: LCV
l

‘ .. I. «r», ,
, wwumunnfirfiu
Emma. ‘
. . a". V
. . .méfi r:
. i S .
. . ‘ a?
. a. ii
i 2., ‘

34:! . .
. unu’nwrnfl was»
, i..- 25. - 1A
.1 .33”:
1%.... u
. 5.1.21.3 .unwflw
. :15...) (I! .1...
. r .13..-.tqusrq
. n. : . ,
a.
Iflki: r:
If.

 

«9:!

mm.

 

2...
1.3.2.?
5.3.3..
1.0.”.
3. x
I. II

 

 

“fl

 

 

 

”WWW

.

 

 

 

 

 

. . .16.. ., “a...
.. H . :5: Sigma”.
. N $3?” a. a??? .,

 
 
 

n .._u 44 EL:
., r

 

., .. .. _._.. m .F.
m, “aw... i. an.

ThEbLS

DOOl

This is to certify that the

thesis entitled

An fMRI Study of Regular and Irregular Inflections
in German

presented by

Carrie Lynn Campbell

has been accepted towards fulfillment
of the requirements for

M.A. . Linguistics
degree in

[jewel

Major professor

 

Date XfiQlflO]

0-7639 MS U is an Affirmative Action/Equal Opportunity Institution

 

 

LIBRARY
Michigan State
University

 

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 c:/CIRC/DateDue.p65-p.15

 

AN FMRI STUDY OF REGULAR AND IRREGULAR INFLECTIONS IN GERMAN
BY

Carrie Lynn Campbell

AN ABSTRACT OF A THESIS

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

MASTER OF ARTS

Department of Linguistics
2001

Associate Professor Alan Beretta

ABSTRACT
AN FMRI STUDY OF REGULAR AND IRREGULAR INFLECTIONS IN GERMAN
By
Carrie Lynn Campbell

Theories about the processing of regular and irregular inflections have
disagreed over whether they should be represented by a dual or single
component model. Traditional linguistic theory assumes the two components of
grammar for regular inflections and memory for regular stems and irregular
inflections. Rumelhart and McClelland (1986) have introduced a computer model
which dispenses with symbolic rules and seeks to explain both regular and
irregular inflections with a single mechanism.

Event-related fMRI was used to test this distinction on German inflections.
Plural formation of regular and irregular nouns and past participle formation of
regular and two different classes of irregular verbs were studied. These theories
have different implications for neural activation. lithe two-component hypothesis
reflects the actual organization of the language faculty, two distinct patterns of
brain activation should be observed for production of irregular and regular noun
and verb inflections. If a single-route model is correct, then there should be no
neural dichotomy between German regular and irregular inflection.

These results show distinct patterns for neural activation for regulars and
irregular inflection in the total level of activation, hemispheric balance of
activation, and specific regions activated by each inflectional class. These

findings are consistent with predictions of the two-component model of inflection.

ACKNOWLDGEMENTS

The author is grateful to the members of her committee, Alan Beretta (chair),
Tom Carr, and Yen-Hwei Lin; to Yue Cao and Jie Huang whose direction and
guidance was crucial to this project; to David McFarlane for training on E-Prime
programming; and to Kiel Christianson and Lothar Schmitt for helping with the
word frequencycounts. She acknowledges also the generous support of Jim
Potchen and the Radiology Department of Michigan State University for fostering
linguistic fMRI research at Michigan State University. This research was funded
in part by a summer graduate assistantship from the Cognitive Science Program

at Michigan State University.

iii

TABLE OF CONTENTS

List of Tables ...................................................................................................... viii
List of Figures ...................................................................................................... vii
1 Introduction .................................................................................................... 1
1.1 The challenge: connectionism ................................................................ 2
1.1.1 Outline of the model ......................................................................... 2
1.1.2 The ensuing debate ......................................................................... 3

1.2 The response: dual routes 5
1.2.1 The default nature of the regular ...................................................... 7
1.2.2 Frequency effects ............................................................................ 8
1.2.3 Morphological priming ...................................................................... 9
1.2.4 Aphasia and impaired systems ...................................................... 10
1.2.5 Acquisition ..................................................................................... 12

1.3 Counterarguments ................................................................................ 13
1.3.1 Successes of connectionism ......................................................... 13
1.3.2 lndecisive neuroimaging studies .................................................... 14
1.3.3 Limitations of English ..................................................................... 19

1.4 German ................................................................................................. 20
1.4.1 Facts of the German system .......................................................... 21
1.4.2 Frequency ...................................................................................... 23
1.4.3 The case for default in German ..................................................... 23
1.4.4 Frequency effects .......................................................................... 27

iv

1.4.5 Priming .......................................................................................... 28

1.4.6 Aphasia .......................................................................................... 29

1.5 Neuroimaging ....................................................................................... 29
1.5.1 Event-related brain potentials (ERP) ............................................. 29
1.5.2 PET ................................................................................................ 30

1.6 Overview of the current study ............................................................... 31

2 Methods ....................................................................................................... 33
2.1 Subjects ................................................................................................ 33
2.2 Materials ............................................................................................... 33
2.3 Language Protocol .................................................. g .............................. 34
2.4 MR Imaging Protocol ............................................................................ 35
2.5 Image Processing ................................................................................. 36

3 Results ......................................................................................................... 39
3.1 Empirical Observations ......................................................................... 39
3.1.1 Common Map ................................................................................ 39
3.1.2 l-R Contrast ................................................................................... 40
3.1.3 R-l Contrast ................................................................................... 41

3.2 Regularity .............................................................................................. 43
3.3 Regional Analysis ................................................................................. 45
3.4 Lateralization ........................................................................................ 45

4 Discussion .................................................................................................... 46
4.1 Levels of activation ............................................................................... 46
4.2 Hemispheric differences ....................................................................... 47

4.3 Regions of activation ............................................................................. 47

5 Conclusion ................................................................................................... 51
REFERENCES .................................................................................................... 52
APPENDIX .......................................................................................................... 57
1. Verbs ........................................................................................................ 57
2. Nouns ....................................................................................................... 59

vi

LIST OF TABLES

Average activation in brain regions common to both regular and irregular data .40
Average activation of brain regions in the l-R subtraction condition .................... 41
Average activation of brain regions in the R-l subtraction condition .................... 42

Areas of activation for regulars and irregulars in previous neuroimaging studies
compared with results from the current study ..................................................... 48

vii

LIST OF FIGURES

Left and right hemisphere contrasts between regular and irregular activation in a
single subject ..................................................................................................... 43

Regular and irregular paradigms in the left hemisphere of two subjects ............ 44

viii

1 Introduction

Traditionally the language faculty has been viewed as a set of stored,
memorized words in addition to a set of combinatorial rules to help put the stored
elements together. This system relies both on rules that manipulate symbols, or
variables, allowing different stored words to utilize the same set of rules. This
system has recently been challenged by those who would model language
acquisition using only a single process of associative memory. The debate
between these two views has focused primarily on inflectional morphology.
Symbol-manipulators argue that two routes are necessary to process inflections:
memory stores individual lexical stems and irregular inflected items while regular
inflections require an additional rule that is applied either to an existing stem or
extended to other stems. However, in an effort to posit as simple a system as
possible, some modelers have eliminated the rule-based route and attempt to
show that human-like performance can be achieved based on memory
processes alone.

Henderson has written (1965), “Research into the activity of the brain is, of
course, considerably handicapped by its almost complete inaccessibility to
objective investigation. It has then to be studied indirectly through the behavior
which results from it; and of all kinds of behavior linguistic behavior is one of the
most revealing for this purpose.” The rules studied by modern linguists have
been meant to do precisely what Henderson suggested; dual-route accounts of
inflection assume that the rules linguists have formulated based on behavior truly

do reflect the activity of the brain. Single-route theories view them merely as

descriptive of behavior and not the underlying processes. A great deal of
evidence has been amassed that supports the notion of a default regular
inflection but, at the same time, associative models have been devised which
their proponents believe are clear advances over earlier models. This study uses

neuroimaging to “ask the brain” how many systems it uses to process inflection.

1.1 The challenge: connectionism

The event sparking this controversy was a computer model of language
acquisition and processing described by Rumelhart and McClelland (1986). This
model, consisting of a network of densely connected units representing features
in input and output forms, used probabilistically weighted connections which
“learned” to associate input with output forms by adjusting connection strengths
and thresholds.

1.1.1 Outline of the model

This model employs a distributed network, where each input and output
node corresponds not to a particular word, but to a feature, or in this case a set
of features. Because Rumelhart and McClelland use Wickelphones, or sets of
three adjacent phonemes, to break up words, they employ sets of simplified
features for three adjacent phones in their representation (Wickelfeatures). Each
node codes for a triple of features, one each from the predecessor, central, and
successor phoneme. Only combinations of different values on the same
dimension for the predecessor and successor phonemes, as well as those
encoding word-boundaries, were used, giving a total of 460 Wickelfeatures.

Thus two sets of 460 nodes were necessary to encode both inputs and outputs.

Node activations are discrete, so units that are on to represent the word
should be set to 1 and all other set to O. The threshold for the activation of each
unit is established through the training procedure.

Connections between these nodes are initially set to 0 so that the input
nodes have no influence on the output. During the learning stage, the model is
presented with both the root form of a verb and the correct target. lt computes,
based on the strengths of its connections, an output given the verb root input,
and compares this with the actual target. The connection strengths are then
adjusted using a variation of the delta rule which adjusts both the connection
strength and the threshold value by a value of one when the output of a node
does not match the target output.

Rumelhart and McClelland trained their model to perform English past
tense inflection by training it on sets of words which arguably reflected the input a
child would receive: early training took place on a set of ten high-frequency verbs
of which 8 were irregular and later training stages involved progressively lower
frequency (and more predominantly regular) verbs. The model was able to
successfully replicate some observed patterns in child language acquisition,
including a U-shaped learning curve with stages of accurate but limited
performance, overregularization, and proper usage.

1.1.2 The ensuing debate

Based on this success, Rumelhart and McClelland declared, “We have

shown that a reasonable account of the acquisition of past tense can be provided

without recourse to the notion of a ‘rule’ as anything more than a description of

the language. We have shown that, for this case, there is no induction problem.
The child need not figure out what the rules are, nor even that there are rules” (p.
267, emphasis theirs). This assurance that rules would be henceforth
unnecessary and challenged the very foundations of modern linguistics and
cognitive science which both rely on symbol manipulation. Because the notion of
rules involving some kind of symbol or variable is fundamental to classical
cognitive science and linguistics, this challenge has sparked much debate.

