Transcript Hagoort, P., Brown, C.M., Groothusen, J. (1993)

Hagoort, P., Brown, C.M., Groothusen, J.
(1993)
The Syntactic Positive Shift (SPS)
as an ERP measure of syntactic
processing
Introduction
The data in this study may prove useful
for resolving an ongoing debate in
parsing research about the nature of
processing of the parser: two approaches
Debate:


Autonomous approach: computation is based on
syntactic principles. Lexical, semantic and
pragmatic information have their influence only
after syntactic parse is delivered 
Separate level for the syntactic structure of a
sentence is required.
Debate:


Interactive approach: syntactic information is
directly integrated with lexical and
semantic/pragmatic information, i.e., non
syntactic sources are used either to direct the
parser initial analysis or to evaluate immediately
the outcome of syntactic analysis on a word-byword basis.
No separate level of representation is required.
How to resolve this debate?



There is no sufficient empirical evidence
Existing chronometric techniques may not pick
on the small effects of the parsing process
Re-analysis can be so fast, that it is difficult to
observe
ERP method - advantages



Empirical evidence from reaction times
Sensitivity of ERP’s to the representational level
(qualitatively different cognitive processes show
up in different waveforms)
If there are separate processing components
producing output at different level of
representation  there are distinct brain
correlates to semantic and syntactic processing
Previous studies dealing with the ERP
responses to syntactic violations



Kutas & Hillyard (1983): first study to
investigate ERP responses to syntactic errors
Osterhout & Holcomb (1992): ERP responses to
the syntactic violations (verb sub-categorization
and phrase structure constraints)
Neville et al. (1991): violations of constraints on
the movement of wh-phrases
Summary of the previous research



Syntactic violations do not show N400 effect
shown for semantic anomaly
Syntactic violations show P600 family
positivities
However, the results are not robust as the
studies have not resulted in the global ERP
index of parsing syntactic operations
Problems with the previous studies



The number of studies is too limited to draw
conclusions regarding the existence and nature
of syntactic ERP responses
Studies on parsing have been done on English
only
The question is whether specific syntactic
violations in English result in similar ERP effect
as the same syntactic violation in other
languages (i.e., Dutch)
Current study



Its goal is to investigate ERP manifestations of
syntactic parsing (electrophysiological response
that is qualitatively different from semantic
parsing) – Syntactic Positive Shift (SPS)
Although it may not provide the answer, it
opens the way for a novel empirical
investigation
Syntactic processing is studied based on the
following three syntactic violations in Dutch:
Choice of syntactic violation



Should help optimize the likelihood that the
subjects detect the violations
All violations can be locally processed by
detecting a mismatch between syntactic
specifications of immediately adjacent elements
or elements separated by a determiner (size of
syntactic buffer is kept to a minimum)
All three types are different at the level of the
grammar and semantics  Will reflect ERP
responses to different types of syntactic
information
Types of syntactic violations in Dutch


Agreement:
subject-NP and finite verb do not agree in
number (*On a rainy day the old man buy a
life insurance.)
Types of syntactic violations in Dutch


Subcategorization:
verb that does not take an object NP is
followed by a noun which has a role of
grammatical object (*The tired young man
elapsed the book on the floor.)
Types of syntactic violations in Dutch


Phrase structure:
obligatory word order violation-in Dutch, in NPs
consisting of Adj and Adv, the Adv should
precede the Adj (*Most of the visitors like the
colourful very tulips in Holland.)
SVO and VSO agreement violation

Violations between verbs and nouns: on either
finite verbs or subject nouns within the same
sentence (i.e., the subject noun could be
singular in combination with plural form of the
verb)
Examples: SVO and VSO agreement
violation (the CW is italicized)

Subcategorization violation


Involves obligatory intransitive verbs, which
take a noun as a direct object
In Dutch, unlike English, no continuation is
possible following the noun in object position
Examples: subcategorization
Phrase structure violation


Consist of nouns preceded by transpositions of
adverbs and adjectives
In Dutch, it’s a violation to have an Adj-AdvNoun sequence
Sample sentences: adjective-adverbnoun sequence
Method




Total of 360 sentences – 50% grammatically
correct and 50% contain grammatical violation
Each incorrect sentence is derived from the
companion correct sentence such that words
preceding and following the word string that
makes the sentence ungrammatical are the
same as the companion correct sentence
The sentences are matched in number of words
Other than specific violations, the sets of 180
correct and incorrect sentences are closely
matched  well controlled
Criteria for violations




Immediate (restricted to two adjacent words in
a clause, or by a triplet of words)
The Critical Word has a counterpart in its
companion correct sentences
The CW is either verb or noun, max length =
9 and min length = 4
At least 3 words preceding the CW and at least
two words following the CW
Experiment design
Two experimental lists 180 sentences each
(90 correct and 90 incorrect, each made of 30
sentences of each violation type)
 Order: pseudo-randomized for each list (no >
than 3 incorrect or correct sentences occur)
 The experiment included two parts:
 Grammaticality Judgment Pretest: to ascertain
that the syntactic violations were perceived by
the subjects as such
 The ERP experiment

