|UFDC Home||myUFDC Home | Help|
This item has the following downloads:
EFFECTS OF SEMANTIC ASSOCIATION ON THE RESOLUTION OF COORDINATION
A THESIS PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF ARTS
UNIVERSITY OF FLORIDA
O 2007 Christopher Barkley
To all those who nurtured my academic curiosity, interests, and pursuit of knowledge throughout
I would like to thank my committee, Dr. Edith Kaan and Dr. Wind Cowles, for their
helpful advice while conducting this research. I would also like to thank my classmates and lab
colleagues for their support and suggestions on this work.
TABLE OF CONTENTS
LIST OF TABLES .........__... ......._. ...............7....
FIGURE ........._.__........_. ...............8.....
AB S TRAC T ......_ ................. ............_........9
1 INTRODUCTION ................. ...............12.......... ......
Sentence Processing Models and Ambiguity Resolution ................ ......... ................12
Syntax-First versus Constraint-Based Models .............. ...............12....
"Good-Enough" Representations in Sentence Processing............... ...............2
Processing Coordination................. .... ..........2
Processing Coordination Ambiguities ...._.._.._ ......._._. ...._.._ ............2
Processing Coordinate Structures ................. ...............31........... ....
2 MATERIALS AND METHODS .............. ...............34....
The Present Experiment............... ...............3
Predictions .............. ...............36....
Questionnaire study .............. ...............37....
Participants .............. ...............37....
M material s ................. ...............3.. 8..............
Procedure ................. ...............39.................
Analy si s ................. ...............3.. 9..............
Results .................... ...............39.
Self-Paced Reading Study .............. ...............39....
Participants .............. ...............39....
M material s ................. ...............40.......... .....
Procedure ................. ...............41.................
Analy si s ................. ...............42.......... .....
3 RE SULT S .............. ...............43....
Comprehension Data .............. ...............43....
Accuracy ................. ...............43.......... ......
Reaction Time .............. ...............43....
Reading Time Data ................. ...............44........... ....
4 DI SCUS SSION ................. ...............47................
Comprehension Data .............. ...............47....
Reading Time Data ................. ...............50........... ....
Interaction of Parallelism and Ambiguity at the Post Disambiguating Verb
Preposition (Position 11)................ ... .. .. ... ..... ........5
Main Effect of Parallelism at the Post-Conjunction Noun (Position 9) ................... .......52
The Relationship between Comprehension and Reading Time Data ........._.._.. ........._.._....56
Syntax-First versus Constraint-Based Models............... ...............56
"Good-Enough" Representations in Sentence Processing" ................ .........................58
5 CONCLU SION................ ..............5
6 FUTURE WORK............... ...............60..
A EXPERIMENTAL MATERIALS................ ...............6
B FILLER SENTENCES .............. ...............65....
LIST OF REFERENCES ................. ...............66................
BIOGRAPHICAL SKETCH .............. ...............69....
LIST OF TABLES
1-1: mean RT and accuracy per condition ................ ...............45.............
1-2: mean (SD) reading times by condition............... ...............4
2-1: mean reading times at each position (by subj ects) ................ ...............46...........
Abstract of Thesis Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Master of Arts
EFFECTS OF SEMANTIC ASSOCIATION ON THE
RESOLUTION OF COORDINATION AMBIGUITIES
Chair: Edith Kaan
In constructions such as (1) the man is temporarily ambiguous: it can either serve as the
second NP of a coordinate obj ect of the first verb (NP-coordination) or the subj ect of the second
verb (S-coordination). Previous research has demonstrated that readers prefer an NP
coordination reading, and garden path at the second verb in the sentence. One account is that the
ambiguous NP is initially attached to the preceding NP to form a coordinate noun phrase forcing
readers to re-analyze when they reach the disambiguating verb, based on syntactic principles. A
recent study, examined effects of non-syntactic information on the resolution of coordination
ambiguities and demonstrated that poor thematic fit between the initial verb in the sentence and
the ambiguous noun phrase can reduce processing difficulty at the disambiguating region but not
completely eliminate it. In the present study, we investigated whether the effect of semantic
association between the relevant NPs can ease or eliminate difficulty at the disambiguating
region in these constructions. In addition, we investigated whether after disambiguation, readers
successfully re-analyze constructions such as (1).
(1) The girl screamed at the woman and the man screamed at the boy.
Thirty-six participants read seven trials each of the four conditions in (2) in a self-paced
reading study, intermixed with filler items. The experimental sentences were either ambiguous
(a, c) or unambiguous (b, d), based on the conjunction, and the two subj ect NPs were either
associated (c, d) or un-associated (a, b) as determined by an independent paper-and-pencil rating.
Care was taken to avoid plausibility confounds by using only neutral, transitive verbs and
animate, human NPs in the experimental sentences. If readers indeed take information regarding
semantic association into account while parsing ambiguous sentences, then processing difficulty
should be reduced or eliminated at the disambiguating verb in the associated NP, ambiguous
We found a main effect of association at the noun following 'and', with the nouns
unassociated with the first subject NP (2a, b) taking longer to read than associated nouns (2c, d).
The direction of the main effect at the post-conjunction NP shows that this effect cannot be
attributed to different degrees of lexical priming between local NPs. We also found an
interaction between association and ambiguity at the post-verbal preposition. Reading times were
longest for the unassociated, ambiguous condition (2a) compared to the remaining conditions,
which did not differ. This suggests that the effect of semantic association can almost completely
override processing difficulty in the disambiguating region. Accuracy data illustrate that
participants may not always re-analyze these constructions completely, as accuracy was
significantly lower to the ambiguous sentences than the non-ambiguous items. Accuracy was
also significantly higher to associated than non-associated conditions, suggesting that semantic
information may ease re-analysis. These results suggest that readers are taking semantic factors
into account online, and using this information to guide parsing decisions, but do not always
successfully reanalyze or overwrite the initial interpretation.
a. Non-associated subj ect NPs, ambiguous
The enemy talked to the queen and the king talked to the people.
b Non-associated subject NPs, non-ambiguous
The enemy talked to the queen but the king talked to the people.
c. Associated subject NPs, ambiguous
The king talked to the enemy and the queen talked to the people.
d. Associated subject NPs, non-ambiguous
The king talked to the enemy but the queen talked to the people.
Sentence Processing Models and Ambiguity Resolution
Syntax-First versus Constraint-Based Models
Language comprehension takes place both rapidly and incrementally. Because a large
part of the comprehension of a sentence is based on its syntactic structure, comprehenders are
forced to make decisions of a structural nature in real time, although they are frequently faced
with syntactic indeterminacy and temporary ambiguity. Traditionally, there have been two
approaches to describing how the parser accomplishes the feat of interpretation under these less
than ideal circumstances, and these approaches have been contrasted extensively in the literature.
The first approach grants primacy to structurally based factors. According to these "syntax-first"
models, listeners/readers employ knowledge of the grammar of their language in order to build a
single structure into which incoming words are placed; without recourse to context, plausibility,
or any other non-structural source of information. Thus, during comprehension the
listener/reader adopts a single interpretation of the unfolding sentence based only on grammatical
category information and syntactic principles and only employs other sources of information in
order to evaluate the appropriateness of the structure and revise it if necessary. This view of
sentence parsing has been very influential and there are many versions of this approach (Ferreira
& Clifton, 1986; Frazier 1979), with perhaps the most influential being Lynn Frazier' s "garden
path model" (Frazier, 1979). In this model, the parser builds a first parse based on only the
grammatical categories of the incoming lexical items. When a sequence of grammatical
categories is consistent with more than one structure, the processor employs parsing principles to
distinguish between them. The principal of minimal attachment states that incoming materials
are attached into the phrase marker with the fewest syntactic nodes consistent with the well-
formedness rules of the language. If two analyses are equally simple, late closure dictates that
incoming information is attached to the constituent currently being processed; this ensures that
only one structure will be pursued. Based purely on grammatical category information and these
structural principles, the parser constructs its initial analysis of the sentence.
In opposition to these "syntax-first" models, constraint-based models (MacDonald
Perlmutter, & Seidenberg, 1994); assert that the process of building the interpretation of a
sentence involves the interaction of numerous factors in addition to syntactic considerations.
Plausibility, frequencies of occurrence and co-occurence, the relative weighting of grammatical
and probabilistic constraints, and animacy are just some of the sources of information that
contribute to the construction of an interpretation. As a sentence unfolds, all possible analyses
are computed in parallel by the parser and evaluated using multiple sources of evidence in the
input. Alternative structures are partially active with the activation level dependent on the
degree to which the input supports a given structure. The interpretation that receives the most
support from the input then "wins" out over its competitors, and the input is assigned this
structure. Processing difficulty arises when alternatives are closely matched in terms of the
support they receive from the input. Grammatical information may indeed be important to this
process and therefore may be a heavily-weighted source of evidence, but its temporal primacy is
removed and it takes no serial precedence over non-structural factors. The role of syntactic
factors is not the only crucial difference between the models, however. Syntax-first models
compute structures in serial order, and thus are often called "serial, depth-first" parsers; the first
parse is computed based on structural factors, and then revisions are computed serially based on
other relevant sources of information in the input. In contrast, constraint-based models assume
that all structures are computed in parallel, generating a type of competition which the "best"
structure wins. The contrast between these two models has motivated a great deal of research,
generating results that appear to strongly support these contrasting views. The present research
aims to add to the debate between these models by examining the effects that semantic
association has on the resolution of coordination ambiguities.
Ferreira & Clifton (1986) employed eye-tracking and a behavioral self-paced moving-
window technique to show that readers experienced difficulty when what was initially
understood as a subj ect plus main verb was disambiguated towards the dispreferred (non-
minimal attachment), less frequent reduced relative clause structure (la, 2a below) This
structure, depending on verbs that are morphologically ambiguous between past participle and
past tense forms, is known as the reduced relative ambiguity and has been studied extensively.
Ferreira and Clifton demonstrated that this processing disruption, the "garden path" effect, was
experienced even when the subj ect plus main verb analysis was highly unlikely because of an
inanimate subj ect in sentence-initial position, as in 2a). Thus, Ferreira and Clifton saw increased
reading times at the disambiguating region in both la and 2a (when compared to unambiguous
counterparts in b):
la) The defendant examined by the lawyer turned out to be unreliable.
lb) The defendant that was examined by the lawyer turned out to be unreliable.
2a) The evidence examined by the lawyer turned out to be unreliable.
2b) The evidence that was examined by the lawyer turned out to be unreliable.
As readers begin to construct their parse of sentences like 1) and 2) above, it is assumed
that they initially adopt a main-clause analysis and assign the role of subj ect and agent to the
defendantdd~~~~~ddddd~~~~ in 1) and the evidence in 2) (Based on the principle of minimal attachment under a
syntax-first model (Frazier, 1979); or based on the higher frequency of this structure under a
constraint-based approach (MacDonald et al. 1994). However, when readers reach the by-phrase
followed by the main verb, the parse under construction becomes inconsistent with the input and
the structure must be revised. This is the cause of the increase in reading time when compared to
the sentences in b) which are unambiguous between main verb and reduced-relative
interpretations because of the appearance of that which unambiguously signals a relative clause.