The debate, however, need not be between networking models
themselves and those who insist on the necessity of some rule involving
symbolic manipulation. As Marcus (2001) shows, models built using
connectionist networks can in fact make use of rule-and-memory systems,
whether explicitly or inadvertently. Westermann and Goebel (1995), for example,
purposely include a short—term memory module to represent the regular, rule-
based path as well as a phonological lexicon for irregulars. Hare et al. (1995)
have a similar two-part network including one feedforward network that works like
othersuch models, without implementing any rules, and a second “clean-up”
network, which copies the stem automatically, preparing it for the application of a
rule. This effectively implements symbolist suggestions about possible
mechanisms for rule-and-memory models: it computes a possible output based
purely on phonological associations, then passes this output and the verb stem
along to the clean-up network. If the feedforward network’s output is strongly
activated, the clean-up network is suppressed, but if weakly activated the clean-

up network is engaged, applying the —ed suffixation.

Thus “connectionist” is not a term which necessarily implies an advocate
of alternatives to symbol manipulation. However, because most connectionist
networks under study currently are in fact advocating such alternatives, the term

connectionist will be used in this paper in that somewhat over-simplified sense.

1.2 The response: dual routes

The first major reply to this connectionist model and related models came
from Pinker and Prince (1988). They propose that neither a totally rule-based
nor a totally associative system can account for the acquisition of regular and
irregular forms, but suggest instead an associative process of acquisition and
access for irregular forms and a rule-based process for regulars used for default
production. In response to the connectionist challenge they point out certain
facts about English regular and irregular morphology that any model must be
able to explain to be considered successful and discuss the inability of the
Rumelhart and McClelland model and any that rely on a single associative
process to address these issues.

The notion of a default is crucial to the dual-route approach. While in
English the concepts of the regular as both a default and common ending are
confounded, the generalizability of certain forms, not their prevalence by token or
type, is the considered crucial characteristic of a default regular in this account.
Irregular inflections apply in a limited way to certain classes of words (which may
be large or small) and should never be generalized to words which have no
similarity to their class. Regular inflections, however, should be able to apply to

all words not included in an irregular class, and may even be used on novel

words which resemble irregulars or be misapplied to words in an irregular class,
thus intruding into “irregular territory”.

This default nature of the regular is a major defense of Pinker and Prince
against associative models. For human language shows that regular endings
can apply not only in circumstances in which the word to be inflected resembles
other regular words (this would be accounted for easily by connectionist
networks), but also when it resembles or is even a homophone to an irregular
word. Given the association of a phonological form with an irregular inflection
already, such a system should not consistently regularize these regular
homophones of irregular words. This pattern of challenge to connectionist
models has been elaborated by others as well, notably by Marcus et al. (1995)
who synthesized Pinker and Prince’s list with other subsequent evidence found
for the distribution of regulars. They find at least 21 distinct circumstances in
which regular inflections are applied as a default, including many that
connectionist models have even yet not been able to account for.

One major difficulty with current connectionist’models is their reliance on a
purely phonological (or Wickelphonological) representation of the input which is
inconsistent with the features of real language. Even while all subsequent
models have abandoned Wickelphones, the reliance on mere phonological input
in almost all current models ignores other aspects of words such as semantics
and does not reflect human language use. Such a system cannot account for .
homophonous words which differ in regularity (ring a bell/ring a campfire) or

forms similar to large classes of irregulars that are regular themselves (throw-

threw vs. flow-flowed). Large classes of nouns that consistently take regular
inflections include members with phonological similarity to or identity with
irregular forms. These words show not only the limitations of the phonological
reliance of connectionist systems, but the generalizability of regular inflection.
Furthermore, phonology-only inputs have no basis for recognizing the stem
within the input because phonetic distinctions often fail to distinguish the stem
from its affixes. This explains the many recent models that show more blends
(e.g. wented) than overregularizations (e.g. Plunkett and Juola 1999, Plunkett
and Marchman 1991, MacWhinney and Leinbach 1991) — contrary to the facts of

child development (Marcus et al. 1992).

1.2.1 The default nature of the regular

Extensive lists of verb classes that are consistently regularized are seen
throughout the literature (see Pinker and Prince 1988, Marcus et al. 1995). The
most prominent and robust are discussed below.

First, words which lack an entry in memory or have no similar entry from
which to draw analogies are consistently regularized. This class includes novel
words (snarfed), low frequency words which may not have a participle entry
(stinted), and unusual nonce words (ploamphed). While connectionist models
built upon systems like English with regular forms in the majority can generalize
to new forms accurately, forms which are completely unrelated to forms

recognized by the model cause trouble for connectionist models (Seidenberg

1992).

Words whose linguistic entry is not a canonical root including the full
complement of phonological, syntactic, and semantic information are inflected
according to the regular paradigm. Such words include onomatopoeia (peeped),
quotations (“man”’s), names (the Chi/d3), some borrowings (kimonos), and
acronyms (CEOs). Many of these forms bear great similarity to irregular forms
and thus risk irregularization by analogy in a connectionist model. Other words
which are derived from a different category share a weakness in mental
representation because they cannot be marked for the proper inflectional
features. Such denominal verbs (high-sticked), deadjectival verbs (n'ghted the
boat), and nominalizations (wo/fs of food) are regularized and cannot be
accounted for using purely phonological features.

In cases of headless compounds, features from the head cannot percolate
to the whole word because they differ in category. This occurs often in

compounds (Iow-lifes) or derivations via a name (Toronto Maple Leafs).

1.2.2 Frequency effects

Frequency effects for irregular participles show that lower frequency
irregulars are significantly more likely to be overregularized than high frequency
irregulars. Frequency is not an effect, however, for error rates in regular
participles (Pinker and Prince 1991). This is in keeping with an account in which
regulars involve a rule which is unaffected by frequency, while irregulars rely on
memory, which would be less reliable for low than for high frequency items.

Furthermore, frequency comes into play in response times. Prasada et al.

(1990) show distinctions between performance on regulars and irregulars when

subjects orally produced past tense forms from visually presented infinitives of
the same verbs. High frequency irregulars were significantly faster than low
frequency irregulars, while regulars showed no such frequency effect. Again, this

fits most clearly into a dual-route account of inflection.

1.2.3 Morphological priming

Priming experiments present subjects with two stimuli while their
relationship is manipulated to determine the effect that processing the first
stimulus will have on processing the second. Repetition of a stimulus facilitates
its lexical access, so the second in a set of identical stimuli is generally
associated with faster response times for tasks performed with the stimulus.
ERP effects are also associated with primed stimuli.

In a study of visual priming, Miinte et al. (1999) showed that regular past
participles in English exhibited a priming effect on their stems while irregulars did
not. ERPs to regular verbs were associated with an N400 reduction in the
primed condition, while ERPs to irregulars showed no such effect. This study
was controlled for phonological and orthographic priming, so the effect observed
is likely due to regularity and not merely to formal similarity of participles to their
stems.

The priming of regular participles by their infinitives and vice versa is
consistent with the dual-route model. Because regular participles are made up of
stem and affix, processing either form will activate the lexical entry for the stem.
Because the presentation of the first instance of the verb accesses the stem, the

second instance will be facilitated in stem access, even if the inflectional aspects

of the word are different. Irregular participles and infinitives, however, are
separate lexical entries under the dual-route model, and so priming should not

occur as the above studies show.
1.2.4 Aphasia and impaired systems

1.2.4.1 Aphasia

ln agrammatic Broca’s aphasia inflectional morphology is often impaired,
presumably by damage to the part(s) of the brain which are responsible for the
processing of inflection. By observing the patterns of breakdown one can
compare the major theories of inflection because these models should not be
equally vulnerable to certain types of impairment. lithe dual-route model is
correct, then it should be possible to selectively impair either regular or irregular
inflection in a language disorder.

Marslen-Wilson and Tyler (1997) describe just such a double—dissociation
in impairment. They tested three agrammatic aphasic subjects on an auditory
priming task in which subjects were asked to decide whether the spoken target
word (following a spoken prime) was a word or not. Morphological priming
effects for inflected regular and irregular primes as well as semantic priming
effects were measured. Control subjects showed priming in all three of these
cases, with no interaction between regularity and priming. However, the aphasic
patients showed a three-way interaction between patient, regularity, and priming,
with two aphasic patients showing priming for irregular past tense and inhibition

for regular past tense and one aphasic producing the opposite pattern.

10

Japanese aphasics have also been studied and have been shown to
exhibit a dissociation by aphasia type in preference for selecting a regular or
irregular nominal suffix in sample discourses (Hagiwara et al. 1999). Broca’s
aphasics selected a regular nominal ending when it was semantically required
significantly less often than did normals or aphasics with other damage. Gogi
aphasics, however, who are characterized by intact inflection but impaired lexical
retrieval, used significantly more regular nominalizations than did normals in a
context that forced irregular suffixation.

While studies in connectionism have succeeded in impairing regular
performance by removing nodes from a model, creating a “lesion”, none have
replicated an impaired irregular system by this process. Although this does not
show decisively that lesions to an associative system could not produce the
above patterns, the “double dissociation” of the aphasia studies described above
reflects more clearly the dual-route analysis which allows for selective

impairment of either route.

1.2.4.2 Alzheimer’s disease

In a variation on Pinker’s (1991) dual-route theory, Ullman has advanced a
theory that the lexicon is part of a “declarative/semantic memory” system that
includes other factual memory and that grammatical rules are stored in a
“procedural/skill system” that also handles some non-linguistic skills. He bases
much of his theory on contrasts among Alzheimer’s patients, who he finds to be
impaired in tasks in his “declarative memory” category (e.g. irregular production

and animal naming) but relatively spared in “procedural” tasks (e.g. past tense

ll

production and tool naming), and Parkinson’s Disease, which shows the opposite
pattern (Ullman et al. 1997, Ullman1999). Thus, while Ullman does not suggest
that language areas in the brain are specialized for language alone, he supports
the idea that rules and memory are distinct components of the system. He even
suggests different neuroanatomical locations for them related to areas of damage
in each of these diseases, identifying the impairment of declarative and lexical
memory with damage to the tempoparietal cortex and that of procedural or
grammatical rules with damage to the frontal/basaI-ganglia system including
Broca’s area.

