out to be an alternative approach
to the puzzles of Chinese RC processing asymmetry. The results that online processing of
SRCs vis-à-vis ORCs did not show greater deterioration when interfered with by greater
digit load lead to our speculation that Chinese SRCs and ORCs might not have as much
varied processing asymmetry as observed in head-initial languages like English. Interfer-
ing the online sentence processing with increasing memory load in effect modulates and
mitigates the putative processing differences between SRCs and ORCs, which could be
intrinsically not so significant.
Although concurrent memory-load involved with RC processing asymmetry is the ini-
tial focus of this study, one should not ignore the intriguing finding of listeners’ single-
site SRC processing advantage on DE marker under no-digit load interference. Taken this
together with the online results that Chinese SRCs and ORCs displayed no-difference
and parallel patterns under the concurrent memory-load interference, current results did
not favor either memory-based or experience-based account. These results directed us to
reconsider the Chinese RC processing asymmetry debated in the literature. We suggest that
Chinese RCs may have no intrinsic processing asymmetry, in view of working memory
involvement. This suggestion points to the need for the competing memory- and experi-
ence-based accounts of language parsing to take into account the interference effects of
working memory in future research.
Acknowledgements An earlier version of this paper was presented at the 21st Annual Conference of the
International Association of Chinese Linguistics (IACL-21), June 7–9, Taipei, 2013. A travel grant was pro-
vided to the first author by National Science Council, NSC 101-2410-H-439-001. Special thanks go to the
anonymous reviewers of JOPR, whose constructive comments have led to significant improvements of our
manuscript in both substance and style.
Funding This study was funded by the Ministry of Science and Technology, R.O.C.
[NSC-101-2410-H-439-001].
Conflict of interest The authors declare that they have no conflict of interest.
13
Author's personal copy
1054 J Psycholinguist Res (2018) 47:1035–1055
References
Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Mixed-effects modeling with crossed random effects
for subjects and items. Journal of Memory and Language, 59, 390–412.
Boersma, P., & Weenink, D. (2013). Praat: Doing phonetics by computer (Version 5.3.43) [Computer pro-
gram]. http://www.praat.org/. Accessed March 27 2013.
Caplan, D., & Waters, G. S. (1999). Verbal working memory and sentence comprehension. Behavioral and
Brain Sciences, 22, 77–94.
Carpenter, P. A., Miyaki, A., & Just, M. A. (1994). Working memory constrains in comprehension: Evi-
dence from individual differences, aphasia, and aging. In M. Gernsbacher (Ed.), Handbook of psycho-
linguistics (pp. 1075–1122). San Diego, CA: Academic Press.
Chen, B., Ning, A., Bi, H., & Dunlap, S. (2008). Chinese subject-relative clauses are more difficult to pro-
cess than the object-relative clauses. Acta Psychologica, 129, 61–65.
Ferreira, F., Henderson, J., Anes, M., Weeks, P., & McFarlane, D. (1996). Effects of lexical frequency and
syntactic complexity in spoken language comprehension: Evidence from the auditory moving-window
technique. Journal of Experimental Psychology. Learning, Memory, and Cognition, 22, 324–335.
Gennari, S., & MacDonald, M. (2009). Linking production and comprehension processes: The case of rela-
tive clauses. Cognition, 58, 161–187.
Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76.
Gibson, E. (2000). The dependency locality theory: A distance-based theory of linguistic complexity. In Y.
Miyashita, A. Marantz, & W. O’Neil (Eds.), Image, language, brain (pp. 95–126). Cambridge, MA:
MIT Press.
Gibson, E., & Wu, H.-H. (2013). Processing Chinese relative clauses in context. Language and Cognitive
Processes, 28, 125–155.
Gordon, P. C., Hendrick, R., & Levine, W. H. (2002). Memory-load interference in syntactic processing.
Psychological Science, 13(5), 425–430.
Hale, J. (2001). A probabilistic earley parser as a psycholinguistic model. In Proceedings of NAACL (Vol. 2,
pp. 159–166). Association for Computational Linguistics.
Hale, J. (2003). The information conveyed by words in sentences. Journal of Psycholinguistic Research, 32,
101–123.
Hsiao, F., & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3–27.
Ishizuka, T., Nakatani, K., & Gibson, E. (2003). Relative clause extraction complexity in Japanese. In The
sixteenth annual CUNY conference on human sentence processing.
Jäger, L., Chen, Z., Li, Q., Lin, C.-J. C., & Vasishth, S. (2015). The subject-relative advantage in Chinese:
Evidence for expectation-based processing. Journal of Memory and Language, 79, 97–120.
Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in work-
ing memory. Psychological Review, 99, 122–149.
Kemper, S., & Kliegl, R. (Eds.). (2007). Constraints on language: Aging, grammar, and memory. Springer.