Crucially, if readers take animacy or plausibility into account to guide parsing decisions, then
the difficulty should be overridden or at least reduced in 2a) when compared to la). The fact that
Ferreira and Clifton observed similar patterns of disruption in la) and 2a) led them to argue for
an informati onally-encap sulated syntax module. They argue that thi s module operates on purely
syntactic principles and is insensitive to other sources of information, such as animacy. Thus,
their results support a syntax-first parser and are inconsistent with a constraint-based process in
which multiple structures are computed in parallel based on the contributions of all possible
sources of information.
Trueswell, Tanenhaus, and Garnsey (1994) influentially challenged the results of Ferreira
and Clifton (1986) by asserting that the implausible (inanimate) subj ects were selected for their
materials without extensive pre-testing. Trueswell et al. argued that many of the implausible
subjects were actually compatible with the main verb interpretation (such as the car towed....),
thus making the results of Ferreira and Clifton hard to interpret and rendering the data
inconclusive. Using paper-and-pencil completion norms, Trueswell et al. had participants rate
how typical animate and inanimate nouns were as agents and themes. Based on the results of
this study, Trueswell et al. constructed the same paradigm employed by Ferreira and Clifton
(1986), except with more highly controlled nouns in sentence-initial subj ect position, and then
employed the same experimental techniques in order to determine if Ferreira and Clifton's
results were replicable. This ensured that all animate subjects were good agents, and all
inanimate subj ects were poor agents and good themes. Trueswell et al. (1994) reported that, with
their modified materials, the difficulty observed at the disambiguating by-phrase of reduced
relative constructions with inanimate subjects disappeared. The authors state that their results
are most consistent with a evidential constraint-based model of sentence parsing in which
grammatical information takes no precedence over non-structural factors, and multiple sources of
information interact during structure-building operations.
Clifton, Traxler, Mohamed, Williams, Morris, & Rayner (2003) extended the research of
Ferreira & Clifton (1986) by employing an eye-tracking methodology to examine the resolution
of reduced-relative ambiguities. They again crossed ambiguity with the animacy of the subject
to form sentence quadruplets like 3) below:
3a) The man paid by the parents was unreasonable.
3b) The man who was paid by the parents was unreasonable.
3c) The ransom paid by the parents was unreasonable.
3d) The ransom that was paid by the parents was unreasonable.
Clifton et al. also manipulated parafoveal preview of the disambiguating by-phrase so
that each condition above was seen by some participants with the disambiguating by-phrase
visible in their parafovea and by others with the disambiguating by-phrase masked. This
manipulation was introduced in response to some criticisms of Ferreira and Clifton (1986);
namely that the grammatical information expressed by the by-phrase may have been used to
guide parsing decisions. By introducing this manipulation into their experimental design as a
factor, they were able to determine any possible effect of the participant seeing the by-phrase
while processing the ambiguous verb. Clifton et al. found evidence of disruption (in terms of
first-pass time, regression frequency and regression duration) at the disambiguating region (again
the by-phrase) in the reduced relative constructions, regardless of animacy or preview. Late
measures of processing seemed to indicate that animacy made the ambiguity especially hard to
overcome, with the reduced relative structures with animate subj ects proving hardest to revise.
These data support the notion that animacy cannot guide parsing decisions, but that it can serve
to ease re-analysis after misinterpretation. While these results initially seem to argue for a
syntax-first parser, Clifton et al. eloquently argue that this is a hasty conclusion. Clifton et al.
state that while the data do cast doubt on the notion that semantic and pragmatic factors can
override syntactic considerations, that this can not be straightforwardly concluded. They contend
that syntactic preferences may be overridden in some domains and not others, and that a
constraint-based theory could also explain their data by simply ranking constraints such that
structural simplicity outweighs semantic plausibility. These arguments actually extend to most
research designed to probe this theoretical distinction; with proponents of each model attributing
effects to different stages of the parsing process. For example, if it appears that animacy can
guide parsing decisions (e.g Trueswell, Tanenhaus, and Garnsey, 1994); then syntax-first
proponents can simply attribute these effects to "rapid re-analysis guided by semantic
principles." For this reason, it is notoriously difficult to interpret data gathered by experiments
designed to test syntax-first versus constraint-based models. The issues raised by Clifton et al.
should be considered when critically evaluating any research in this domain.
One issue that arises when critically evaluating these studies is the use of animacy as a
semantic factor. Trueswell (1994), Clifton et al. (2003), and Ferreira & Clifton (1986) all
employed animacy to argue either for or against semantic effects on parsing decisions. Animacy
is not the ideal factor for this purpose, as it is somewhat controversial as to whether it is a
completely "non-structural" factor. When considering what sort of information could influence a
restricted domain, or syntax-first parser, there was considerable debate about whether to include
thematic information such as animacy. Animacy is morphologically marked in some languages
with clear morpho-syntactic consequences, and in some languages, such as Japanese, structural
decisions on subject-hood are based on animacy information. It is also incorporated into a verb's
argument structure under some accounts, granting animacy structural consequences. The
semantic manipulation in the present experiment was designed to avoid this issue.
Trueswell (1996) examined the effect of frequency on the resolution of ambiguous
reduced relative clauses. The specific type of frequency that he was interested in was the
frequency with which a verb that is morphologically ambiguous between past tense and past
participle appears in each of these forms. Crucially, constraint-based models predict that
difficulty in assigning a reduced relative clause interpretation to a structure should be modulated
by the frequency with which a morphologically ambiguous verb form is employed in this
construction. Trueswell conducted two self-paced reading experiments in which readers read
relative clause ambiguities which were manipulated along three dimensions: whether the
sentence contained a reduced or unreduced relative clause, whether the subj ect was animate or
inanimate, and whether the main verb had a high or low frequency as a past participle. Results
of these experiments supported the notion that readers are subj ect to frequency effects when
building the first-parse of a sentence. Trueswell found that when the relative clause
interpretation received support from semantics (in the form of an inanimate subj ect), this
information served to override processing difficulty for verbs with high participle frequencies but
not those with low participle frequencies. When semantics provided support for the main clause
interpretation (in the form of an animate subj ect), there was evidence of processing difficulty for
both verbs, though ambiguity resolution appeared to be faster in the high participle frequency
group. Trueswell states that these results support a constraint-based model of sentence parsing,
in which verb-form frequency interacts with other structural and non-structural sources of
information during parsing. Trueswell argues that, based on this data, processing models should
take statistical frequency into account in addition to other non-structural factors.
Garnsey, Pearlmutter, Myers, & Lotocky (1997) examined the effects of verb bias and
plausibility on the resolution of sentences containing temporary ambiguities. Garnsey et al.
manipulated the plausibility that the post verbal NP would be interpreted as a direct obj ect of the
verb, the bias towards the transitive reading induced by the verb, and the presence or absence of
ambiguity. These manipulations resulted in the paradigm seen in 5):
Direct-obj ect (DO) biasing verbs:
Sa) The talented photographer accepted (that) the money could not be spent.
5b) The talented photographer accepted (that) the fire could not have been prevented.
Subordinate clause biasing verbs
Sc) The ticket agent admitted (that) the mistake had been careless and stupid.
5d) The ticket agent admitted (that) the airplane had been late in taking off.
Se) The sales clerk acknowledged (that) the error should have been detected earlier.
5f) The sales clerk acknowledged (that) the shirt should have been marked down.
Garnsey et al. found that verb bias has a rapid effect on ambiguity resolution, with
readers experiencing the most difficulty at the disambiguating region in sentences with DO
biasing verbs. Garnsey et al. also found an effect of plausibility, with implausible objects
reducing difficulty at the point of disambiguation, and this effect interacted with verb bias during
comprehension. The authors interpreted their results as best fitting into a constraint-based model
of sentence parsing, which takes into account non-structural factors (plausibility) and factors
generated by comprehender' s experience with transitivity/subordinate clause frequencies of
"Good-Enough" Representations in Sentence Processing
To this point, I have focused my description of sentence processing models on the factors
that may or may not interact to guide parsing decisions during comprehension. Verb bias,
plausibility, animacy, and frequency are among the sources of information that may possibly
influence structure building operations; and there is evidence pointing either to effects generated
by these factors or a lack thereof. A great deal more research is required to resolve this debate.
There is another debate, orthogonal to the debate between syntax-first and constraint-based
models, and it concerns the representational nature of the output of the parsing processes. An
implicit assumption of both syntax-first and constraint based models is that the end result of
parsing (with or without re-analysis) is a syntactic frame that is fully compatible with the input.
Regardless of whether the relative activation of this frame can be influenced by non-syntactic
information as the parser moves through the sentence, the interpretation of a sentence must be
based on a syntactic frame. The parser is traditionally viewed as a compositional algorithmic
processor that continues processing until the correct structure is built. The meaning of a sentence
is derived from the composition of the meaning of its parts; and the parser analyzes to the extent
that it can reach this interpretation. Recent research has begun to challenge this assumption by
illustrating that the end result of comprehension is often an incomplete or inaccurate
representation. One source of data that supports this view is the M~oses illusion (Erickson &
Mattson, 1981). When asked, "How many animals of each type did Moses bring on the ark?"
people often respond "two," without noticing the glaring error. While this evidence is frequently
employed to demonstrate the fallibility of memory, the degree to which listeners do not fully
process the input is striking. Barton and Sanford (1993) documented a similar effect, noticing
that people rarely notice the problem with the question, "Where did the authorities bury the
survivors?" While these sources of information are somewhat anecdotal, they do lend support to
the view that language processing is sometimes "shallow" and that compositionality does not
always hold. I will now summarize some research that lends support to this shallow, heuristic
Ferreira (1996) used simple, unambiguous sentences to investigate the issues raised
above. Ferreira examined this issue with the paradigm in 6), which manipulates bias,
plausibility, and grammatical voice. The sentence was either presented in active or passive
voice, and the arguments could be presented in one of two orders. In the reversible conditions,
reversing the order of the arguments led to a possible but implausible arrangement (as can be
seen in 6b). In the non-reversible condition, reversing the order of the arguments led to a
semantically anomalous arrangement, and in the symmetrical condition rearranging the
arguments had no effect on plausibility. The full paradigm can be seen in 6):
6a) biased, reversible, plausible
The dog bit the man. The dog was bitten by the man.
6b) biased, reversible, implausible
The man bit the dog The man was bitten by the dog.
6c) nonreversible, plausible
The mouse ate the cheese. The cheese was eaten by the mouse
6d) nonreversible, implausible
The cheese ate the mouse. The mouse was eaten by the cheese.
6e) symmetrical, one order
The woman visited the man. The man was visited by the woman.
6f) symmetrical, other order
The man visited the woman. The woman was visited by the man.
Sentences were presented to the participants aurally, and the participants were then
probed with comprehension questions that forced them to identify the thematic roles in the
sentence (i.e., DO-ER?). The results show that there is a cost for the passive, so that for all
sentence types, the participants made more errors on the questions probing the passive voice than
they did on the questions probing the active voice, and also responded to these sentences more
slowly. It appears that, even in relatively simple, unambiguous sentences, comprehenders
frequently misassign thematic roles leading to inaccurate representations. The authors conducted
a second experiment to ensure that it was not surface frequency of the construction generating
this effect, passive being less frequent than active, by comparing simple active sentences and
subject-cleft constructions. This was indeed not the case as subjects performed equally well to
questions probing actives and subj ect clefts, even though the subj ect-cleft construction is
extremely infrequent in English. A third experiment demonstrated a parallel between obj ect and
subj ect clefts and actives and passives, with participants performing poorly on questions probing
the obj ect clefts. The parallel between the obj ect-cleft construction and the passive is instructive.