1.2.5 Acquisition

Given the evidence described above, there are solid grounds for claiming
that adults make a distinction between a default regular and a structured lexical
entry for irregulars. It seems logical, as well, to ask how this distinction came to
be, and whether the patterns of acquisition can give any hints about the
processing of inflection.

On one level, it is quite impressive that children can possibly learn that
one inflection is a default, since irregular words make up a majority of the high-
frequency children encounter early on in acquisition. Furthermore, inflections are
non-salient, being monosyllabic (or less) and spoken as unstressed bound
morphemes. Eventually, it seems, children must notice that only regulars can
apply to any phonological pattern and in unusual circumstances (names,

borrowings, onomatopoeia, etc). Since this observation relies on experience, it

12

may take children some time to fully identify and grasp the default in any given
language.

However, research has shown that children are in fact quite adept at
identifying defaults in English; by age four they use the -ed past tense for novel
words. In fact, children characteristically overapply the default, in
overregularization errors like holded (Marcus et al. 1992). Yet these
overregularizations actually seem to follow certain patterns, showing sensitivity to
frequency and similarity (fewer errors take place on high-frequency or large-
family irregulars), and ultimately accounting for only 10% of all inflections.
Children also reflect adult behavior in compounding words, allowing irregular but
not regular plurals in compounds like rat-eater (not rats-eater, but mice-eater is
okay). In all these respects children’s behavior mirrors that of adults and make
clear the need for any successful model of language acquisition to be able to

recognize and properly apply default inflections.

1.3 Counterarguments

Given all the data above that are consistent with a dual-route account of
inflection and pose challenges to connectionism, why has eliminative

connectionism not yet been eliminated?

1.3.1 Successes of connectionism

Since Rumelhart and McClelland’s 1986 model, many new models have
been developed which have succeeded in dealing with the individual problems
pointed out by Pinker and Prince. For example, MacWhinney and Leinbach

(1991) describe a model that can account for the treatment of homonyms to

13

irregulars as regular, but it doesn’t succeed in consistently adding —ed to
unusual-sounding, unfamiliar words or in avoiding blends as real speakers do.
Plunkett and Juola (1999) present a model that properly shows frequency effects
for both response time and error rate, but it fails to avoid blends and the authors
do not demonstrate clearly whether the model can generalize —ed to unusual-
sounding unfamiliar words.

Thus these new models are often very good at the particular problem for
which they are tailored, but still fail to demonstrate how a comprehensive single-
route connectionist model can account for all the various facts of English
inflection. Yet each new partial success serves to keep the controversy alive.

Furthermore, neuroimaging research to date has been inadequate and
difficult to interpret as described above. First, the use of block design makes it
unclear what is being tested. Additionally, many studies (and all discussed thus
far) have made use of English, which is not an ideal language to decide between
the competing accounts of inflection.

1.3.2 lndecisive neuroimaging studies

Seidenberg writes that “The nature of language is determined by how it is
acquired and used and therefore needs to be explained in terms of these
functions and the brain mechanisms that support them.” (1997, p. 1601)

If the goal of each both associationists and symbolists is to describe the
underlying brain mechanisms, then a useful area to search for answers in this
debate is that of neuroimaging, in effect asking the brain how many different

processes it thinks it has. In this section several recent neuroimaging techniques

14

and findings pertaining to this debate will be reviewed. However, because of
several confounds of design and findings that are difficult to interpret make the
data obtained in these studies indecisive as regards the symbolist/connectionist
debate.

1.3.2.1 Positron emission tomography (PET)

PET measures metabolic activity in the brain by detecting positrons
emitted from an isotope ingested by the subject. Scans acquired in this way can
be used to provide estimates of regional cerebral blood flow (rCBF). One major
drawback to this method of neuroimaging is its inability to look at real-time stimuli
and requirement for block design.

Using this technique, Jaeger et al. (1996) found that both regular and
irregular English verbs activate the left middle and inferior frontal gyri (including
Broca’s area) and bilaterally activate the precuneus. Regular verbs in their study
were associated with left dorsolateral prefrontal, left anterior cingulate, and
anterior left inferior parietal activation, while irregular verbs activated the left
superior frontal gyrus, left superior parietal Iobule, posterior left inferior parietal
Iobule and left middle temporal gyrus. However, this study used a block design
that may have resulted in subjects forming response strategies to the regular and
irregular conditions. Penke et al. (1997) note that, given the block design of the
study, it is impossible to know whether the observed activity is truly due to a
regular-irregular distinction or merely the production of different kinds of lists.

In their critique of Jaeger et al., Seidenberg and Hoeffner (1998) note

several ways in which block design confuses the issue actually being tested.

15

First, the large body of priming research already demonstrates that performance
on one trial of an experiment is affected by its similarity or dissimilarity to
preceding trials (see section 1.2.3). Second, if blocks that are predictably made
up of regular verbs the subject need not processa word beyond the final
phoneme to know which inflection to apply (the regular) and to determine which
allomorph of the —ed ending to use. For irregular blocks, generating the correct
past tense form would still require the recognition of the word as a particular
lexical item. When items are presented randomly all words must be processed
equally deeply to determine which inflection to use. Jaeger et al.’s own RT data
actually support the claim that their subjects employed a strategy that did not
require identifying words in regular blocks as distinct lexical items, but only
accessing superficial features. In their study, it took subjects 38 ms on average
to produce a regular past tense, while other research suggests that 100 ms is
needed to identify a word, and that fixation on a word in text normally lasts 200-
250 ms. Thus it is plausible that Jaeger et al.’s subjects, rather than identifying
words in regular blocks and processing them normally, were taking advantage of
the fact that all words required the same inflection and merely attended to

enough of the word to choose the proper allomorph.

1.3.2.2 Magnetoencephalography (MEG)
MEG uses a special detection device to measure time—varying magnetic

fields generated naturally by neural activity. It has the advantage of observing

brain activity in “real-time” while maintaining detailed spatial resolution.

16

Rhee et al. 1999 used MEG to examine production of past tense verbs
and found dipoles in the left temporoparietal region for both regular and irregular
verbs with slightly later dipoles in the left frontal regions only for regular verbs.
These results are presented as consistent with a dual-route model in which
regular verbs stems must be accessed in memory (temporal) and then inflected
in the grammar (frontal), while irregulars find an appropriate inflected form in
memory, blocking the grammatical rule to create past tense. However, the
prediction of these findings is that regulars must always be slower than
irregulars, which is contrary to fact (see section 1.2.2).

While vague, in terms of localization these findings are somewhat at odds
with the PET study described above. Jaeger et al. (1996) found no common
activation in the temporoparietal region though some temporal and parietal areas
were activated by irregulars and left frontal regions were activated in regulars.
However, the block design of the former and the potentially confounding
sentence-frame protocol of the latter study make disagreement with their findings
a less serious matter.

However, the results of Rhee et al. are also confusing in light of RT
studies which find the English irregular verbs take significantly more time to
inflect than regulars (Jaeger et al. 1996, Seidenberg 1992, Prasada et al. 1990).
If regulars are inflected more quickly than irregulars, it is unclear why they would

be characterized by a later dipole than irregulars, as Rhee et al. report.

17

1.3.2.3 Functional magnetic resonance imaging (fMRI)
Functional MRI is a neuroimaging technique sensitive to changes in

deoxyhemoglobin in brain tissue. Using powerful magnets to polarize molecules
in the body, fMRI detects the variable differences in transverse magnetization
decay time (T2) and the local magnetic field homogeneity (T2*) of water in the
brain tissues. Because the ration of deoxygenated to oxygenated hemoglobin in
the blood in a small region, or voxel, influences the local T2*, detection of T2*
signal levels can be used to track blood flow and oxygen balance in tissue. fMRI
gives good anatomical resolution, though its temporal resolution is far less than
that of ERP. It also allows for an event-related study that avoids the pitfalls of the
block design as described above. Furthermore, fMRI has been shown by Pugh
et al. (1997) to be predictive of psychological performance in a word identification
study. These features make fMRI a good tool to use to investigate the nature of
the inflectional system.

Ullman et al. (1997) take advantage of this tool in a study which utilized a
silent production task. Subjects (all male) were shown verb stems in regular and
irregular blocks and were asked to silently produce their past tense forms.
Comparisons of both regular and irregular to fixation showed bilateral activation
in the inferior frontal gyrus (including Broca’s area), caudate nucleus, superior
temporal gyrus, and supramarginal and angular gyri. The frontal cortex and the
basal ganglia showed more activation for the irregular than for the regular
conditions, while posterior regions did not show this distinction. A decreased
signal for irregular compared to fixation was found in the temporal and temporo-

parietal areas. This decrease in signal is difficult to interpret and it is not clear

18

how it plays into either a single- or dual-route account of inflections. While areas
of differential activation for regulars and irregulars initially suggest distinct
processes involved in regular and irregular production, the increased activation in
the frontal cortex for irregulars compared to regulars seems to contradict a
general association of anterior language areas with syntactic function (e.g.
inflectional rules) and posterior areas with memory functions. This study is
problematic not only in the interpretation of its findings, but also in its block
design.
1.3.3 Limitations of English

Two major confounds are present in English which limit its ability to
distinguish between single— and dual-route accounts. Regularity interacts with
both frequency and affixation so that the characteristics of regular, default forms
described above could also be attributed to these differences.
1.3.3.1 Frequency

In English, the regular forms of both nouns and verbs are also the most
frequent, and thus overregularization could be due to frequency of input and not
a default assignment. Plunkett and Marchman (1991, 1993) show that altering
the ration of regular to irregular items greatly influences the performance of their
network model respecting proper generalization of the regular. Specifically, they
find that “generalization is virtually absent when regulars contribute less than
50% of the items overall” (Plunkett and Marchman 1993, p.55). Thus, the
prevalence of regulars in English is crucially linked to the ability of connectionist

models to represent proper generalization of the regular.

19

1.3.3.2 Affixation
English verbal and nominal affixes are similar in that each involves a suffix

added to the stem with no stem change. Irregulars, however, are characterized
by stem changes for both nouns and verbs, sometimes in addition to other
changes (e.g. child-children, dig-dug). While there are some exceptions that
have no stem change for the irregular (e.g. fish-fish or let-let), there is no case of
simple affixation for irregular without an accompanying stem change. Thus it is
possible to argue that the distinctions between regular and irregular described
above result at least in part from the difference between processing affixes and
stem changes.