Kwon, N., Gordon, P. C., Lee, Y., Kluender, R., & Polinsky, M. (2010). Cognitive and linguistic factors
affecting subject/object asymmetry: An eye-tracking study of prenominal relative clauses in Korean.
Language, 86(3), 546–582.
Kwon, N., Polinsky, M., & Kluender, R. (2004). Processing of relative clause sentences in Korean. In Poster
presented at the Architectures and mechanisms for language processing (AMLaP), 2004, France.
Kwon, N., Polinsky, M., & Kluender, R. (2006). Subject preference in Korean. In Baumer, D., Montero, D.,
& Scanlon, M. (Eds.), Proceedings of the 25th West Coast conference on formal linguistics (WCCFL-
25) (pp. 1–14). Somerville, MA: Cascadilla Press.
Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106, 1126–1177.
Lin, C.-J. C. (2006). Grammar and parsing: A typological investigation of relative-clause processing. Ph. D.
dissertation, University of Arizona, Tucson.
Lin, C.-J. C. (2014). Effect of thematic order on the comprehension of Chinese relative clauses. Lingua,
140, 180–206.
Lin, C.-J. C., & Bever, T. G. (2006a). A universal account of relative-clause processing in typologically
diverse languages. In Abstract presented at the 14th annual conference of the international associa-
tion of Chinese linguistics and the 10th international symposium on Chinese languages and linguistics
joint meeting. Taipei: Academia Sinica.
Lin, C.-J. C., & Bever, T. G. (2006b). Subject preference in the processing of relative clauses in Chinese.
In Baumer, D., Montero, D., & Scanlon, M. (Eds.), Proceedings of the 25th West Coast conference on
formal linguistics (WCCFL-25) (pp. 254–260). Somerville, MA: Cascadilla Press.
13
Author's personal copy
J Psycholinguist Res (2018) 47:1035–1055 1055
Lin, C.-J. C., & Bever, T. G. (2007). Processing head-final relative clauses without garden-paths. In Paper
presented at the international conference on processing head-final structures. Rochester, NY: Roches-
ter Institute of Technology.
Miyamoto, E. T., & Nakamura, M. (2003). Subject/object asymmetries in the processing of relative clauses
in Japanese. In G. Garding & M. Tsujimura (Eds.), WCCFL 22 proceedings (pp. 342–355). Somer-
ville, MA: Cascadilla Press.
Miyamoto, E. T., & Nakamura, M. (2013). Unmet expectations in the comprehension of relative clauses in
Japanese. In Proceedings of the 35th annual meeting of the cognitive science society.
Nakamura, M., & Miyamoto, E. T. (2003). Subject/object gaps and the processing of relative clauses in
Japanese. In The sixteenth annual CUNY conference on human sentence processing.
Packard, J. L., Ye, Z., & Zhou, X. (2011). Filler-gap processing in Mandarin relative clauses: Evidence from
event-related potentials. In H. Yamashita, Y. Hirose, & J. L. Packard (Eds.), Processing and producing
head-final structures (pp. 219–240). Dordrecht: Springer.
R Development Core Team. (2007). R: A language and environment for statistical computing [software].
Vienna, Austria: R Foundation for Statistical Computing. Retrieved February 5, 2018 from http://
www.r-project.org/.
Sung, Y. T., Cha, J. H., Tu, J. Y., Wu, M. D., & Lin, W. C. (2015). Investigating the processing of rela-
tive clauses in Mandarin Chinese: evidence from eye-movement data. Journal of Psycholinguistic
Research, 45, 1089–1113.
Sung, Y. T., Tu, J. Y., Cha, J. H., & Wu, M. D. (2016). Processing preference toward object-extracted rela-
tive clauses in Mandarin Chinese by L1 and L2 speakers: An eye-tracking study. Frontiers in psychol-
ogy, 7, 1–14.
Ueno, M., & Garnsey, S. M. (2008). An ERP study of the processing of subject and object relative clauses in
Japanese. Language and Cognitive Processes, 23(5), 646–688.
Waters, G. S., & Caplan, D. (2001). Age, working memory, and on-line syntactic processing in sentence
comprehension. Psychology and Aging, 16, 128–144.
Waters, G. S., & Caplan, D. (2002). Working memory and online syntactic processing in Alzheimer’s dis-
ease: Studies with auditory moving window presentation. Journal of Gerontology: Psychological Sci-
ences, 57B, 298–311.
Waters, G., Caplan, D., & Yampolsky, S. (2003). On-line syntactic processing under concurrent memory
load. Psychonomic Bulletin & Review, 10(1), 88–95.
Yang, C. L., Perfetti, C. A., & Liu, Y. (2010). Sentence integration processes: An ERP study of Chinese sen-
tence comprehension with relative clauses. Brain and Language, 112(2), 85–100.
13