Both of these constructions force comprehenders to assign thematic roles in an atypical order.
The fact that this atypical order of assignment is so problematic for readers points to a processor
that employs heuristic strategies (one such strategy, the NVN strategy, listeners readers to
identify the pre-verbal noun as a proto-agent and the post-verbal noun as the proto-patient)
caused rather than detailed algorithms; strategies which are prone to breakdown in atypical or
infrequent situations. Ferreira interprets these results as evidence of a heuristic processor, she
states that comprehenders often rely on shallow processing that is simply "good enough," rather
than always relying on complex syntactic algorithms. In some cases, as in the experiment, non-
canonical or atypical structures cause the parser to assign thematic roles to the wrong syntactic
constituents. The result is an underspecified, incomplete representation that may never be
completely revised. Crucially, Ferreira argues that these results are inconsistent with any
processing model (Frazier, 1979; MacDonald et al. 1994) that assumes that the parser only
computes structures that are consistent with the syntactic information encoded in the input.
Whether non-structural information can influence the computation of the structure or not, both
types of model argue that the structure must be built based on a syntactic frame. Ferreira argues
against this position convincingly.
Christianson, Hollingworth, Halliwell, and Ferriera (2001) examined incomplete or
inaccurate representations by examining whether readers ever fully delete their initial incorrect
interpretation of a sentence after re-analysis. In order to examine this issue they employed
transitivity ambiguities (with no commas to enhance the effect) as seen in (7). The reader will
initially interpret the baby as the object of dressed. However, when they encounter the verb spit,
they will encounter difficulty as the parse under construction is inconsistent with the input. They
will then have to revise their interpretation; and the authors were interested in the fate of the
7) While Anna dressed the baby spit up.
Previous research has assumed that, after mis-analysis, readers fully repair the
representation to the extent that the parse becomes fully consistent with the input string. After
reading sentences such as 7), subjects were probed with questions such as DidAnna dress the
baby? and, Did the baby spit up? Questions such as these can only be answered correctly if,
after mis-analysis, the parser detaches the baby from the slot designating it as the obj ect of the
verb and re-attaches it to subject position of the second verb. Christianson et al. found that
subjects were almost 100% accurate in stating that the baby spit up. However, a surprising
number of errors were found in which subj ects took the baby to be the obj ect of dressed. This is
evidence that their Einal representation is not consistent with the input. This could be caused by a
failure to re-analyze and assign thematic roles to the correct constituents, or by interference
between the unsurpressed initial incorrect interpretation and the correct one. It also may provide
evidence that a comprehender' s Einal representation isn't always consistent with linguistic well-
formedness conditions, as some participants interpreted the baby as both the subj ect of spit up
and the obj ect of dressed, an interpretation that violates structural principles such as the theta-
criterion. Though as stated above, it could also be the case that the accuracy pattern described
may be caused by the incorrect analysis persisting in memory to the extent that it can compete
with the subsequent analysis and influence comprehension. In the second experiment, the
researchers compared garden-path sentences such as 7) with unambiguous controls. The results
suggest that readers often do not completely re-analyze a construction, or that if they do, the
incorrect analysis persists to the extent that it can influence comprehension. These results nicely
tie language comprehension to other areas of cognitive science, such as vision, where recent
Endings have shown that the visual system does not construct a true global image of a visual
scene leading to change blindness, which refers to the observation that many participants are
insensitive to changes in a visual scene as they are viewing it. These data suggest that the visual
system may also rely on representations that are not always fully detailed. It may be the case that
the heuristic strategies discussed by Ferreira (1996) and Christianson et al. (2001) are
advantageous to an organism who must interact with its environment rapidly, in which case these
types of strategies would be employed in all domains of cognition and decision making. Further
evidence for the persistence of the initial mis-analysis and its effects on comprehension comes
from a study by Van Gompel, Pickering, Pearson, & Jacob (2006). They had participants read
sentences including verbs that were ambiguous between transitive and intransitive structures; in
one condition the transitivity of the verb was disambiguated with a comma but it in the other
condition no disambiguating information was present leading subjects to initially interpret an
intransitive verb as transitive. This led participants to garden path and temporarily activate a
transitive reading of the verb in the condition with no comma. After having read the sentence,
participants were prompted to complete a sentence fragment out loud. Participants were more
likely to produce transitive completions after the ambiguous condition in which they had initially
misinterpreted the verb as transitive, indicating that the initial misinterpretation persisted to the
extent that it could influence production. The authors were unable to distinguish between the
different accounts that provide an explanation for this effect (persistence of a memory trace, lack
of re-analysis, or application of a heuristic).
In sum, Ferreira (1996), Christianson et al. (2001), and Van Gompel et al. (2006)
demonstrate that the parser often constructs incomplete representations that are not the result of
exhaustive analysis. Whether it is the result of applying a heuristic to the processing of simple
sentences, failing to completely re-analyze the input, or not suppressing the initial interpretation;
evidently the parser is often satisfied with constructing representations that are only "good
The research that I have presented lays out the issues surrounding the debate between
syntax-first and constraint based processing models, and between algorithmic and heuristic
processors. As we have seen, these debates center on what types of information (syntactic and
non-syntactic) can influence initial parsing decisions and whether the end result of this parse is a
linguistic representation that is fully specified at all levels.
The purpose of the current study is to collect data that help to resolve these issues
utilizing coordination ambiguities and a manipulation that varies which previous noun phrase in
a sentence an ambiguous noun phrase is semantically associated with. A coordination ambiguity
is a sentence like (8) below; in which the NP the man is ambiguous between being the second NP
in a coordinate obj ect of the main verb of the first clause and the subj ect of the main verb of the
second clause. The sentence is disambiguated at the verb talked, when it becomes clear that the
man is in fact the subject of a second clause.
8) The woman talked to the boy and the man2 talked to the girl.
Previous research on this construction has demonstrated the so called "NP coordination
preference" (Frazier, 1979; Kaan & Swaab, 2003; Hoeks, 2006). This causes readers to initially
interpret the ambiguous NP as part of a coordinate obj ect of the first verb, and therefore initially
adopt the wrong analysis of the sentence. This triggers processing difficulty and re-analysis at
the second verb, when the parser realizes that the input has been misanalyzed.
For instance, Kaan and Swaab (2003) provided evidence for the NP coordination preference
using an ERP technique. Part of their experimental manipulation compared coordinated
sentences with ungrammatical continuations versus continuations that were "dispreferred."
Importantly, the dispreferred condition is only dispreferred if participants initially adopt an
analysis where the house and the garage form a coordinated NP. For example:
9) The man was painting the house and the garage were already finished.
10) The man was painting the house and the garage is already finished.
Kaan and Swaab (2003) found that both dispreferred and ungrammatical continuations
elicited a LAN effect at the verb after the ambiguous NP, this ERP component indexes early
syntactic processing difficulty that may reflect local phrase structure violations induced by what
the parser perceives as word category errors (Friederici, Pfeiffer, and Hahne 1993; Neville,
Nicol, and Barss 1991, Hahne and Friederici 1999). Crucially, for the purposes of this paper,
these data suggest that the parser initially attached the ambiguous NP to the preceding NP to
form a complex obj ect. When it reaches the disambiguating verb, the parser experiences
difficulty and initially interprets the string as ungrammatical because the verb cannot be
incorporated into the phrase structure currently under construction. While these data do reveal
that the parser is prone to attaching the ambiguous NP to form an NP coordination, they have
nothing to say about what underlies this effect. I will quickly summarize two approaches to
explain this parsing preference: syntactic principles and frequency based accounts.
Following the syntactic principle of minimal attachment, Frazier (1979) predicts that
readers will initially attach the ambiguous noun phrase to a second obj ect slot of the first verb,
creating a coordinate noun phrase reading. This occurs because to create a full S-coordination
requires the creation of one extra S node to head the new clause, thus violating minimal
attachment ("create the fewest nodes possible") and causing the parser to choose the NP-
coordination analysis. After the reader adopts this structure, they will garden path at the second
verb and have to re-analyze their initial mis-interpretation, presumably with measurable
processing cost resulting. This is a syntax-based account of the NP coordination preference.
Hoeks (2006) adopted a frequency based approach in attempt to explain this preference.
He conducted a corpus study examining the prevalence of S- and NP-coordinations in the Dutch
newspaper TROUW. In a set of 1000 occurrences, NP coordination was by far the most
prevalent, making up 46% of the instances using the Dutch connective en (equivalent to and). S-
coordination only made up 10% of these instances. So, perhaps it is not necessarily a syntactic
principle that underlies the NP coordination preference, but rather just frequency of occurrence
which causes readers to adopt the most likely structure based on frequency.
While the underlying mechanism of the NP coordination preference may be unclear, it is
a robust finding in sentence processing experiments. Whether motivated by frequency or
syntactic principles, this preference causes readers to incorrectly assign an NP coordination
reading to S-coordination structures and garden path at the disambiguating verb. The purpose of
the current research is to attempt to modulate this attachment preference in coordination
ambiguities with semantic factors. By varying whether the ambiguous noun phrase is
semantically associated with the subj ect or the obj ect of the first verb, the present study aims to
determine whether this preference can be overridden and processing difficulty reduced. If this is
the case, processing disruption would be reduced at the disambiguating verb in sentences like
Before summarizing the present study, it is first necessary to discuss the previous
research that has been examined factors that can influence the resolution of coordination
ambiguities as well as the factors that influence the processing of the second conjunct of a
Processing Coordination Ambiguities
Hoeks (1997, 1999, 2002) has conducted a series of experiments examining factors that
have an effect on the online resolution of coordination ambiguities. Before examining these
factors, Hoeks (1997) employed eye-tracking and ERP methodologies to examine the NP-
conjunction preference predicted by minimal attachment (Frazier, 1979). The data from these
parallel experiments suggest that readers do indeed demonstrate this processing preference, as
ERP and eye-tracking indices of parsing disruption were evident at and immediately following
the disambiguating region of the coordination ambiguity. Next, Hoeks examined whether a
semantic manipulation could eliminate processing difficulty by overriding the NP-conjunction
preference. He chose to investigate this issue by manipulating the thematic fit between the verb
and the ambiguous noun phrase, leading to the experimental manipulation seen in 9):
9a) The boatman repaired the mainsail and the skipper painted the mast of the boat.
9b) The sheriff saw the Indian and the cowboy noticed the horse in the bushes.