There is a language which avoids these confounds, and thus provides a
better test case for questions about symbolist vs. connectionist systems:

German.

1.4 German

Although much of the early work that addressed the processing of
irregular and regular forms explored English in particular despite the confounds
discussed abouve, a wider range of languages is now being considered including
Japanese (Hagiwara et al. 1999), Arabic (Wunderlich and Fabri 1995), and Italian
(Say and Clahsen in press). Recently German has been suggested as an ideal
test case for the issue of inflectional processing leading to a series of studies of

this problem in German (Marcus et al. 1995).

20

1.4.1 Facts of the German system

1.4.1.1 Verbs

An outline of the German verb paradigm is given in (1). There are three forms for
each verb: infinitive, preterite, and participle, listed in that order. lnfinitives are
composed of a stern and infinitival suffix —(e)n. Participles are more common
than preterites in informal speech, thus they are the inflected form that will be
considered here. All participles of the forms considered in this experiment have
a ge- prefix; this is necessary phonologically but does not differentiate the
morphology and thus will not be considered further. Weak verbs consist of a root
and a suffix, -t, strong and mixed verbs may involve stem changes from the
infinitive and have suffixes —en and —t, respectively. For theoretical reasons
(Wunderlich and Fabri 1995) as well as their default nature (described below),
they are considered to be the regular form, while strong and mixed verbs
represent different irregular forms. The characterization of the strong irregulars
into three classes (shown in (1)) is based on the pattern of stem change
associated with each class.
(1) a. Regular (“weak”)

spiel-en spiel-te ge-spieI-t

‘to play’ ‘played’ ‘(has) played’

b. Irregular (“strong”)

ABA (e.g. au-ie-au)

lauf-en lief ge-Iauf-en

‘to run' ‘ran’ ‘(has) run’

ABB (e.g. ei-i-r)

pfeif-en pfiff ge-pfiff-en
‘to whistle’ ‘whistled’ ‘(has) whistled’

21

ABC (e.g. eh-ing-ang)

geh-en ging ge-gang-en
‘to go’ ‘went’ ‘(has) gone’
c. Irregular (“mixed”)

brenn-en brann-te ge-brann-t
‘to burn’ ‘burnt’ ‘(has) burnt’

Thus, unlike those in English, all German participles involve overt
affixation. This is crucial for studies contrasting regular and irregular forms
because it ensures that findings of regular-irregular difference are not merely the
result of differing affixation characteristics. While dual-route theorists would still
argue that only the regular verbs have an underlying representation as stern and

affix (irregulars being stored in memorY). the forms would naturally be dealt with

under one process in single-route models.

1.4.1.2 Nouns

German has five plural suffixes: -(e)n, -s, -e, -er, and —0 given in (2).
Three of these suffixes are associated with changes on the verb stem as well,
but this seems to be due to an independent morphophonemic rule and so will not
be considered here. No known systematic pattern adequately explains a
relationship between semantic properties and the plural suffix a noun takes;
some correlations between gender and morphophonemic properties of the root
do exist, but exceptions are extensive. In the case of nouns as for verbs, all
formsiconsidered in our experiment have segmentable affixes (no zero inflected

nouns are used in this study).

22

(2) a. Regular

Auto Auto-s
car ‘cars’

b. lrregulars

Daumen Daumen
‘thumb’ ‘thumbs’
Hund Hund-e
‘dog’ ‘dogs’
Kind Kind—er
‘child’ ‘children’
Frau Frau-en
‘woman’ ‘women’

1.4.2 Frequency

While English regulars represented the most common class of either
nouns or verbs, in German regulars are in the minority, while the remaining forms
are split among several irregular classes. Nouns are nearly 100% irregular by
type or token, while verbs are less ovenlvhelmingly so. Within the 1000 most
common words, the regular verbal ending accounts for 45% of types (Marcus et
al. 1995). Thus German avoids the confound of frequency and regularity found

in English: defaults cannot be accounted for simply by high frequency.

1.4.3 The case for default in German

As discussed in section 1.2, the key feature of a default is its ability to
extend to new words regardless of similarity. This included novel words,
nonsense words, noncanonical words, unusual—sounding words, etc. The next
section will demonstrate that German does indeed have a default inflection for

both nouns and verbs that patterns much like the English regular.

1.4.3.1 Extension to nonsense words
Clahsen (1997) used the preterite form of nonsense verbs to indicate

which regular or irregular class the word should belong to in order to examine
potential regularization (or irregularization) of novel words. Subjects were
presented with an infinitive and preterite of a nonsense verb in a sentence
context, thus determining the class to which the verb should belong. Their task is
to use the preterite form of this verb to fill a gap in a second sentence and the
participle to fill a gap in a third sentence. In the third sentences, 97% of
nonsense verbs that patterned like regulars in the example sentence got the
regular —t ending, while only 31% of those that patterned like irregulars got the

. irregular —n ending. Sixty-nine percent of those exemplified as irregulars were
eventually regularized, while only 3% of the “regular” nonces were irregularized.
This is a case of massive encroachment by regulars into irregular territory, with
almost no reciprocal irregularization.

In terms of frequency, the —s inflection is very rare, accounting for no more
than 5% of nouns by token or type, so this inflection is by no means regular in
terms of commonality. However, it does behave as a default. Marcus et al
(1995) presented singular nonsense nouns to subjects in the context of a
sentence. These nonsense nouns were taken from lists that either rhymed with
existing German irregular nouns or did not rhyme with any German noun.
Subjects were asked to rate the possible German plurals of the noun, also
presented in sentence form. When the nonsense noun rhymed with an existing
irregular, irregular plurals received significantly higher ratings than for non-

rhyming nouns. Conversely, regular plurals received significantly higher ratings

24

for non-rhyming nouns than for rhyming ones. Thus irregular inflections were
generalized based on similarity only, while regular inflections were generalized to
new nouns when no similarity existed. This generalizability where there is no
similarity is the key feature of a default.

1.4.3.2 Extension to noncanonical words by adults

With respect to noncanonical roots, German also shows a default pattern
much like that of English. Marcus et al. (1995) presented new denominal verbs
in sentences with the regular —t or the irregular —n ending. Acceptibility ratings
showed significantly better acceptance of the —tthan -n ending. This shows that
the regular ending is extended even to denominal verbs homophonous with
existing irregular verbs, a clear case of the regular encroaching on regular
territory.

Further, Marcus et al. presented nouns as borrowings or names in a
similar task and found that the regular —s plural ending got significantly higher
acceptability ratings than did the irregular -n ending. Thus both verbs and nouns
with non-canonical roots resort to the default affixation rule.
1.4.3.3 Errors

In a study that investigated the effects of purposely using an incorrect
form, Clahsen et al. (1997) presented subjects with sets of two stimuli and asked
them to decide whether the items were the same or different. Reaction times
should correlate with well—formedness: well-formed words should be judged more
quickly than ill-formed ones. The sets of stimuli consisted of nonsense verbs

which had already been memorized by the subjects as either regular or irregular.

25

Pairs of these verbs were presented in participle form with either a —tor a —n as
an affix. While “regular” nonsense verbs were associated with longer RTs for an
—n ending (irregularization) than a -t ending (correct), “irregular” nonsense words
showed no difference in RTs for either a —n (correct) or a —t (overregularized)
ending. This is logical if the —t ending is a default: an irregularization adding -n
to a regular word constitutes an affixation error increasing RT, while a default -t
affixed to an irregular word is not a true error as the default can apply across the
board.

1.4.3.4 Child Acquisition

Because of the confounds of English described above, one could account
for children’s overregularization patterns to an extent without employing a default.
The frequency of the regular in English makes it possible for overregularizations
to be based merely on the number of times a child is exposed to the regular
ending. Furthermore, studies finding that children use only singular regulars
within compounds, dropping the plural —s could be explained in terms of a
preference for bare forms rather than a real adult-like use of inflection. However,
German-speaking children behave similarly to English-speaking ones, and these
counterarguments do not apply to German.

Clahsen and Rothweiler (1993), for example, studied spontaneous speech
from 1 year 4 months to 3 years 9 months in nine children and found that
incorrect participle endings in German constituted only 10% of children’s total
participle endings, just as in English. Almost all of these incorrect endings (93%)

were overapplications of the default regular —t. German children also have been

26

shown to use —t overwhelmingly for novel words and incorrect participle endings
in elicited production (Weyerts and Clahsen 1994, Weyerts 1997).

Likewise, with nouns, children predominantly overgeneralize —s, the
default ending. Clahsen et al (1996) found that children presented with low-
frequency words from all five noun classes overused —s in 58% of their errors
(the next most common error accounted for 26% of errors). This is clearly not a
usage proportionate to the approximately 5% of nouns the —s plural actually
applies to. In the same experiment, children were asked to form a compound
noun describing someone who eats the item they had just produced a plural for.
Following the adult constraint on the use of regular plurals within compounds, the
children omitted the overregularized forms in compounds significantly more than
the non-overregularized forms.

1.4.4 Frequency effects

Lexical decision times in German also show similar frequency effects to
those in English: low frequency irregular nouns and verbs are associated with
greater lexical decision times than regulars. Clahsen et al. (1997) describe a
study which involved a word/non-word decision task, measuring RTs for visually
presented verbs and nouns. Regulars and irregulars were matched for
frequency in high and low frequency groups. High frequency irregulars produced
faster RTs than low frequency irregulars, while high and low frequency regulars
showed no distinction in RT.

These data can be explained in terms of a default regular because

memory effects that come into play in irregular inflection when searching for the

27

correct participle form but are not needed to find low-frequency regular forms
computed by a default rule.
1.4.5 Priming

Sonnenstuhl et al. (1999) investigated RTs for cross-modal priming of
German nouns and verbs. Primes were presented orally in participle form while
targets were a visual presentation of the 1St person singular verb form. The
subject was to perform a lexical decision task on the target word. RTs were
compared for (I) identical primes and targets, (ll) morphologically related primes
and targets (, and (III) semantically and phonologically unrelated primes and
targets. Condition (I) represented full priming and (Ill) no priming, so that the
experimental condition (ll) could be compared to both baselines. Verbs were
either regular or irregular.

For regular nouns and verbs, the experimental condition (II) was the same
as (l) and significantly faster than (Ill), showing full priming. Irregular nouns and
verbs showed (ll) still significantly faster than (Ill), but significantly slower than (I),
i.e. partial priming. This full priming for regulars in German is consistent with the
claim that that in regulars stem and affix are decomposed. Partial priming with
irregulars is consistent with the claim that irregulars are in structured lexical
entries so that roots and inflected forms are related more closely than unrelated
words, but still do maintain separate entries. If no structure were present in the
lexical entries, no priming would be predicted here.