Grammaticality Judgment Pretest:
Results



The subjects correctly identified the violations
The majority of responses were either to CW or
the word immediately following the CW
 Choosing a CW as a point of
ungrammaticality in a sentence is a valid
method
The ERP experiment



Subjects: 34 University students, all native
speakers of Dutch, mean age=23 years old,
range 18-28 years
Procedure:
The stimuli were displayed in the same manner
as in a pretest (word by word, each word was
presented for 300 ms with an ISI=300 ms, the
ITI depended on number of words in a
sentence)
The ERP experiment






EEG activity has been recorded using an
Electrocap with seven scalp electrodes, each
referred to the left mastoid
Sampling started 150 msec before the
presentation of the first word of each sentence
Total sampling epoch = 8550 msec
Task: comprehending the whole sentence
No other task assigned
The subjects were told that there were
incorrect sentences, but no information on
violations was provided
Results




Average waveforms were computed by subject
for correct and incorrect sentence for each of
the violation types
Calculations are done separately for each of the
seven electrodes
The baseline for CW and its preceding position
was chosen
Each violation analyzed separately
Results by violation: Agreement

Grand average waveforms by electrode site for
the critical word in the correct and incorrect
agreement condition (see page 454-455)
Agreement: Data analysis
CWs have widely distributed positive shift
(the waveforms for the incorrect CW have a
widely distributed positive shift in comparison
to the correct words, which starts at around
500 msec following the onset of the CW and
continues throughout the following word )
 Analysis on each position (to test if the shift
occurs on both grammatical words and CWs):
 Positions preceding the CW
 Penultimate and sentence-final positions
  the shift is different on those positions (p.
456-457)

Subcategorization: Data analysis




The waveforms do not show the same positive
shift to the CW
However, the waveforms show a sustained
positive shift preceding and on the CW
This shift is most prominent at the Fz and right
anterior sites, where it is in a position following
the CW
Frontal positivity is absent or marginal at the
Cp or P lateral sites (p. 459-461)
Phrase structure: Data analysis



The waveforms at the CW are characterized by
a positive shift with the broad scalp distribution
for the incorrect CW in the latency window of
500-700 msec
The positivity is present from the immediate
onset of the CW (p. 463-465)
The negative shift for the incorrect condition is
very similar to subcategorization and
agreement conditions
Summary of the results



The CW in the incorrect agreement and
phrase-structure conditions have a positive
shift compared to the correct conditions
Most probably it results from the syntactic
violation on the CW
There is no significant positive shift in the
subcategorization condition
Summary (continued)


All three violations have a significant negative
shift for the incorrect condition on the
sentence-final, and penultimate positions
The negativity most probably results from the
semantic analysis problems originating in
syntactic violations
Discussion



Major result – the widely distributed positivity
elicited by two of three syntactic violations:
starts at about 500 msec, with a centro-parietal
maximum (very similar to the P600 reported by
Osterhout and Holcomb)
SPS is different from N400 resulting from
semantic violations
Brain seems to process syntactic information
differently from semantic one
Discussion (continued)


The SPS is not an ERP response to violations
only: in phrase structure violations, it occurs
one word BEFORE the syntactic violation (due
to the parser’s attempt to entertain the
possibility of statistically infrequent but
nevertheless grammatical construction
“Determiner-Adj-Adv-Adj-Noun”)
This is consistent with other empirical evidence
that the parser avoids keeping all the
possibilities open until the disambiguating
information is received 
Discussion (continued)



 The parser operates on the basis of the
principle of computational economy or on the
basis of the frequency of alternative syntactic
constructions
The SPS occurs in the absence of other tasks 
the “surprise” account of the observed positivity
is less likely
Both open- and closed- class of words trigger
SPS  the possibility that SPS is related to a
word class is ruled out
Discussion: problems encountered


The subcategorization violation results which
differ from two other violation types are not
only syntactic, but also semantic by necessity
(verb meaning and its syntactic aspect are
closely intertwined)  it is possible that
“semantic” negativity and “syntactic” positivity
occurred on the same word
However, it may also be a methodological
fallacy: it should be also tested on the closed
class of words or using “syntactic prose”
Discussion: problems encountered
(continued)



It is hardly possible to avoid the constellation of
semantic and syntactic violation
The obtained results do allow for strong claims
regarding temporal relation between parsing
operations and semantic integration processes
It is unclear if SPS is time-locked to the initial
structural assignment or to the processor’s
rejection of this first assignment
Autonomous or interactive parsing?
Although the results do not allow for
conclusions that can resolve this debate, they
show that there is some intermediate level of
syntactic representation, which is computed, or
a process of syntactic re-analysis initiated upon
encountering a structural violation. At the level
of language processing there is a relation
between syntactic processing and “syntactic”
positivity in the ERP waveform.
Autonomous or interactive parsing?




Positive shift is a manifestation of a functionally
distinct process in computation of syntactic
structures
A new label is assigned to this ERP response:
SPS
The obtained positivity can be related to the
domain of syntactic parsing
It is a common brain response to very different
types of syntactic violations
The end