If the parser computes thematic fit online and uses this information to guide parsing
decisions, then the processing difficulty encountered at the underlined region of 9b) should be
reduced at the underlined region of 9a). In 9a), the NP the skipper is not a good argument of the
verb repair because it is animate, and Hoeks (1997) was interested in whether this type of
information was taken into account when parsing these constructions. He found that all effects
of processing difficulty in the disambiguating region disappeared as a result of this manipulation,
and used these data to argue against syntax-first models. While he acknowledges the argument
that re-analysis may have proceeded very quickly in these cases, indicating that the ambiguous
construction was still mis-analyzed; he argued that if this were the case, then the ERPs would
have shown signs of residual processing difficulty, which they did not (however see Hoeks
(2006), in which he argues that signs of residual difficulty are in fact present). One issue that is
evident in regards to these experimental materials is that perhaps the conjunction of animate and
inanimate NPs is so unnatural that the readers did not consider the NP-conjunction analysis.
This argument is backed up by an examination of language corpora which demonstrate that that
coordination of this type very rarely occurs in natural language (Hoeks, 2006).
Hoeks (2002) examined the effect that discourse structure in the preceding context has
on the resolution of coordination ambiguities. He argued that the NP conjunction preference
demonstrated by himself and other researchers is motivated by the fact that this interpretation is
simpler in terms of discourse structure. Hoeks first conducted a completion study which
illustrated the NP-coordination preference. He found that when completing fragments such as
The photographer spoke 0I ithr the model and the actress...., only 15% of completions represented
S-coordination. However, when the prior context introduced two subjects (see 10a), the number
of S-coordination completions rose to 80%. He then examined these issues with eye-tracking
and self-paced reading techniques. He had participants read ambiguous S-coordination
sentences, that were either preceded by neutral contexts or contexts that introduced two NPs in
coordinate subject position into the discourse. The paradigm is seen in 10)
10a), biasing context:
When they met the designer at the party afterward, the model and the photographer were very
10a'), neutral context:
It was not surprising that the party after the fashion show was exhilarating.
10b), target sentence:
The model embraced the designer and the photographer laughed.
Hoeks found that readers encountered difficulty at the disambiguating region of the
coordination ambiguities when they were preceded by neutral contexts. When the ambiguous
target sentence was preceded by a sentence introducing two subj ects, all traces of processing
difficulty disappeared. He used this data to argue for the immediate effect of context on the
resolution of coordination ambiguities and against models of sentence processing that give
precedence to structural sources of information.
Hoeks (1997, 1999, 2002) documented that thematic fit and discourse structure have
effects on the resolution of coordination ambiguities, and it is essentially his work that will be
extended in this paper. Let us first examine this issue from a slightly different perspective. Lyn
Frazier and colleagues have conducted a series of experiments examining what factors can
influence the processing of the second conjunct in a coordinate phrase, and I will now summarize
Processing Coordinate Structures
In earlier work, Frazier et al. (1984) conducted a series of self-paced reading experiments
examining the coordination of sentences. Frazier et al. were interested in whether they could
find an effect of structural parallelism (two constituents are said to be structurally parallel when
they share the same internal structure) when processing coordinated phrases. They investigated
this in many syntactic environments, such as active/passive, minimal/non-minimal attachment,
shifted/non-shifted heavy NPs, agent/theme, and animate vs. inanimate NPs. Across all
environments, Frazier et al. found an effect of parallelism. Therefore an active construction was
read more quickly when it was preceded by an active than when it was preceded by a passive, a
non-minimally attached structure was read more quickly when it was preceded by another non-
minimally attached structure than when it was preceded by a minimally attached structure, etc.
Frazier et al. used these data to argue that coordination constitutes a special syntactic case, but
does not and cannot conclude what the cause of the parallelism effect is. Nonetheless, this
research demonstrates that parallelism is a factor that can affect the processing of coordinate
Frazier, Munn, and Clifton (2000) conducted as series of four experiments examining the
processing of coordinate structures. They utilized an eye-tracking methodology to investigate
factors that influenced the processing of these structures. They found that coordination of unlike
categories was just as acceptable as processing two conjuncts of the same category. They used
this data to argue against a grammatical constraint that only constituents of the same category
can be conj oined. They did, however, find processing facilitation (in terms of reading time) at
the second of two conjuncts when it was either parallel in terms of category or internal structure.
For example the underlined portion of (9) was read faster than the underlined portion of (10), due
the fact that it is preceded by a conjunct with the same internal structure.
9) Bill talked to the tall man and the short woman.
10) Bill talked to the man and the short woman.
Additionally, Frazier et al. (2000) found a similar effect of syntactic category, leading to
a reduction in reading time for a constituent that was preceded by a constituent of the same
category. This effect would lead to the underlined portion of 11) being read faster than the
underlined portion of 12):
11) John walked with great speed and with great care.
12) John walked slowly and with great care.
However, the same facilitation was not found between two noun phrases that served as
the subj ect and obj ect of a verb. The authors use these data to argue that coordination enj oys
some sort of special status, as effects were not seen when using the same NPs in different
constructions. The authors successfully demonstrate that internal structure of the conjuncts is
one factor that influences the processing of coordinate structures, and hint at the fact that
coordinate structures may be influenced by factors that do not necessarily influence other
structures. This echoes Clifton' s (2003) argument that factors may have an influence on parsing
decisions in some domains, but not others.
While Frazier's primary interest is in the processing facilitation induced by structural
parallelism in coordination structures, and the current experimental paradigm will manipulate
whether an ambiguous noun phrase is semantically associated with the subj ect or object of the
main verb, her work still provides crucial background to the present study. The present study
aims to elucidate the factors that can influence the interpretation of the ambiguous NP in
coordination ambiguities. Research into these factors can be conceptualized in two ways. First,
it is a continuation of Hoeks' work that has already documented the effect that discourse
structure and thematic information can have on the attachment of this noun phrase. Secondly,
the present experiment has much in common with the work of Frazier and colleagues. Crucially,
the ambiguous noun phrase is a second conjunct and its interpretation can be influenced by
characteristics of the first conjunct as documented by Frazier. If semantic parallelism has a
similar effect to structural parallelism between conjuncts, then the attachment of the second
conjunct should be influenced by which other NP in the sentence it is semantically parallel with.
So, the attachment of this ambiguous NP can be influenced in two ways: by factors that affect
ambiguity resolution and by virtue of its semantic parallelism with a preceding conjunct. While
these two accounts are difficult to tease apart, I will keep both in mind while analyzing the data
resulting from the present experiment.
MATERIALS AND IVETHODS
The Present Experiment
The present experiment aims to add to the two debates central to the sentence processing
literature as well as extending the results of Hoeks (1997, 1999,2002), and determining whether
lexical/semantic parallelism has an effect on the resolution of coordination ambiguities. While
Hoeks's results are intriguing, there are issues with his experimental design that I intend to
remedy. First, as he himself acknowledges, the data that he uses to argue that semantic
information can override structural preferences in coordination ambiguities are flawed. He
employs materials that have an animate NP and an inanimate NP surrounding the conjunction
and. Based on corpus studies, coordinate phrases of this type are extremely rare. Therefore, his
data may only indicate that structural preferences were overridden by frequency or plausibility
factors. While this is interesting in and of itself, it does not constitute evidence that structural
preferences have been overridden by semantic factors. Secondly, animacy itself is not purely
semantic. As I discussed earlier, in some languages animacy has morpho-syntactic consequences
and the possibility that it can have similar influences in English cannot be excluded. I aim to
remedy these issues with the paradigm seen in (11) which manipulates whether the ambiguous
noun-phrase is semantically associated with the subj ect of the first clause or the obj ect of the first
11a) The enemy talked to the king and the queen talked to the people.
Q: Did the enemy talk to the queen?
11Ib) The enemy talked to the king but the queen talked to the people.
Q: Did the enemy talk to the queen?
11c) The king talked to the enemy and the queen talked to the people.
Q: Did the king talk to the queen?
11d) The king talked to the enemy but the queen talked to the people.
Q: Did the king talk to the queen?
This paradigm will allow me to examine the same issues that are central to the work of
Hoeks, but has the following advantages: 1. All NPs are human and animate. This allows the
reader to create equally natural conjunctions of both noun phrases and sentences. 2. The verbs
are neutrally biasing. In combination with the selection of nouns, this leads to elimination of
plausibility confounds and eliminates differences in selectional restrictions or thematic fit
between nouns and verbs across conditions. The manipulation is unequivocally semantic in
nature, in contrast to Hoeks (1997). He manipulated the animacy of the ambiguous noun phrase,
while the paradigm in 11) manipulates which previous noun-phrase the ambiguous noun-phrase
is associated with. If indeed, information on semantics/lexical association is taken into account
to guide parsing decisions, then difficulty in the disambiguating region should be reduced when
the ambiguous NP is semantically associated with the subject of the first clause. If this reduction
in difficulty is observed, it would constitute compelling evidence against serial, depth-first
models of parsing as in no way does the above manipulation have syntactic or structural
ramifications. I also aim to extend the work of Ferreira (1996) and Christianson (2001), by
examining the representational nature of the result of the parsing process. The process of
resolving coordination ambiguities is similar to the process of resolving the transitivity
ambiguities in Christianson et al. (2001). I will examine accuracy data in an attempt to
determine whether full re-analysis is undertaken, if the end result of parsing is a fully-detailed
representation, if the semantic manipulation can have an effect on accuracy, and if the end result
of parsing is fully specified and consistent with linguistic well-formedness conditions.
The paradigm in 11) above crosses ambiguity (but versus and) with the semantic
manipulation that varies which prior noun phrase the post conjunction noun is associated with.
In the parallel conditions, the ambiguous noun phrase is highly associated with the subj ect of the
main verb, while in the non-parallel conditions the ambiguous noun phrase is associated with the
obj ect of the main verb. If the preference for NP coordination is replicated in this study, then
reading times will be elevated at the second verb in the ambiguous conditions when compared to
the non-ambiguous conditions. This increase in reading time reflects the fact that the non-
ambiguous conditions do not license the NP coordination reading which is licensed in, and
causes difficulty in, the ambiguous conditions. If there is no such preference, we will not see
differences between ambiguous and non-ambiguous conditions. If semantic information is taken
into account and helps guide parsing decisions then reading time at the second verb (and
downstream) in the parallel, ambiguous condition should be reduced when compared to the non-
parallel, ambiguous condition. This reduction may manifest itself as a main effect of parallelism
or an interaction between parallelism and ambiguity at the point of disambiguation and
subsequent positions. However, if this information is ignored during initial parsing decisions
then reading time will be comparable at this position. The reading time data gathered in the
ambiguous region and post-ambiguity will determine whether information on semantic
association is reducing processing difficulty in these constructions. If semantic information is
strong enough to trigger S-coordination; we would expect an increase in reading time for non-
parallel conditions when compared to parallel conditions. Accuracy data will also be collected
comparing ambiguous and non-ambiguous conditions on questions that probe the grammatical
role of the ambiguous noun phrase. For example, the sentence "The king talked to the enemy
and the queen talked to the people," would be probed with the question "Did the king talk to the
queen?" If participants fully recover from the ambiguity and accurately re-assign theta roles or
successfully suppress the misinterpretation, then there will be no difference between these
conditions. However, if participants do not fully recover from a misinterpretation of the
ambiguity, we will observe significantly lower accuracy to the ambiguous than non-ambiguous
constructions. These data will help determine whether participants undertake full re-analysis
when recovering from a coordination ambiguity.
In order to generate the experimental paradigm described above with adequate control, it
was first necessary to create noun sets that contained two associates and one non-associate. For
this purpose, a questionnaire study which was designed to collect association ratings on pairs of
nouns was conducted.