Weyerts et al. (1996) have also shown that regular German past

participles preceded by their infinitive show distinctive ERP waveforms

28

characteristic of priming by an identical word, while irregulars show no such
evidence of priming.
1.4.6 Aphasia

In English-speaking aphasics, it is unclear whether a dissociation in
performance on regulars and irregulars actually reflects impaired affixation or a
true effect of regularity. To test this in German where affixes are present on both
regular and irregular forms, Penke et al. (1999) studied eleven German speaking
Broca’s aphasics. Of these, eight were found who had unimpaired performance
for regular participles but overregularized irregular verbs extensively,
demonstrating that aphasia can impair regular or irregular verbs as a class even

when both involve affixes.

1.5 Neuroimaging
1.5.1 Event-related brain potentials (ERP)

ERP’s are small voltage fluctuations measured from a set of points on the
scalp and timed with the presentation of linguistic stimuli. Variables of amplitude,
polarity, location, and latency are used to investigate psycholinguistic issues.
While this method offers limited spatial resolution, its temporal resolution is very
fine.

Using ERP, different brain potentials for German noun regularizations than
for irregularizations have been found by Weyerts et al. (1997). While
regularizations produce a left anterior negativity (LAN) effect, said to be
characteristic of morphosyntactic violations, irregularizations are associated with

an N400-like effect, which resembles the effect of pronounceable non-words and

29

 

is associated with semantic violations. This can be explained under a dual-route
model where regularizations are the misapplication of a morphological rule
adding an affix to a stem present in the lexicon, the irregular base word. Under
that model irregularizations are essentially nonce words, because any affixes or
stem changes associated with them are not stored in the grammar.

Penke et al. (1997) found similar LAN and N400 effects for regularization
and irregularization in German verb past participles. A similar dissociation
between reactions to incorrect regular and irregular forms was also found by
Clahsen (1997). These dissociations are consistent with a dual-route model of
inflection but are not explained in single-route theories. If a unitary method is
used for processing both regular and irregular words no dissociation is predicted
between the effect of over-applying either an irregular or a regular ending. In
particular, the similarity of the effects found for both nouns and verbs in these two
studies suggests that the differences in regularity of the words is the source of
the different effects, and not some accidental features of frequency or word-form.
1.5.2 PET

While a PET study of German inflection by lndefrey et al. (1997) avoided
the confounds associated with English faced by Jaeger et al. (1996), it still used
a block design. Within regular or irregular blocks, a randomized presentation of
frames requiring past tense and participle formation was used to prevent the
formation of possible response strategies. However, it is unclear whether this
innovation completely avoids the confounds of block design or how the

introduction of additional syntax, words, and complexity of task affected the

30

 

results. Thus the twelve areas identified as activated by one set of conditions or
the other are in question. Additionally, the finding of no overlap between areas
activated in regular or irregular sentence frames seems suspect because all
current models of inflection assume at least some memory processes for lexical

access must be common to both irregular and regular inflection.

1.6 Overview of the current study

This study seeks to overcome some of the problems of earlier research by
using event-related fMRI to investigate neural activation of regular and irregular
verbs and nouns in German. Because of its high level of irregularity and parallel
affixation for regulars and irregulars, German seems a system ideally suited to a
connectionist approach. Thus a dichotomy found between regular and irregular
forms in German can be more destructive to a single-route theory than such a
dichotomy is in English. German speaking subjects were presented visually with
a list of regular and irregular verb infinitives and noun singular forms and asked
to produce the participle or plural silently. Data were then analyzed to determine
whether any difference was present between regular and irregular forms in
general. Stimulus words were presented to subjects in isolation.

While presentation within a sentence mimics “real-life” generation to some
extent and is used by some researchers (lndefrey et al. 1997), such a form
introduces too many uncontrolled variables in the extra words to be feasible.
This design introduces no other grammatical processes into the stimulus other
than the one being studied. The presentation of individual words also makes it

possible to identify more precisely the moment at which the linguistic activity

31

being studied is taking place, since fMRI does not have the fine temporal
resolution of other imaging techniques such as ERP. An event-related design
with regular and irregular words presented in random order was used to ensure
that subjects could not use an alternate strategy to complete the task. This is
crucial because it ensures that activation is in fact reflecting processing of
regulars and irregulars - the issue at hand -— and not some other activity.
Functional MRI provides a tool to allow researchers to “ask the brain” how
it processes language more directly than previous experiments were able to.
While neither major theory, in its general form, makes a specific prediction of
what the architecture of the inflectional system in the brain would look like under
MRI, they each suggest a different broad form. In general a single-system theory
would predict that regular and irregular forms would localize to a specific area of
the brain or be similarly diffused throughout cortical areas. In either case, the
prediction would be that regular and irregular forms pattern in the same way.
This contrasts with the general prediction of a dual-route model, which expects a

different distribution of activation for regular and irregular forms.

32

2 Methods

2.1 Subjects

Subjects were eleven native speakers of German, six female and five
male, recruited from among faculty members and students at Michigan State
University. All were righthanded, free of neurological or psychological disorders,
and had no previous knowledge of the cognitive issues involved in the study.
Handedness was assessed using the Edinburgh Handedness Inventory (Oldfield
1971 ). Written, informed consent was obtained in accordance with protocols
approved by the MSU UCRIHS board, lRB # 99-093. Subjects ranged in age
from 23-47 years, mean age 29.5.

In total, three of these subjects were excluded from our final analysis.
One subject was excluded from analysis because he admitted to having fallen
asleep during the scan and did not remember several of the target words during
a review afterwards. Motion correction analysis also indicated movement
suggestive of sleep. The data from a second subject showed artifacts from the
equipment that rendered analysis impossible, and a third subject was excluded
because noun scans showed virtually no measurable activation. The eight
remaining subjects whose data is presented here were four females and four

males, age range 24-45, mean age 29.

2.2 Materials

One test and one practice word list was developed for nouns and one of each for

verbs (see Appendix 2 for the lists of words used). The test lists contained

33

regular and irregular stimulus words matched for token frequency of the inflected
form, following Clahsen et al (1997), using the CELEX database (Baayen et al
1995). The noun paradigm included 12 regular and 12 irregular targets, while the
verb paradigm included 12 regular targets and 12 each of two irregular classes of
verbs. Irregular class A is inflected with the prefix ge- and suffix —en as
described above in section 1.4.1.1. Class B irregulars are inflected in the same
way but involve an additional vowel change in the stem. Because this is a
preliminary study, the two classes of irregular are considered together, though
eventually a separate analysis will be made. Nonetheless, under either a single-
or a dual-route model both classes should pattern together based on a shared
associative process to access irregulars. Practice word lists consisted of 10
words of both regular and various classes of irregular morphology.

Because regular and irregular forms pattern similarly at high frequencies
(Pinker and Prince 1988), low frequency words were chosen. Words in the noun
list all showed a plural frequency no greater than 20 per million and in the verb
list no word had a participle frequency greater than 254 per million. While 254
tokens per million is not generally considered a low frequency, this indicates the
lowest possible frequency to match regular and irregulars. Thus, the word lists
consisted of the lowest set of matched frequency regulars and irregulars

possible.

2.3 Language Protocol

During the functional MRI scan, subjects performed both a verb and a

noun generation task. In the noun task words were presented to the subjects

34

using an event-related paradigm to randomly mix regular and irregular words. A
stimulus word from one of the lists described above was shown to the subject on
a computer screen within the magnet. Stimulus words were displayed
sequentially every 14 seconds for 1 second each in white letters on a black
background. Between these items a fixation cross was displayed. The subjects
were told to silently generate the plural of each word as rapidly as possible while
maintaining accuracy. Each test paradigm was preceded by a practice scan of at
least five words to familiarize the subject with the task. Within the test paradigm
two dummy words preceded the randomized targets to allow the subject time to
adjust to the task. Data collected during presentation of these first two words
were discarded in the analysis.

The verb task was identical in design and subjects were asked to silently
generate the past participle of the word. The order of presentation of the noun
and verb trials was counterbalanced across subjects.

The paradigms were synchronized with MRI data acquisition using a
device, placed on the floor of the magnet room, which detects each RF signal
generated by the EPI pulse sequence. This RF signal was sent via an RF filter to
the PC computer running the experimental paradigm so that the computer could

coordinate timing of stimulus presentation with scan acquisition.

2.4 MR Imaging Protocol
Scans were performed on a 1.5 Tesla clinical scanner in the Department
of Radiology, Michigan State University. A T1-weighted MRI axial scan provided

guidance in locating areas of subjects’ brain anatomy. After auto-shimming, a

gradient echo EPI pulse sequence was used to acquire T2*-weighted images (a
64 X 64 matrix over a 24cm field of view (FOV)). During each language
paradigm, these pulse sequences acquired 21 (or 19) sagittal sections with
7.0mm thickness, spacing 0.0mm, Flip Angle 90 degree, and TE/TR =
50/2000ms.

After the language paradigms were completed 21 (or 19) T1-weighted spin
echo images were collected at the same anatomical locations, with TEfl' R =
14/500ms , flip angle 90 degree, FOV = 24cm, matrix 256x192, and slice
thickness/spacing = 7.0/0.0mm. These images were used for anatomical
orientation during preliminary analysis.

A set of 3D spoiled gradient echo T1-weighted MR images Of the whole
brain were collected, TE/T R = 3.3/8ms, flip angle 30 degree, slice-
thickness/spacing 1.2/0.0mm, 124 slices, matrix size 256x192, and FOV = 24cm.
These images provided a three-dimensional anatomical map used for the

localization of activated pixels in the final stage of data collection.

2.5 Image Processing

Images collected were first corrected for motion by co-registration for both
in-plane translational and rotational movement (Cao et al. 1993). No subjects
exhibited translational movement and none showed more than 1.46 degrees
rotational movement. Each image in the series was registered with the first
baseline image by applying a range of test transformations consisting of planar
translations and rotations and selecting the transformation yielding the minimum

difference in magnitude between the first image and the transformed image.