Fifty participants from the University of Florida population participated in the
experiment. Data from ten participants were excluded due to their responses on an
accompanying language questionnaire that showed that they were not monolingual native
speakers of English. Participants gave informed consent according to University of Florida IRB
procedures, and were not compensated for their participation.
For the purpose of selecting the nouns for the self-paced reading study, forty-two noun pairs
were chosen based on data from the University of South Florida association database
(http://w3.uusf edu/FreeAssociation/). Nouns were selected based on the following criteria: The
noun refers to a descriptive or occupational role that a human could fill, the noun is an animate
count noun, the noun has a noun with the above characteristics on its list of associates, and the
cue-to-target association strength between the two nouns is at least .20 as listed in the association
database. A cue-to-target strength of .20 indicates that when prompted with the cue, 20% of
participants responded with the target. The original forty-two nouns and their associates thus
made up forty-two pairs of associated nouns. Based on intuition, a third, unassociated, noun was
added to each pair yielding triplets. Crucially, this noun could not appear on the list of associates
of either of the other two nouns. By splitting the triplets into their constituent pairs (ABC &
AB, CB, e.g, king-queen; enemy-queen), eighty-four pairs of nouns were created; forty-two that
were associated and forty-two that were unassociated. These eighty-four noun pairs were
randomly intermixed with eighty-four filler pairs, yielding 168 pairs of nouns in total. These
distractors were designed to mask the experimental manipulation as well as to encourage the
participant to utilize the entire rating scale. For example, the participant could be asked to rate
the association between "donkey and gymnast," or "coal and tree." The noun pairs were pseudo-
randomized so that no experimental pairs followed each other, and the pages of the questionnaire
were presented randomly to eliminate the effect of presentation order on the participants' ratings.
The participants' task was to provide a rating of how "associated" the two nouns were by
circling a number between 1 ("completely unassociated") and 7 ("completely associated").
Participants were given instructions to answer based on their initial intuitions, and given
examples in order to familiarize them with the rating scale. The participants were also
encouraged to employ all parts of the rating scale. The questionnaire study lasted 15 minutes.
We first calculated mean ratings for all eighty-four experimental pairs (42 associated
pairs and 42 unassociated pairs). Then t-tests were conducted examining the difference between
mean ratings of associated and unassociated pairs across items and comparing the associated and
unassociated pairs by item.
The results of the t-test comparing the associated and unassociated conditions revealed
that overall there was a significant difference between the ratings [t(1,41)=20.3, p < .001). The
results of the individual t-tests between the constituent pairs of each triplet yielded twenty-eight
triplets whose rating difference was significant at an alpha level of .001 (mean associated rating
= 5.68, mean unassociated rating 3.25). These twenty-eight triplets were then employed to create
the experimental materials for the self-paced reading study.
Self-Paced Reading Study
Forty-three participants (26 female, 17 male) from the University of Florida population
participated in the experiment for course credit, and gave informed consent according to
University of Florida IRB procedures. All participants had no known speech or hearing disorders
as indicated by self report. The mean age of participants was 20.4 years (range 18 to 28). Data
from seven participants were excluded from analysis based on responses to the accompanying
the language questionnaire that showed that they were not monolingual native speakers of
English. Thus, the data from 36 participants was analyzed.
The twenty-eight noun triplets from the questionnaire study were inserted into the
experimental paradigm (11i), yielding twenty-eight sets of experimental items. Note that in the
critical region starting from the determiner after the conjunct, the four conditions are physically
Across the twenty-eight sets, four neutral verbs ('talk to', 'walk with', 'meet with', 'ask
about') were each repeated seven times. This was done to ensure that there were no differences
in thematic fit between noun and verb across items, and that the ambiguous noun in post-
conjunction position could serve equally well as the agent and the patient of the verb. Each
participant saw seven items from each experimental condition as well as fifty-six fillers. The
fillers were constructed with the following aims: to balance out the number of S and NP
coordinations across the entire set of experimental materials thus precluding the development of
implicit processing strategies by the participant, to vary the clause boundary so that it could not
be predicted purely by linear position, to ensure that all conjunctions not signal a clause
boundary, and to form as natural a set of materials as possible by introducing different verbs and
types of nouns.
Four item groups were created by dividing the twenty-eight experimental items into
groups of seven. Across these groups, the difference between associated and unassociated pairs
was held constant such that a comparison of these conditions yielded non-significant results. A
one-way ANOVA with the factor group was conducted, yielding a p-value of .878. This
confirmed that across item groups, there was no significant difference in the difference between
associated and unassociated pairs. Unfortunately, because of the manner in which the critical
nouns were selected and the limited number of creatable noun pairs with the desired
characteristics; we were unable to balance length and frequency across lists or item groups. We
instead focused on balancing association strength, which is critical to the experimental
The item groups were then distributed across four participant lists, and counterbalanced.
In this way each participant only saw each item in one condition, and all items were seen an
equal number of times in all conditions. Fillers and experimental items were pseudo-randomized
such that no two experimental items followed each other.
Participants were probed with comprehension questions after each sentence. These
questions were designed to probe the role assigned to the ambiguous noun phrase after the
conjunction in the experimental conditions, in order to determine whether the reader recovered
from the garden path. For example, the sentence, "the enemy talked to the king and the queen
talked to the people," would be probed with the sentence, "Did the enemy talk to the king?" The
remainder of the comprehension questions probed relationships between the other arguments in
the sentences. The complete set of experimental materials and some examples of fillers sentences
are attached in the appendix.
Participants were individually tested. Stimulus presentation and data acquisition was
controlled using E-Prime (Psychology Software Tools) running on a PC. The experiment
employed a self-paced reading, moving window paradigm. Before beginning the experiment,
each participant was familiarized with the task with instructions and a brief practice block of five
items. After we were certain that the participants understood the task, the experiment began and
proceeded as follows: The trial began with a series of dashes on the screen which represented
each word of the sentence (with the number of dashes corresponding to the number of letters in
each word), and the participant pressed the "enter" button on the button box for the first word
(presented in size 15, courier new; black text on white screen) to appear. Each successive press
of the "enter" button caused the current word to disappear and the next word in the sentence to
appear. This continued to the end of the sentence at which point the question related to the
previous sentence appeared on the screen with instructions to press the "Y: button for yes, and
the "N" button for no. Once the participant answered the question, a "press X for next" screen
appeared. Once the participant pressed the "X" button, the screen disappeared and the next trial
proceeded as described above.
Reaction time data from the remaining thirty-six participants were first trimmed using the
following procedures: Any data point exceeding 3000 msec was treated as missing during the
analysis, this procedure affected .007% of the data. Additionally, means were calculated at each
word position for each condition for each participant and any data point that exceeded the mean
plus or minus 2 standard deviations was replaced with the mean, affecting .05% of the data. A
repeated measures ANOVA was conducted on the participant (F l) and item means (F2) with the
factors ambiguity (2 levels) and parallelism (2 levels) as within subj ect (item) factors, and item
as a between item factor and list as a between subject factor. Question accuracy and RT data
were analyzed using a two-way ANOVA with factors ambiguity and parallelism.
Overall, questions probing the experimental materials were answered quite accurately.
Table 1 summarizes comprehension data on the four experimental conditions. The results of
the two-way ANOVA revealed a significant effect of ambiguity [Fl(1,35)=14.316, p < .001;
F2(1,27)=4.44, p < .05], with accuracy on questions probing ambiguous items significantly lower
(mean =89.3%) than accuracy on non-ambiguous items (mean= 93.7%). There was also an
effect of parallelism, that was significant by subj ects but not by items [F l (1, 3 5)=6. 171, p < .05;
F2 (1,27)=2.766, p=.108], with accuracy on questions probing parallel items higher (93.7%) than
on the non-parallel items (88.8%). Additionally, the interaction between ambiguity and
parallelism was significant [Fl (1,35)=36.594, p < .001; F2(1,27)=17.016, p < .001]. The pattern
of the interaction reveals that there is a large accuracy difference between the two non-parallel
conditions, while the mean accuracy to the parallel conditions is much closer. Post-hoc Tukey's
pairwise comparisons on accuracy by condition indicated that accuracy on questions probing the
non-parallel, ambiguous condition was significantly lower than for on all other experimental
conditions (p < .001 in all three comparisons).
Question response times pattern with the accuracy data. The results of the ANOVA
indicated that the interaction between parallelism and ambiguity was significant [Fl
(1,35)=13.367, p < .001; F2 (1,27)= 14.008, p < .001]. The pattern of the interaction reveals that
there is a large reaction time difference between the two non-parallel conditions, while the mean
reaction time to the parallel conditions is much closer There was a marginal main effect of
ambiguity [Fl1(1,3 5)=3.723, p=.062; F2 (1,27)=3.089, p=.090], though the means do illustrate
that participants took longer to answer questions probing ambiguous (2778 ms) than non-
ambiguous (2547 ms) items. There was no main effect of parallelism [Fl1(1,35)=1.347, p=.254,
F2 (1,27)=. 868, p=.360]. Mean reading times do however indicate that subj ects took longer to
answer questions probing non-parallel items (2742 ms) than parallel items (2578ms).
Reading Time Data
A GLM repeated measures ANOVA was performed on all experimental items with the
within-subj ect factors ambiguity (ambiguous or not) and parallelism (parallel or not), and the
between-subj ect factors list and item group. An analysis including only correct items did not
yield qualitatively different results, and will therefore not be discussed. Table 2 presents the
reading times at position 5 (the determiner of the pre-conjunction noun phrase) and each
subsequent position and Eigure 1 summarizes reading times at all positions in the experimental
The ANOVA revealed an effect of parallelism at the pre-conjunction noun that was
significant by items [F l(1,32)=2.98, p=.09; F2(1,24)=7.12, p< .05] and an effect at the post-
conjunction noun that was significant by subj ects but marginal by items [Fl1(1,32)=5.125, p< .05;
F2(1,24)=3.199, p=.086], with the non-parallel items taking longer to read than the parallel items
at both positions. There was also an interaction between ambiguity and parallelism at position
11, which was significant by subjects only [Fl(1,24)=4.575, p < .05; F2(1,32)=1.45, p=.24].
The pattern at this position reveals that the non-parallel ambiguous condition takes the longest to
SThe discrepancy between p-values in the F1 and F2 analyses can be explained by the manner in which the items
were controlled. Controlling for association between associated and non-associated pairs across lists and item
groups dictated that frequency and length could not be adequately controlled for. We believe that this led to the
differences in significance in the F1 and F2 analyses.
Table 1-1: mean RT and accuracy per condition
NonParallel, Amb NonParallel, NonAmb Parallel, Amb Parallel, NonAmb
Accuracy 81.1% (.39) 96.8% (. 12) 97.2% (. 15) 90.6 %(.29)
RT 3051 (1069) 2433 (785) 2505 (813) 2652 (1021)
Table 1-2: mean (SD) reading times by condition
the N conj. the N V P the N.
N-P, A 356(14) 510(41) 426(19) 364(11) 442(29) 421(19) 364(12) 346(11) 725(51)
N-P, N-A 370(15) 473(32) 411(16) 367(13) 404)414(22) 345(9) 337(10) 688(35)
P, A 363(24) 463(34) 446(22) 377(14) 417(19) 411(18) 346(11) 348(10) 702(51)
P, N-A 361(12) 459(32) 425(21) 383(18) 436(25) 393(14) 346(9) 353(10) 719(47)
read, while the parallel ambiguous condition and the two non-ambiguous conditions have
virtually identical means.
The analysis revealed no main effects of ambiguity at any position. Main effects of and
interactions with list and item group ranged from p=.018 to p=.969, and all effects at other
positions were non-significant.
-* -N-P A
- --o-- -N-P, N-A
---0--- P, N-A
Figure 2-1: mean reading times at each position (by subj ects)
The goal of the present study was to provide data that would help resolve two of the more
prevalent debates in the sentence processing literature. By employing a manipulation that varied
which previous noun phrase the ambiguous noun phrase in a coordination ambiguity was
semantically associated with, and probing the attachment of this ambiguous noun phrase with
comprehension questions; the present study was designed to provide data that distinguished
between constraint based versus syntax-first parsers, and between exhaustive, algorithmic
parsing and parsing that outputs merely "good-enough" representations. We hypothesized that if
semantic information is taken into account during parsing, reading time at the disambiguating
region in the non-parallel, ambiguous condition would be elevated in comparison with the
parallel, ambiguous condition; and that differences in accuracy to ambiguous and non-ambiguous
conditions would provide evidence for a parser that sometimes outputs "good-enough"
representations. Comprehension and reading time data will be discussed before they are placed
in the context of models of processing and representation.
The data reveal that subjects answered questions probing non-ambiguous items more
accurately than questions probing their ambiguous counterparts. Subj ects were also more
accurate when responding to questions in the parallel conditions than when responding to the
non-parallel conditions. These two factors also interacted significantly, with a large accuracy
difference between the two ambiguous conditions that was almost twice the difference between
the two non-ambiguous conditions.
The demonstration that participants respond more accurately to questions probing non-
ambiguous (93.7%) than questions probing ambiguous items (89.3%) supports the notion that
participants sometimes do not overwrite or suppress their initial interpretation of a sentence, even
when that interpretation is erroneous. Extending results of Ferreira et al. (2002) and Christianson
et al. (2001) with a different construction, these results appear to demonstrate that the initial
misinterpretation of a sentence can persist even in the face of disambiguating input. It appears
that in the ambiguous conditions, readers sometimes attach the ambiguous noun phrase as an
obj ect of the first verb (based on either minimal attachment or the higher frequency of the NP
coordinate structure). Even after the main verb of the second clause is encountered and it
becomes clear that the ambiguous noun phrase is actually a subj ect, this reading survives to the
extent that it can effect question answering by lowering accuracy. The present results cannot
distinguish between the different accounts of this effect (underrepresentation, a lack of re-
analysis, or persistence of the initial analysis), but they do seem to indicate that readers
sometimes rely on representations that are not the result of exhaustively analyzing the input until
the correct structure is built. The comparison of accuracy to parallel and non-parallel conditions
also reached significance, as did the interaction between ambiguity and parallelism. The
accuracy to parallel conditions (93.7%) was higher than accuracy to non-parallel items (88.8%).
These results provide evidence that purely semantic information can affect comprehension of a
sentence. The data on the accuracy to questions probing non-parallel ambiguous (81.1%) and
parallel, ambiguous trials (97.2%) are also very interesting. This large difference in accuracy to
syntactically identical ambiguous constructions may be because participants are relying on
semantics when answering these questions. In the case of the parallel ambiguous conditions,
semantic information supports the correct structure in which the associated nouns are subjects of
their own clauses. Accuracy is high in this condition. In the non-parallel conditions, the
semantic information supports the wrong structure in which the associated nouns are both
objects, and consequently accuracy is low in this condition. If this is the case, then it can provide
some insight into a heuristic processor that does not always decommit from its initial
interpretation and/or revise after mis-analysis. This type of processor may output a
representation that is structurally underspecified due to a lack of re-analysis, or it may output
multiple structures that represent all interpretations entertained during the parsing process
leading to interference between correct and incorrect structures. Regardless of the exact
mechanism leading to deficits in structural representation, it appears that the language system
can recruit other sources of information to mask these inadequacies. When probed on a sentence
of which it has an inadequate representation, the system resorts to semantic sources of
information. When these cues are present and supportive, they are enough to override the
difficulty engendered by an inadequate representation. This is evidenced by the high accuracy to
the parallel, ambiguous condition (97.2%). However, when these cues are not present or support
the wrong interpretation, the system is left to deal with its deficits in structural representation and
accuracy drops sharply (81.1%). The reaction time data gathered from the comprehension
questions supports the error patterns described above. While comparisons of the experimental
factors do not reach significance, the pattern of the means supports the above conclusions. The
non-parallel, ambiguous condition which yielded the lowest accuracy also yielded the highest
reaction time (3051 ms). This indicates that there was no speed-accuracy tradeoff for the
participants. The difference between parallel, ambiguous (2505 ms) and non-parallel, ambiguous
(3051) mirrors the large difference in accuracy to these conditions and against suggests that
semantics can be recruited to facilitate off-line comprehension. When semantic cues are present,
subjects respond quickly and accurately. When these cues are not present, subjects respond
slowly and less accurately.
Reading Time Data
The data reveal a main effect of parallelism at both the pre- and post-conjunction nouns,
with non-parallel items taking longer to read than parallel items at both positions. There was
also a significant interaction between parallelism and ambiguity at the post disambiguating verb
preposition which resulted in a three versus one pattern. The non-parallel, ambiguous condition
showed an increase in reading time at this position when compared to all other conditions.
Interaction of Parallelism and Ambiguity at the Post Disambiguating Verb Preposition
As our interest is most sharply focused in the disambiguating region of the sentence, let
us begin a discussion of the results there. The interaction between parallelism and ambiguity at
this position is perhaps the most intriguing finding in this study. While the interaction was only
significant by subjects and not by items (see footnote 1), this effect may provide evidence that
semantic information can override processing difficulty in a syntactically ambiguous
environment. At the preposition, the pattern of the means reveals that the parallel, ambiguous
condition patterns with the two non-ambiguous conditions. Relative to these conditions, the non-
parallel ambiguous condition shows an elevation in reading time of approximately 20 ms
reflecting processing difficulty. These data appear to demonstrate that semantic information can
eliminate the garden path effect at the disambiguating verb, as no evidence of processing
disruption is seen in the parallel, ambiguous condition. This pattern of reading times is best
explained with a constraint-based account of parsing that resolves ambiguity through the process
of structure competition. Constraint based models of sentence processing state that as the parser
incrementally interprets the input, multiple structures consistent with the input are built based on
all information currently available to the parser. These structures "compete," in that they
become relatively more or less active than their competitors as the sentence unfolds and more
information becomes available, and at the end of the sentence the structure with the highest
activation "wins." Crucially, this type of model predicts that when two analyses receive
approximately equal support from the input, severe competition will occur and processing
difficulty will result. In contrast, when one analysis receives virtually all of the support,
processing will be facilitated because of a lack of competition. The crucial sources of
information here are the NP-coordination preference, semantics, and processing biases that
develop throughout the experiment. Due to the variable contributions of these sources of
information at prior positions in the sentence, the parser reaches the disambiguating region with
structures that are competing to different degrees in the two ambiguous conditions. The differing
degree to which these structures compete can explain the data at the preposition, and explain why
the parallel, ambiguous condition patterns with the non-ambiguous conditions at this position.
As the parser processes the conjunction and the post-conjunction region, both S and NP
coordinate structures become active. Their relative levels of activation are influenced in this
region by semantics, general processing principles, and parsing preferences that develop
throughout the experiment. Different cues are available in each condition, and this leads to
variable levels of activation (and therefore competition) between the two competing structures
across the conditions. As the parser reaches the preposition in the non-parallel, ambiguous
condition the two structures are competing greatly due to the two licit structures being activated
to approximately the same extent. This causes an elevation in reading time when the parser
encounters the disambiguating verb and it becomes evident that the S-coordination reading is
correct. However, in the parallel ambiguous condition, the S-coordinate structure is already
highly active due to the contribution of semantics at previous positions in the sentence. This
eliminates processing difficulty in the disambiguating region as the correct structure is already
highly active. The observation that the parallel, ambiguous condition shows no increase in
reading time relative to the two non ambiguous constructions. suggests that the contribution of
semantics is enough to boost the activation of the S-coordination reading to the extent that it is
virtually identical to the activation of the S structure when it is unambiguously signaled by the
preposition but. These results are hard to reconcile with a syntax-first parser, even one that
claims that semantic information can be recruited rapidly to facilitate re-analysis. The sentences
are syntactically identical at this position, licensing the same two structures, so differences
among the conditions are unlikely to be based on syntactic considerations. Additionally, a claim
that a syntax-first parser is simply recruiting semantic information to re-analyze at this position
would be unfounded. If this were the case, then both ambiguous conditions would show an
elevation in reading times compared to the non-ambiguous conditions, although perhaps this
effect would be smaller for the parallel conditions, since semantic information could ease re-
analysis. This pattern is not present in the data, as only the non-parallel ambiguous condition
shows an elevation compared to the non-ambiguous items. We therefore believe that the patterns
at this position are generated by the variable contributions of semantics in a constraint-based
competition model that also calculates the probability of alternative structures based on general
parsing principles and processing strategies that are sensitive to the frequency of occurrence of
alternative structures in the materials set.
Main Effect of Parallelism at the Post-Conjunction Noun (Position 9)
The main effect of parallelism at this position is also interesting. The direction of the
effect, with the non-parallel items taking longer to read than their parallel counterparts,
eliminates the possibility that this is simple local priming between noun-phrases as the effect is
in the wrong direction. Here is the experimental paradigm again (with position 9 in bold),
provided for clarity of discussion:
The enemy talked to the king and the queen talked to the people.
The enemy talked to the king but the queen talked to the people.
The king talked to the enemy and the queen talked to the people.
The king talked to the enemy but the queen talked to the people.