36

Images associated with regular and irregular forms were sorted out and
separately analyzed. After testing, subjects had been questioned about their
responses to the targets. When errors occurred in which the subjects’ responses
were no longer appropriate to the target’s category (e.g. regularizing an irregular
word) the images corresponding to these targets were included in the category
corresponding to the subjects’ reported responses. There were six cases of this
sort of subject ‘mistake’, five of which had the effect of treating a class-A irregular
verb like a class—B irregular and one- of which gave an irregular plural for a
regular noun.

Statistical analysis of activated pixels was based on a combination of
temporal cross-correlation with a reference function and cluster-size
thresholding. All pixels were thresholded at a level of cc magnitude equal to
0.44, which had a one-tailed type I error of < 0.01 per pixel. Pixels passing the
co threshold were subject to cluster—size thresholding at 23 pixels.

The reference function was generated based on the gamma function S(t)
= A(t-to)6e’("‘°)’T where A is the amplitude, to the time delay, 6 the rise time of the
curve, and T the width. Our reference function was S(t) = (t-1)2'3e““"’°'65, based
on activation in the angular and supramarginal gyri expected because of the
written lexical processing involved in our task but not under investigation here.
Because of the time delay associated with each successive slice as the magnet
proceeds from left to right, a delay proportional to this time lapse was added to

the function to create in effect one reference function for each sagittal section.

37

In order to explore effects related to regularity alone and not merely to
category (noun-verb), noun and verb images were collapsed to obtain an
average regular and irregular image for each subject. Subtractions of these
images were then performed, creating maps showing the voxels only activated
by regular but not irregular words (R-l), irregular but not regular words (l-R), and
those voxels activated in both paradigms (Common).

The activation images were overlaid on the 3Dspgr images using the
research computer software Analysis of Functional Neuroimages (Robert W.
Cox, Medical College of Wisconsin 2000). Localization of the activated pixels in
cortical regions was computed using this software and Human Brain Anatomy in

Computerized Images (Damasio 1995).

38

3 Resuhs

These results discuss both the empirical description of areas of activation '
for regular and irregular conditions and a statistical analysis of the extent of
acfivafion.

3. 1 Empirical Observations

The average results for all cortical areas studied in each of the three
contrasts are shown in tables 1, 2, and 3. Because of possible mismatch in
activation maps due to subtle head movement, a comparison of activation in the
subtraction contrasts (l-R and R-l) and the activation in the common map was
made. It was found that the average total activations for the subtraction
conditions were considerable in relation to that of the common map: average
total l-R activation was 146% that of the common map, R-l 41% that of the
common map. This makes it clear that the activation unique to each of the
conditions (regularity or irregularity) is considerable and not simply the result of
mismatches of a few voxels along the edges of otherwise similar regions. Thus it
is safe to assume that, in terms of the area activated, the difference between
regular and irregular conditions is real and not an artifact.

3.1.1 Common Map

Regions activated by all subjects in the common map were the left
premotor area, left postcentral gyrus, left supramarginal gyrus, and right
supramarginal gyrus. Other areas activated by a more than half of the subjects
were left and right superior, middle and inferior frontal gyri, right premotor area,

left and right precentral gyri, left paracentral gyrus, left and right cingulate gyri,

39

left and right angular gyri, left and right superior parietal lobules, left anterior and
posterior superior temporal gyrus, left posterior middle temporal gyrus, and right
insula. Broca’s area is not found to be consistently activated, though the left

posterior superior temporal gyrus (Wernicke’s area) is activated in most subjects.

 

 

 

 

Region Left Right Total
Superior frontal gyrus 22.88” 20.13” 43.00"
lMiddle frontal gyrus 15.50* 7.75* 23.25"
Inferior frontal gyrus 16.88* 10.75“ 27.63*
Broca's area 5.63 1.50 7.13"
Premotor area 7.75*” 6.00* 13.75““
Precentral gyrus 9.38* 6.25” 15.63 **
Paracentral gyrus 9.50” 4.75 14.25
Cingulate gyrus 7.50* 4.25” 11.75“."
Postcentral gyrus 6.63“ 3.25 9.88"
Supramarginal gyrus 38.63" 15.63" 54.25"
Angular gyrus 9.88* 13.63” 23.50"
Superior parietal Iobule 14.50* 18.38* 32.88*
Precuneus 5.25* 2.50“ 7.75*
Superior Temporal gyrus: Anterior 3.25” 2.00 5.25*
Superior Temporal gyrus: Posterior 4.13* 0.88 500*
Middle Temporal gyrus: Anterior 0.50 0.25 0.75
Middle Temporal gyrus: Posterior 5.00* 3.50 8.50”
Inferior Temporal gyrus: Anterior 0.00 0.75 0.75
Inferior Temporal gyrus: Posterior 6.25 1.63 7.88
Insula 1.25 5.25* 6.50“
Total 190.25 129.00 319.25

 

 

 

Table 1. Average activation in brain regions common to both regular and
irregular data.

** activated in all subjects * activated in a majority of the subjects

3.1.2 l-R Contrast
The I-R contrast shows more activation overall than the common map, as

well as more consistent activation. The lengthy list of the areas activated by all
or most subjects is provided in Table 2. While some of the regions which are
activated reliably in the l-R contrast overlap with those in the common map, it is

important to note that this activation is necessarily different locations within the

40

overlapping regions for the different contrasts because of the subtraction process

used to obtain the maps.

 

 

 

 

 

Region Left Right Total
Superior frontal gyrus 29.75" 33.13" 62.88“
lMiddle frontal gyrus 24.25" 16.50” 40.75"
Inferior frontal gyrus 1525* 16.00" 31.25”
Broca’s area 4.38“ 3.75* 8.13*
Premotor area 9.88“ 9.88" 19.75“
Precentral gyrus 6.13* 1 1.88** 18.00“
Paracentral gyrus 5.25“ 5.63* 10.88**
Cingulate gyrus 17.88* 15.75" 33.63"
Postcentral gyrus 9.13* 16.50“ 25.63*
Supramarginal gyrus 26.25 ** 19.50“ 45.75"
Angular gyrus 11.38" 21 .50* 32.88"
Superior parietal Iobule 17.38" 30.25" 47.63”
Precuneus 10.50* 13.50* 2400*
Superior Temporal gyrus: Anterior 1.63 0.50 2.13
Superior Temporal gyrus: Posterior 363* 4.38 800*
Middle Temporal gyrus: Anterior 0.38 0.75 1.13
Middle Temporal gyrus: Posterior 6.63" 6.88* 13.50“
Inferior Temporal gyrus: Anterior 1.38 0.38 1.75
Inferior Temporal gyrus: Posterior 5.13* 7.38" 12.50"
Insula 10.88" 14.25* 25.13"
Total 217.00 248.25 465.25

 

 

Table 2. Average activation of brain regions in the l-R subtraction condition.

** activated in all subjects * activated in a majority of the subjects
In this condition Broca’s area is activated in the majority of subjects both in the
right and left hemispheres. The posterior superior temporal gyrus is also
activated in the left hemisphere for most subjects.
3.1.3 R-l Contrast

In the R-l condition, the overall activation is much lower than in either of

the previous contrasts, making identification of areas activated across subjects
more difficult. Activation in many anatomical area were close to zero, so it is

possible that variation in extent of activation was below threshold in some

41

subjects. Nonetheless several areas were still consistently activated. These
included the left middle frontal gyrus, the left precentral gyrus, and the left
supramarginal gyrus. The left posterior superior temporal gyrus showed
activation in a majority of subjects. Notably absent is consistent activation in

Broca’s area.

 

 

 

 

Region Left Right Total
Superior frontal gyrus 4.00“ 11.63“ 15.63“
Middle frontal gyrus 4.50““ 7.25“ 11.75““
Inferior frontal gyrus 3.50“ 4.13 7.63“
Broca's area 2.00 1.00 3.00“
Premotor area 3.63“ 0.63 4.25“
Precentral gyrus 11.50““ 1.63“ 13.13“
Paracentral gyrus 2.00 0.75 2.75“
Cingulate gyrus 6.50“ 3.00 9.50“
Postcentral gyrus 7.00“ 4.75“ 1 1 .75*
Supramarginal gyrus 10.13““ 6.50“ 16.63““
Angular gyrus 4.88“ 2.13 7.00“
Superior parietal Iobule 3.88 6.75“ 10.63“
Precuneus 1.00 1.50 2.50
Superior Temporal gyrus: Anterior 0.25 0.00 0.25
Superior Temporal gyrus: Posterior 1.88“ 0.63“ 2.50“
Middle Temporal gyrus: Anterior 0.00 0.00 0.00
Middle Temporal gyrus: Posterior 1.00 0.00 1.00
Inferior Temporal gyrus: Anterior 0.00 0.88 0.88
Inferior Temporal gyrus: Posterior 1.25 0.63 1.88
Insula 4.00“ 4.75“ 8.75“
Total 72.88 58.50 131.38

 

 

 

Table 3. Average activation of brain regions in the R-l subtraction condition.
*“ activated in all subjects * activated in a majority of the subjects

42

Irregular

 

Regular

 

Figure 1. Left and right hemisphere contrasts between regular and irregular
activation in a single subject. Sidebar indicates the level of correlation of activity
with the reference function.

Because Broca’s area and the posterior superior temporal gyrus have
been implicated for activation in grammatical processing in both aphasia and
imaging studies, an ANOVA was performed for each area, but both were shown
to have no significant effects by subtraction contrast.

3.2 Regularity
An omnibus ANOVA was performed on the voxel activation data (before

subtraction) considering three variables as possible sources of variation:

43

hemisphere (left or right), region (anterior or posterior), and regularity (regular or
irregular). Anterior regions were those in the frontal lobs; posterior regions
included those in the parietal lobe and in the temporal lobe, posterior to the
center of the ventricles.
A main effect for regularity was found, with more voxels activated for
irregulars than for regulars, F (1, 62) = 25.29, p < 0.001 (see Figure 2).

Regularity did not interact with any other variables.

Irregular Regular

Subject B

Subject A

 

Figure 2. Regular and irregular paradigms in the left hemisphere of two subjects.
Sidebar indicates the level of correlation of activity with the reference function.

 

 

3.3 Regional Analysis
In order to identify the particular regions that exhibit the greatest

distinction between regulars and irregulars, eight major regions were identified in
each hemisphere, based on general functional similarities and distinctions and,
particular to the present study, commonalities and distinctions in patterns of
activation. Anterior regions included Broca’s area, the precentral gyrus, the
anterior cingulate, and the remaining portions of the prefrontal cortex (i.e., not
including Broca’s area). Posterior regions included the postcentral gyrus, the
insula, the posterior temporal lobe (treated as one unit), and the areas of the
supramarginal gyrus, the angular gyrus, and the superior parietal Iobule
combined as one unit (SMG/AG/SPL). Regions showing a significant effect for
regularity were the right precentral gyrus (F (1,14) = 4.66, p < 0.05), the left
prefrontal cortex (F (1,14) = 9.66, p < 0.01), the right SMG/AG/SPL (F (1,14) =
8.81, p < 0.02), the right posterior temporal lobes (F (1,14) = 8.89, p < 0.01, and
the left insula (F (1,14) = 4.74, p < 0.05). In each of these cases irregular
activation was significantly more widespread than regular.
3.4 Lateralization

The ANOVA also revealed a trend toward a greater activation in the left
hemisphere than in the right, F (1, 62) = 2.72, p = 0.10 (see Figure 1). However,
when the regular and irregular contrasts were examined in relation to the eight
major regions described above, it was found that the regular condition showed a
significant effect for hemisphere (F (1,126) = 4.37, p < 0.05) while the irregular

did not (F (1,126) = 1.15 p = 0.29).

45

4 Discussion
The goal of the study was to determine whether fMRI analysis of participle

and plural formation in German could shed light on the nature of the mechanism
for inflection. Both theoretical paradigms are able in many cases to account for
the same set of facts — it is for this reason that discussion still continues as to
which model best represents the human language system. Proponents of each
theory can in cases use the same data to draw different conclusions. However,
the results of this study show distinctly different patterns of activation for regular
and irregular German inflection which is most consistent with the dual-route
theory of inflection.

4.1 Levels of activation

The distinction in regular and irregular patterns is seen in several areas of
the study. First, irregulars are associated with significantly higher levels of
activation than are regulars. This is a well attested pattern, and seems to point to
a higher processing load involved in irregulars (e.g. Jaeger et al. 1996, lndefrey
et al. 1997). This also corresponds to higher RTs often observed for low-
frequency irregulars (e.g. Seidenberg 1992). Jaeger et al. in fact demonstrate a
correlation between long RTs and extent of neural activation.

Here there is a broad distinction between regular and irregular activation
where single-route connectionist arguments predict none. On a dual process
view, this could be interpreted as showing higher processing costs for memory
retrieval tasks than for symbolic manipulation which is blind to the specifics of
individual tokens or their frequencies. Operating on variables (verb and noun

stems), the regular rule can be seen as less demanding of neural resources than

46

associative processes which are sensitive both to frequency and token
specificity. A connectionist model for English might explain the difficulty of
irregulars in terms of frequency effects, as more common regular endings would
naturally require less processing. In German, however, the predominance of
irregular forms makes such an argument impossible.
4.2 Hemispheric differences

These data indicates a significant difference between regular and irregular
verbs by hemisphere. Areas activated by regulars predictably show more left
hemisphere activation than right, as right-handed individuals are predominantly
left hemisphere dominant for language. However, areas activated during
irregular inflection are distributed bilaterally (no significant difference by
hemisphere. Figure 1 in section 3.1.3 illustrates these hemispheric differences in
activation. If one assumes, following a dual-route account, that the regular forms
utilize a linguistic, rule-based process and that irregulars do not, then it is
reasonable that they would show a the characteristic linguistic left hemisphere
dominance. Purely memory based processes, like the processing of irregular
words, are associated with no such hemisphere preference. This differentiation
in lateralization is inconsistent with a single-route account but follows naturally
from a dual-route account.
4.3 Regions of activation

The pattern of activation common to both regular and irregular forms

shows some agreement with previous studies. In table 4 the localization of

47

activation in the current study is compared with that of three earlier neuroimaging

studies.

Jaeger et al. lndefrey et al. Ullman et al. Current study
1996 1997 1997

Left R ht Left R ht Bilateral Left R' ht

refrontal cortex

u frontal

iddle frontal
nferior frontal

’5 area
remotor area
recentral s
tral
ulate

ma inal
ular

u tal Iobule
nferior Iobule

RGUS
u tem

iddle tern

nferior tern l R
nsula I
Table 4. Areas of activation for regulars and irregulars in previous neuroimaging
studies compared with results from the current study.
I: Activation for irregulars R: Activation for regulars B: Activation for both

 

Previous studies show mixed findings for activation in Broca’s area.
Jaeger et al. (1996) find activation for both regular and irregular verbs in Broca’s
area, while Rhee et al. (1999) only see activation in Broca’s area for regulars.
This data shows the unexpected effect of greater and more consistent activation
for irregulars in Broca’s area than for regulars. Because Broca’s area is
generally associated with grammatical rules (e.g. Grodzinsky 2000), a dual
process account would predict activation of this area particularly for regulars,

while the opposite is observed. Neither would a single-route account explain this

48

finding: according to such an account the activity involved in both regular and
irregular processing should be the same.

This confusing finding may be explained in part by the low level of
activation in the regular paradigms compounded by cluster thresholding after
subtraction may have increased type I error for some regular activation. It is also
possible that the variation in definition of Broca’s area may account for some of
the different findings. The definition of Broca’s area used for localization of
regional activation in this study may have been more conservative than the
definition in other studies.

The left superior temporal gyrus is activated in the common map,
corroborating findings of Pinker et al. (1999) that both regular and irregular
inflection activated this area. Jaeger et al. (1996) points out that a major area of
activation known to be involved in linguistic memory traces is the middle temporal
gyrus. In the present study, as in that of Jaeger et al., this area is found to be
activated consistently in the l-R contrast but not in the R-I contrast. This is
consistent with a dual-route theory which associates memory primarily with
irregular verbs. The middle temporal gyrus is also activated by common areas,
which is predicted by both theories because all indicate that some memory must
be accessed during inflection. Thus the middle temporal activation in the
common area could be accounted for in a dual-route theory by necessary
accessing of the lexicon for the uninflected form, and the activation specific to

irregulars to the accessing of the inflected form. This latter activation would be

49

absent in regular inflection because the inflected form is produced by a rule and
not located in the lexicon.

With regards to the larger major regions discussed in 3.3, it is noteworthy
that every region which showed significant differences between the extent of
regular and irregular activation had significantly greater activation for irregulars
than regulars. This is compatible with the finding that there is more extensive
irregular activation in the brain overall. While it is not clear why particular regions
show more distinction between regular and irregular than others, the fact of such
a distinction is still clearly more compatible with a dual-route than with a single-

route account.

50

5 Conclusion
The findings reported in the present study are clearly preliminary. Further

analysis will determine if differences exist between regular and irregular nouns
and verbs and, at a finer level of analysis, between different classes of irregular
verbs. These analyses may yield a more precise understanding of the nature of
inflection. Nonetheless, even these preliminary findings tell us in gross terms
that the brain behaves differently for regulars than for irregulars. To summarize:
(1) irregulars show greater activation than regulars both overall and, where
significant differences were found, within each of the eight major regions
investigated, (2) regulars show greater lateralization to the left hemisphere than
do irregulars, and (3) brain regions differentially respond to this regular-irregular
distinction, with some regions showing more distinction than others. These
broad findings as most consistent with a dual-route, symbol-manipulating

account of Inflection.

51

REFERENCES

Baayen, H., Piepenbrock, R., & van Rijn, H. 1995. The CELEX lexical database
(CD-ROM). Linguistic Data Consortium, University of Pennsylvania.

Cao, Yue, V. L. Towle, D. N. Levin, and J. M. Balter. 1993. Functional mapping
of human motor cortical activation by conventional MRI at 1.5 T. Journal of
Magnetic Resonance Imaging 3.869-875.

Chomsky, Noam. 1986. Knowledge of Language. NY: Praeger.

Clahsen, Harald. 1993. lnflectional rules in children’s grammars: evidence from
the development of participles in German. Yearbook of Morphology 1992,
ed. by G. Booij and J. van Marie, 1-34. Dordrecht: Kluwer.

Clahsen, Harald. 1997. The representation of participles in the German mental
lexicon: evidence for the dual-mechanism model. Yearbook of Morphology
1996.73-95.

Clahsen, Harald. 1999. Lexical entries and rules of language: a multidisciplinary
study of German inflection. Behavioral and Brain Sciences 22.991-1060.

Clahsen, Harald, S. Eisenbeiss, and Martina Penke. 1996. Lexical learning in
early syntactic development. Generative perspectives on language
acquisition, ed. by Clahsen, Harald. Benjamins.

Clahsen, Harald, S. Eisenbeiss. and l. Sonnenstuhl-Henning. 1997.
Morphological structure and the processing of inflected words. Theoretical
Linguistics 23.201-249.

Clahsen, Harald and Monika Rothweiler. 1993. lnflectional rules in children’s
grammars: evidence from the development of participles in German.
Yearbook of Morphology 199221-34.

Clahsen, Harald, Monika Rothweiler, Andreas Woest, and Gary F. Marcus.
1992. Regular and irregular inflection in the acquisition of German noun
plurals. Cognition 45.225-255.

Damasio, H. 1995. Human brain anatomy in computerized images. Oxford:
Oxford University Press.

Grodzinsky, Y. 2000. The neurology of syntax: Language use without Broca's
area. Behavioral and Brain Sciences 231-71.

52

Hagiwara, Hiroko, Yoko Sugioka, Takane Ito, Mitsuru Kawamura, and Jun-lchi
Shiota. 1999. Neurolinguistic evidence for rule-based nominal suffixation.
Language 75.739-763.

Hagoort, Peter, and Peter lndefrey, Colin Brown, Hans Herzog, Helmuth
Steinmetz, and Rildiger J. Seitz. 1999. The neural circuitry involved in the
reading of German words and pseudowords: a PET study. Journal of
Cognitive Neuroscience 11.383-398.

Hare, M., Elman, J., & Daugherty, K. 1995. Default generalization in
connectionist networks. Language and Cognitive Processes.10, 601-630.

Henderson, E. J. A. 1965. The domain of phonetics (inagural lecture). London.

lndefrey, P., Colin Brown, Peter Hagoort, Hans Herzog, M. Sach, and Rtidiger J.
Seitz. 1997. A PET study of cerebral activation patterns induced by verb
inflection. Neurolmage 5.8548.