The difference in reading time between the non-ambiguous conditions can again be
explained through a constraint based model of parsing. While competition is limited because the
conjunction but only licenses S-coordination, the contribution of semantics can still explain the
increase in reading time in the non-parallel, non-ambiguous condition compared to the parallel,
non-ambiguous condition. In the parallel condition, the semantics of the noun phrase support the
S-coordination that is highly activated, while in the non-parallel condition semantics provides
interference against the S-coordinate reading. This leads to a reduction of the activation of the S-
coordinate structure, and thus an increase in reading time to the non-parallel, non-ambiguous
The difference between the two ambiguous conditions at this position is perhaps more
interesting. One of the primary goals of the experiment was to determine if semantic association
could influence the attachment of this ambiguous noun phrase. It appears as if it can, and the
interaction of semantics, general parsing principles, and experiment specific strategies can
explain the differences between ambiguous conditions at this position. In contrast to the non-
ambiguous items above, when the parser encounters the conjunction and two structures become
active as and licenses both S and NP coordination. The initial pattern of activation should favor
the NP-coordinate structure, as it is this structure that is predicted by minimal attachment and the
observation that NP coordination enj oys a large advantage over S-coordination in terms of real-
world frequency of occurrence (Hoeks, 2006). As the parser reaches the post conjunction noun
in these constructions, the contributions of experiment specific biases and semantics generate a
different pattern of activation across the ambiguous conditions. This leads to a high degree of
competition in the non-parallel condition, and a considerable advantage for the S-coordinate
structure in the parallel condition. This variable level of competition can explain both the
elevation in reading time at this position for the non-parallel items as well as provide the reason
that these two conditions reach the disambiguating region with different levels of competition
between alternatives as described above. The contribution of semantics is clear; it supports S-
coordination in the parallel conditions and NP-coordination in the non-parallel conditions, but
the contribution of experiment-specific parsing principles must be elucidated. As stated earlier,
NP coordination is statistically much more frequent in natural language, this assertion is
supported by completion and corpora studies (Hoeks, 2006). However, following psycho-
linguistic research methods, the number of S and NP coordinations in the present study was
balanced to inhibit the development of processing strategies throughout the experiment based on
the probability of encountering one of the two alternative structures that and licenses. Based on
the data, it appears that this balancing had the opposite effect and actually induced the parser to
modify its attachment preferences. It appears that the parser is sensitive to the discrepancy
between the probability of encountering one of two alternate structures in the real world and the
same probability within the materials set. This discrepancy causes the parser to modify general
principles to lower processing difficulty across the board. This leads to an attenuation of the NP-
coordination preference, and enables semantics to have more of an effect in the post-conjunction
region. This idea is supported by an analysis that included "time" as a factor to determine
whether participants are reacting to the disambiguating region differently at the beginning and
end of the experiment. The experimental sentences that occurred during the first 25% of the
experiment show a robust garden path effect at the disambiguating verb in the two ambiguous
conditions indicating that the parser is initially creating NP coordinate structures when
encountering and, this effect averages 45 ms. However, this effect has been reduced to 7.5 ms in
experimental trials that occur during the last 25% of the experiment. While the interaction of
ambiguity and "time" does not reach significance at the verb due to the necessity of only
analyzing a subset of items [Fl1(1,3 5)=1.91, p=. 168], this pattern is suggestive that the parser
employed the frequency with which S and NP coordinate structures occur within the experiment
to modify the NP coordination preference and reduce processing difficulty as the experiment
progressed. This interpretation generates testable predictions. Namely, that by modulating the
frequency of alternative structures within an experimental set either towards or away from
preferences that have been established by the parser in response to real world frequencies;
varying results will be observed. If this prediction is proven true, it would be of great
methodological consequence. For instance, if a set of materials was created in which S
coordinations were much more frequent than their NP counterparts (opposite of the distribution
in natural language), one would predict garden path effects in the disambiguating region of the
NP coordinate structures. These results are difficult to interpret under a syntax-first view, as
semantic information appears to be having an effect before the disambiguating region triggers
The main effect of parallelism at the pre-conjunction noun is difficult to interpret. At this
position, non-parallel conditions take longer to read than parallel conditions. This effect is not
due to length or frequency differences between the nouns in the two conditions, and cannot be
explained through recourse to syntax-first or constraint based parsing models. One speculative
account is that the participants have more trouble constructing a mental model in the non-parallel
conditions due to the arrangement of the noun phrases. In this case, plausibility at the "real
world" level is the crucial factor. This account would have to be further pursued by conducting
norming studies on the experimental materials.
The Relationship between Comprehension and Reading Time Data
When interpreting the reading time and comprehension data simultaneously, a clear
picture emerges. It appears that the semantic manipulation in the experimental materials has an
effect on both online and omfine linguistic processing. In terms of online processing, we see
clear effects of the semantic manipulation at both the post-conjunction noun and the post-verbal
preposition. At these positions, the additional semantic information provided by the
manipulation facilitates processing and reduces the difficulty that is encountered in the
disambiguating region of the sentence by reducing the level of competition between alternative
structures. In terms of offine processing, this same semantic information increases accuracy and
decreases response time when answering questions. The experimental condition that generates
the most processing dimfculty online also causes the most dimfculty omfine, as subj ects spend
longer and answer less accurately than to any other condition.
Syntax-First versus Constraint-Based Models
The results of this experiment are most compatible with a constraint-based model of
sentence processing in which alternative structures compete based on evidence from the input. It
is very dimfcult to reconcile this data with a syntax-based parser that grants primacy to structural
factors. At the post-conjunction noun in the experimental materials there is a main effect of
parallelism. At this position non-parallel conditions show an increase in reading times compared
to parallel conditions (both ambiguous and non-ambiguous). Crucially, the non-parallel
ambiguous condition takes longer to read than the parallel, ambiguous condition. This effect is
dimfcult to explain without recourse to some non-structural factors. At this position in the
sentence, when comparing parallel to non-parallel ambiguous, the structures are equally
ambiguous. In both sentences at this position, both the S and NP coordination readings are
licensed by the rules of syntax. The only difference between these items is in the lexical
association between the lexical item currently being parsed and a lexical item somewhere else in
the sentence. What then causes the increase in reading time? We believe that this elevation is a
result of increased competition between alternatives receiving equal amounts of support in the
non-parallel ambiguous condition. An explanation based purely on syntax fails here, as does an
account in which a syntax-based parser recruits semantic information during re-analysis as the
disambiguating region has yet to be reached and re-analysis has therefore yet to be initiated. At
the post-disambiguating verb preposition, there is a significant interaction between parallelism
and ambiguity that we also believe to be evidence for constraint-based parsing in these
constructions. At the preposition, evidence of processing difficulty only appears in the non-
parallel, ambiguous condition. The parallel ambiguous condition does not show this disruption
and the mean reading time at this position is virtually identical to the means of the non-
ambiguous conditions. We feel that this lack of difficulty in the parallel ambiguous condition is
due to the heightened activation of the S reading over the NP reading due to the relative
contributions of constraints at the post-conjunction noun. Even proponents of a syntax-based
parser that recruits semantic information for rapid re-analysis would have trouble explaining this
effect at this position. Perhaps this is the case and the methodology employed lacked the
sensitivity to detect residual processing difficulty, but the reading time data appear to illustrate
that the semantic information provided by the prior context is enough to eliminate processing
difficulty entirely. For these reasons, we feel that our data are most consistent with a constraint
based parser that considers all available sources of information to create multiple structures in
parallel, with the structure that receives the most support from the input mapping on to the
meaning of the sentence.
"Good-Enough" Representations in Sentence Processing"
A secondary aim of this research was to discriminate between fully detailed and "good
enough" representations in sentence processing. While the data collected here is relatively
limited in this regard, it does point to a sentence processor that outputs representations that are
not always the result of exhaustively analyzing the input. Our accuracy data indicate that
participants do not always re-analyze their interpretation or suppress it to the extent that it cannot
influence comprehension. It appears that this initial mis-interpretation can persist to the extent
that it can affect comprehension. Interestingly, our data also support the notion that this type of
heuristic syntactic processing may be aided by semantic cues in a manner that masks its
representational deficits. When support from semantics was not present in the ambiguous
constructions, participants experienced a huge drop in accuracy when compared to when the
semantic cues were present. We feel that this type of processor may use these additional cues to
mask when structural processing is not complete and only "good enough," in the same way that
employing semantic cues help Broca's aphasics make structural decisions when their receptive
grammatical deficits prevent them from undertaking a complete analysis. The degree of fall-off
once these cues are taken away is not as dramatic in the current participants, but conceptually we
believe the recruitment of semantic cues to mask structural deficits may be the same. In
conclusion, our accuracy data are not conclusive in either direction, but do appear to support the
notion that comprehenders often rely on representations that are not fully detailed in terms of the
correct structure of the input either because of a lack of full re-analysis or an inability to suppress
the representation that resulted from an incorrect parse.
The reading time study described here was designed to investigate whether information
regarding semantic association would facilitate processing in the disambiguating region of a
coordination ambiguity and also whether the end result of parsing this construction is fully
specified. The answer to the first question appears to be yes. During the experiment, processing
difficulty appeared to be eliminated by the semantic manipulation to the extent that the parallel,
ambiguous condition patterned with the two non-ambiguous conditions at the post-
disambiguating verb preposition. There was also an effect of the semantic manipulation in the
pre-disambiguating region, providing further evidence for the influence of semantics on the first-
pass parse. In terms of the nature of the output of parsing these constructions, we observed
significant differences in accuracy to ambiguous and non-ambiguous constructions; indicating
that perhaps participants weren't completely re-analyzing the input or were failing to suppress
the initial incorrect parse. Evidence was also found that suggests that the parser may modify
attachment preferences in response to the discrepancy between the frequency of alternative
structures in the real world and the materials set.
These Eindings provide support for a constraint-based parser that takes semantic
information into account when building the parse, and that is also sensitive to the probability of
encountering possible alternate structures in the real world and in the context of an experiment.
The parser may output underspecified representations or fail to suppress representations that lead
to clashes in terms of thematic role assignment and deficits in comprehension.
The data presented here represent only an initial investigation into the effects that non-
structural sources of information can have on the attachment of the ambiguous noun phrase in a
coordination ambiguity. Future projects include investigating the manipulation described herein
with more sensitive experimental techniques, such as eye-tracking or Event Related Potentials
(ERPs). I also plan to investigate the effects that a fine grained plausibility manipulation or verb
differences at the discourse level may have on the parser' s preferences at this noun phrase. A
study currently under preparation will investigate the role that the relative frequency of NP and S
coordination within the materials set can have on the attachment of this noun phrase.
Conclusions drawn from the data generated by this study will be of great theoretical interest.
1. h eie tle t eetii a t pub tle t frmn
a. The resident talked to the electrician and the plumber talked to the foreman.
b. The elresident talked to the relietrican bu the plumber talked to the foreman.
c. The electrician talked to the resident and the plumber talked to the foreman.
a. The enemy talked to the queen and the king talked to the people.
b. The enemy talked to the queen but the king talked to the people.
c. The king talked to the enemy and the queen talked to the people.
d. The king talked to the enemy but the queen talked to the people.
3. Tegettle otewf n h ubn akdt h oas
a. The guest talked to the wife and the husband talked to the locals.
b. The gufes talked to the wifes but the husband talked to the locals.
c. The wife talked to the guest and the husband talked to the locals.
a. The client talked to the architect and the builder talked to the contractor.
b. The client talked to the architect but the builder talked to the contractor.
c. The architect talked to the client and the builder talked to the contractor.
d. The architect talked to the client but the builder talked to the contractor.
5. Tevci akdt h oiea n h iea akdt h aiy
a. The victim talked to the policeman and the fireman talked to the family.
b. The voictima talked to the poiciman bu the fireman talked to the family.
c. The policeman talked to the victim and the fireman talked to the family.
a. The parent talked to the teacher and the student talked to the classmate.
b. The parent talked to the teacher but the student talked to the classmate.
c. The teacher talked to the parent and the student talked to the classmate.
d. The teacher talked to the parent but the student talked to the classmate
7. Teauttle oteepr n h oietle otecas
a. The adult talked to the expert and the novice talked to the class.
b. The adulrt talked to the expert but the novice talked to the class.
c. The expert talked to the adult and the novice talked to the class.
a. The coach walked with the swimmer and the diver walked with the fan.
b. The coach walked with the swimmer but the diver walked with the fan.
c. The swimmer walked with the coach and the diver walked with the fan.
d. The swimmer walked with the coach but the diver walked with the fan.
9. Temse akdwt h adadtebte akdwt h is
a. The master walked with the maid and the butler walked with the kids.
b. The maste walked with the maid butn the butler walked with the kids.
c. The maid walked with the master and the butler walked with the kids.
a. The parents walked with the sister and the brother walked with the grandparents.
b. The parents walked with the sister but the brother walked with the grandparents.
c. The brother walked with the parents and the sister walked with the grandparents.
d. The brother walked with the parents but the sister walked with the grandparents.
a. The friend walked with the blonde and the brunette walked with the salesman.
b. The friend walked with the blonde but the brunette walked with the salesman.
c. The blonde walked with the friend and the brunette walked with the salesman.
d. The blonde walked with the friend and the brunette walked with the salesman.
a. The servant walked with the lady and the gentleman walked with the children.
b. The servant walked with the lady but the gentleman walked with the children.
c. The lady walked with the servant and the gentleman walked with the children.
d. The lady walked with the servant and the gentleman walked with the children.
a. The neighbor walked with the son and the daughter walked with the cop.
b. The neighbor walked with the son but the daughter walked with the cop.
c. The son walked with the neighbor and the daughter walked with the cop.
d. The son walked with the neighbor but the daughter walked with the cop.
a. The trainer walked with the pitcher and the catcher walked with the umpire.
b. The trainer walked with the pitcher but the catcher walked with the umpire.
c. The pitcher walked with the trainer and the catcher walked with the umpire.
d. The pitcher walked with the trainer but the catcher walked with the umpire.
a. The scout met with the actress and the actor met with the audience.
b. The scout met with the actress but the actor met with the audience.
c. The actor met with the scout and the actress met with the audience.
d. The actor met with the scout but the actress met with the audience.
a. The reporter met with the judge and the jury met with the criminal.
b. The reporter met with the judge but the jury met with the criminal.
c. The judge met with the reporter and the jury met with the criminal.
d. The judge met with the reporter but the jury met with the criminal.
a. The intern met with the producer and the director met with the extra.
b. The intern met with the producer but the director met with the extra.
c. The producer met with the intern and the director met with the extra.
d. The producer met with the intern and the director met with the extra.
a. The candidate met with the faculty and the staff met with the dean.
b. The candidate met with the faculty but the staff met with the dean.
c. The faculty met with the candidate and the staff met with the dean.
d. The faculty met with the candidate but the staff met with the dean.
a. The student met with the chemist and the biologist met with the technician.
b. The student met with the chemist but the biologist met with the technician.
c. The chemist met with the student and the biologist met with the technician.
d. The chemist met with the student but the biologist met with the technician.
a. The slave met with the prince and the princess met with the ruler.
b. The slave met with the prince but the princess met with the ruler.
c. The prince met with the slave and the princess met with the ruler
d. The prince met with the slave but the princess met with the ruler
a. The boss met with the receptionist and the secretary met with the representative.
b. The boss met with the receptionist but the secretary met with the representative.
c. The receptionist met with the boss and the secretary met with the representative.
d. The receptionist met with the boss but the secretary met with the representative.
a. The passenger asked about the co-pilot and the pilot asked about the stewardess.
b. The passenger asked about the co-pilot but the pilot asked about the stewardess.
c. The co-pilot asked about the passenger and the pilot asked about the stewardess.
d. The co-pilot asked about the passenger but the pilot asked about the stewardess.
a. The officer asked about the killer and the victim asked about the punishment.
b. The officer asked about the killer but the victim asked about the punishment.
c. The killer asked about the officer and the victim asked about the punishment.
d. The killer asked about the officer but the victim asked about the punishment.
a. The writer asked about the hero and the villain asked about the family.
b. The writer asked about the hero but the villain asked about the family.
c. The hero asked about the writer and the villain asked about the family.
d. The hero asked about the writer but the villain asked about the family.
a. The referee asked about the winner and the loser asked about the crowd.
b. The referee asked about the winner but the loser asked about the crowd.
c. The winner asked about the referee and the loser asked about the crowd.
d. The winner asked about the referee but the loser asked about the crowd.
a. The customer asked about the chef and the cook asked about the busboy.
b. The customer asked about the chef but the cook asked about the busboy.
c. The chef asked about the customer and the cook asked about the busboy.
d. The chef asked about the customer but the cook asked about the busboy.
a. The relative asked about the bride and the groom asked about the band.
b. The relative asked about the bride but the groom asked about the band.
c. The bride asked about the relative and the groom asked about the band.
d. The bride asked about the relative but the groom asked about the band.
a. The child asked about the minister and the preacher asked about the congregation.
b. The child asked about the minister but the preacher asked about the congregation.
c. The minister asked about the child and the preacher asked about the congregation.
d. The minister asked about the child but the preacher asked about the congregation.
1. The lion hunted for a zebra and a buffalo in the long grass.
2. The manager spoke with the assistant and the janitor during their lunch break.
Or NP coordination
3. Alex had to meet with Jason or Stephen to work on their new design.
4. The antelope was eaten by a leopard or a hyena during the night.
Since S coordination
5. Since Julie covered for Sara Whitney had to come in to work today.
6. Since a citizen shouted at the mayor the governor has been more careful.
Because S coordination
7. Because the scientist lied about the data the public did not know the truth.
8. Because the accountant mixed up the numbers the banker lost his new j ob.
LIST OF REFERENCES
Altmann. 1998. Ambiguity in Sentence Processing. TICS 2(4)146-152.
Barton S.B., and Sanford A.J. 1993. A case study of anomaly detection: Shallow
semantic processing and cohesion establishment. Memory and Cognition 21: 477-
Christianson, Hollingworth, Halliwell, & Ferreira. 2001. Thematic roles assigned along the
garden path linger. Cognitive Psychology 42.
Clifton, Charles, Traxler, Matthew J., Mohamed, Mohamed Taha, Williams, Rihana S., Morris,
Robin K., and Rayner, Keith. 2003. The use of thematic role information in parsing:
Syntactic processing autonomy revisited. Journal of2~emory and Language 49:3 17-3 34.
Erickson T.A., and Mattson M.E. From words to meaning: a semantic illusion. Journal of verbal
learning and verbal behavior 20: 540-552.
Ferreira, Fernanda. 2003. The misinterpretation of non-canonical sentences. Cognitive
Psychology 47: 164-203.
Ferreira, Fernanda, Bailey, K., & Ferraro, V. 2002. Good-Enough Representations in Language
Comprehension. Current directions in psychological science 11: 11-15.
Ferreira, Fernanda, & Clifton Jr., Charles. 1986. The independence of syntactic processing.
Journal of2~emory and Language 25:348-3 68.
Fodor, J.D., & Inoue, A. 1998. Attach anyway. In J.D Fodor & Fernanda Ferreira (Eds.)
Reanalysis in sentence processing (ppl 01-141). Dordrecht: Kluwer.
Frazier, Lyn. 1979. On comprehending sentences: Syntactic parsing strategies, Univeristy of
Connecticut: Ph.D. Thesis.
Frazier et al. 1984. Parallel structure: a source of facilitation in comprehension. Journal of
Memory and Cognition 12(5).
Frazier, L., Munn, A., & Clifton, C. 2000. Processing coordinate structures. Journal of
Psycholinguistic research 29(4): 343-366.
Friederici, A., Pfeiffer, E., & Hahne, A. 1993. Event-related brain potentials during natural
speech: processing effects of semantic, morphological, and syntactic violations.
Cognitive Brain Research 1(1): 183-192.
Garnsey, Susan M, Pearlmutter, Neal J, Myers, Elizabeth, & Lotocky, Melanie A. 1997. The
Contributions of Verb Bias and Plausibility to the Comprehension of Temporarily
Ambiguous Sentences. Journal of2~emory and Language 3 7:58-93.
Frazier, L and Clifton, C. 1981. Parsing coordinates and ellipis: copy alpha. Syntax 4(1): 1-22.
Hahne, A., and Friederici, A. 1999. Electrophysiological evidence for two steps in syntactic
analysis: early automatic and late controlled. Journal of Cognitive Neuroscience 11(2),
Hoeks, J.C.J. 1997. Processing coordinate structures. Proceedings of the CLS opening academic
year 1997, Nimegen, NL.
Hoeks, J.C.J. 1999. The processing of coordination: semantic and pragmatic
constraints on ambiguity resolution. Doctoral Dissertation. University of
Nijmegen, Nijmegen, The Netherlands.
Hoeks, J.C.J., Hendriks, P., Vonk, W., Brown, C.M., & Hagoort, P. 2006. Processing
the NP- versus S-coordination ambiguity. Thematic information does not
completely eliminate processing difficulty. Quarterly Journal ofExperintental
Psychology, 59(9), 1581-1599.
Hoeks, J.C.J., Vonk, W., & Schriefers, H. 2002. Processing coordination in context:
The effect of topic-structure on ambiguity resolution. Journal of~entory and Language,
Kaan E & Swaab TY 2003. Electrophysiological evidence for serial sentence
Processing: a comparison between dispreffered and ungrammatical continuations
Cognitive Brain Research. 17(3): 621-635.
MacDonald, M. C., Perlmutter, N. J., and Seidenberg, M. S., 1994. The lexical nature
of syntactic ambiguity resolution. PsychologicalReview, 101, 676-703.
Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. 1998. The University of South
Florida word association, rhyme, and word fragment norms.
http://www.u sf edu/F reeAs soci ati on/.
Neville, H., Nicol, J., and Barss, A., Foster, K., and Garret, M. 1991. Syntactically based
Sentence processing classes: evidence from event-related brain potentials.
Journal of Cognitive Neuroscience 3, 151-165.
Trueswell, J. 1996. The role of lexical frequency in syntactic ambiguity resolution.
Journal of~entory and Language, 35, 566-585.
Trueswell, John C., Tannenhaus, Michael K., & Garnsey, Susan M. 1994. Semantic
influence on parsing: use of thematic role information in syntactic ambiguity
resolution. Journal of~entory and Language 33:285-318.
Van Gompel, R., Pickering, M., Pearson, J., and Gunnar, J. 2006. The activation of
inappropriate analyses in garden path sentences: evidence from structural priming.
Journal of2~emory and Language, 55, 335-362.
Van Pettern, C., Weckerly, J., McIsaac, H., and Kutas, M. 1997. Working memory capacity
dissociates lexical and sentential context effects. Psychological Science, 8, 238
Christopher Barkley was born on January 22nd 1982 in Edinburgh, Scotland. After
spending four years in Scotland, Christopher moved to Ingrave, a small village just outside of
London, England. When his father was transferred to a banking position in America,
Christopher moved to Jacksonville, FL, in 1992. Chris attended elementary, middle, and high
school in Jacksonville until he graduated from the Paxon School for Advanced Studies in 2000.
After graduating high school, Chris enrolled at the University of Florida. He obtained his
bachelors degree with a maj or in linguistics in 2005 and his Master of Arts in 2007. Chris is
currently enrolled in the interdisciplinary program in Linguistics and Cognitive Science at the
University of California, San Diego.