Jaeger, Jeri J., Alan H. Lockwood, David L. Kemmerer, Robert D. Van Valin, Jr.,
Brian W. Murphy, and Hanif G. Khalar. 1996. A positron emission
tomographic study of regular and irregular verb morphology in English.
Language 72.451-497.

MacWhinney, Brian and Jared Leinbach. 1991. Implementations are not
conceptualizations: revising the verb learning model. Cognition 40.121-157.

Marcus, Gary P, Steven Pinker, Michael T. Ullman, M. Hollander, T.J. Rosen,
and F. Xu. 1992. Overregularization in language acquisition. Monographs
of the Society for Research in Child Development 57 (4, Serial No. 228).

Marcus, Gary F., Ursula Brinkmann, Harald Clahsen, Richard Wiese, and Steven
Pinker. 1995. German inflection: the exception that proves the rule.
Cognitive Psychology 29.189-256.

Marcus, Gary F. 1998. Rethinking eliminative connectionism. Cognitive
Psychology 37.243-282.

Marcus, Gary F. 1998. The Algebraic Mind: Integrating Connectionism and
Cognitive Science. Cambridge, MA: MIT Press.

Marslen-Wilson, William D., and Lorraine K. Tyler. 1997. Nature 387.592-594.

McClelland, James L., and Mark S. Seidenberg. 2000. Why do kids say goed
and brang? Science 287.47-48.

Mante, Thomas F., Tessa Say, Harald Clahsen, Kolja Schiltz, and Marta Kutas.
1999. Decomposition of morphologically complex words in English: evidence
from event-related brain potentials. Cognitive Brain Research 7.241-253.

53

Oldfield, RC. 1971. The assessment and analysis of handedness: the
Edinburgh Inventory. Neuropsychologia 9.97-113.

Penke, Martina, Helga Weyerts, Matthias Gross, Elke Zander, Thomas F. MUnte,
Harald Clahsen. 1997. How the brain processes complex words: an event-
related potential study of German verb inflections. 1997. Cognitive Brain
Research 6.37-52.

Penke, Martina, Ulrike Janssen, and Marion Krause. 1999. The representation
of inflectional morphology: evidence from Broca’s aphasia. Brain and
Language 68.225-232.

Pinker, Steven. 1991. Rules of language. Science 253.530-535.

Pinker, Steven. 1997. Words and rules in the human brain. Nature 387.547-
548.

Pinker, Steven. 1999. Words and rules: the ingredients of language. New York:
Basic Books.

Pinker, Steven, and Alan Prince. 1988. On language and connectionism:
analysis of a parallel distributed processing model of language acquisition.
Cognition 28.73-193.

Plaut, David C., James L. McClelland, Mark S. Seidenberg, and Karalyn E.
Patterson. 1996. Understanding normal and impaired word reading:
computational principles in quasi-regular domains. Psychological Review
103.56-115.

Plunkett, Kim, and Patrick Juola. 1999. A connectionist model of English past
tense and plural morphology. Cognitive Science 23.463-490.

Plunkett, Kim, and V. Marchman. 1991. U-shaped learning and frequency
effects in a multi-layered perceptron: implications for child language
acquisition. Cognition 38.43-102

Plunkett, Kim, and V. Marchman. 1993. From rote learning to system building:
acquiring verb morphology in children and connectionist nets. Cognition
48.21-69.

Prasada, Sandeep, Steven Pinker, and William Snyder. 1990. Some evidence
that irregular forms are retrieved from memory but regular forms are rule-
generated. Poster presented at the 31St Annual Meeting of the Psychonomic
Society, New Orleans.

54

Pugh, Kenneth R., Bennet A. Shaywitz, Sally E. Shaywitz, Donald P.
Shankweiler, Leonard Katz, Jack M. Fletcher, Pawel Skudlarski, Robert K.
Fulbright, R. Todd Constable, Richard A. Bronen, Cheryl Lacadie, and John
C. Gore. 1997. Predicting reading performance from neuroimaging profiles:
the cerebral basis of phonological effects in printed word identification.
Journal of Experimental Psychology: Human Perception and Performance
23.299-318.

Rhee, Jaemin, Steven Pinker, and Michael T. Ullman. 1999. A
magnetoencephalographic study of English past tense production. Paper
presented at the sixth annual meeting of the Cognitive Neuroscience Society,
Washington, DC.

Rumelhart, David E., and James L. McClelland. 1986. On learning the past
tense of English verbs. Parallel distributed processing: explorations in the
microstructure of cognition, ed. by James L. McClelland, David E. Rumelhart,
and the PDP research group. Volume 2: Psychological and biological
models, 216-271. Cambridge, MA: MIT Press.

Say, Tessa and Harald Clahsen. In Press. Words, rules and stems in the Italian
mental lexicon. In S. Nooteboom, F. Weerman, & F.Wijnen (Eds). Storage
and computation in the language faculty. Amsterdam: Kluwer.

Seidenberg, Mark S. 1992. Connectionism without tears. Connectionism:
theory and practice, ed. by Steven Davis, 84—122. Oxford: Oxford University

Press.

Seidenberg, Mark S. 1997. Language acquisition and use: learning and
applying probabilistic constraints. Science 275.1599-1603.

Seidenberg, Mark S. and James H. Hoeffner. 1997. Evaluating behavioral and
neuroimaging data on past tense processing. Language 74.104-122.

Ullman, Michael T. 1999. Evidence that tool naming and grammatical rule use
are subserved by frontal/basal-ganglia circuits. Paper presented at the sixth
annual meeting of the Cognitive Neuroscience Society, Washington, DC.

Ullman, Michael T., Ruth Bergida, and Kathleen M. O’Craven. 1997. Distinct
fMRI activation patterns for regular and irregular past tense. Neurolmage
5.S549.

Ullman, Michael T., Suzanne Corkin, Marie Coppola, Gregory Hickok, John H. _
Growdon, Walter J. Koroshetz, and Steven Pinker. 1997. A Neural
dissociation within language: evidence that the mental dictionary is part of
declarative memory, and that grammatical rules are processed by the
procedural system. Journal of Cognitive Neuroscience 9.226-276.

55

Westermann, G., & Goebel, R. 1995. Connectionist rules of language.
Proceedings of the seventeenth annual conference of the Cognitive
Science Society, ed. by JD. Moore & J.F. Lehman, 236-241. Mahwah, NJ:
Erlbaum.

Weyerts, Helga, and Harald Clahsen. 1994. Netzwerke und symbolische
Regeln im Spracheniverb: Experimentelle Ergebnisse zur Entwicklung der
Flexionsmorphologie. Linguistiche Berichte 154.430-460.

Weyerts, Helga, Thomas F. Milnte, Henderikus G. O. M. Smid, and Hans-Jochen
Heinze. 1996. Mental representations of morphologically complex words: an
event-related potential study with adult humans. Neuroscience Letters
206.125-128.

Weyerts, Helga. 1997. Regulare und irregulare Flexion: Psycholinguisische und
neurophysiologische Ergebnisse zu Eniverb, Verarbeitung und mentaler
Representation. Dissertation. University of Dilsseldorf.

Weyerts, Helga, Martina Penke, Ulrike Dohrn, Harald Clahsen, and Thomas F.
MUnte. 1997. Brain potentials indicate differences between regular and
irregular German plurals. NeuroReport 8.957-962.

Wunderlich, Dieter, and Ray Fabri. 1995. Minimalist morphology: an approach
to inflection. Zeitschrift filr Sprachwissenschaft 14.236-294.

56

1. Verbs

Ten Practice Items:

APPENDIX

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Two dummy items to get subjects started:

 

 

 

 

Infinitive Participle
Jaresented produced
sehen gesehen
stellen gestellt
gehen gegangen
schliessen geschlossen
tragen getragen
brennen gebrannt
weisen gewiesen
machen gemacht
ziehen gezogen
kommen gekommen

Infinitive Participle
presented produced
geben gegeben
fragen gefragt

 

 

57

 

 

Test items presented randomly

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Participle
Infinitive Participle frequenc

word type presented produced y
1lrregrlarABA braten gebraten 7
lrregularABB kriechen gekrochen 6
Regular knallen geknallt 7
2lrregular ABA fressen gefressen 27
lrregularABB fliessen geflossen 26
Regular tilgen getilgt 27
3IrregularABA blasen geblasen 28
lrregularABB frieren gefroren 28
Regular pflanzen gepflanzt 28
4IrregularABA heissen geheissen 33
lrregularABB fechten gefochten 32
Regular rauben geraubt 33
5lrregularABA graben gegraben 35
lrregularABB weichen gewichen 39
Regular hoffen gehofft 35
6lrregularABA essen gegessen 48
lrregularABB schreiten geschritten 44
Regular Iagern gelagert 49
7lrregular ABA waschen gewaschen 48
lrregularABB leiden gelitten 47
Regular schleppen geschleppt 51
8lrregular ABA schlafen geschlafen 79
lrregularABB streichen gestrichen 78
Regular reisen gereist 79
9Irregular ABA stossen gestossen 171
lrregularABB schiessen geschossen 170
Regular handeln gehandelt 172
10IrregularABA Iaufen gelaufen 216
lrregularABB treiben getrieben 186
Regular zahlen gezahlt 208
11lrregularABA fangen gefangen 221
lrregularABB greifen gegriffen 211
Regular planen geplant 223
12lrregularABA messen gemessen 235
lrregularABB stehen gestanden 254
Regular holen geholt 238

 

58

 

2. Nouns

Ten Practice Items:

 

 

Singular Plural
presented produced
Kind Kinder

 

Komitee Komitees
Wagen Wéigen

 

 

 

 

 

 

 

Tisch Tische

Bild Bilder

Stuhl Stiihle

Ei Eier
Krankheit Krankheiten
Bein Beine

 

 

 

 

Arbeiter Arbeiter

 

Two dummy items to get subjects started:

 

 

 

 

Singular Plural
presented produced
Tasse Tassen
Restaurant Restaurants

 

 

 

59

Test items presented randomly

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

plural
Singular Plural frequenc

word type presented produced y
1lrregular Flirstentum Ftirstentiimer 2
Regular Bikini Bikinis 2
2lrregular Biest Biester 2
Regular Piano Pianos 2
3lrregular Viech Viecher 3
Regular Kobra Kobras 3
4Irregular Maul Mauler 5
Regular Krimi Krimis 6
5lrregular Lamm Lammer 6
Regular Oma Omas 6
6Irregular Altertum AltertiJmer 9
Regular Moskito Moskitos 9
7lrregular Leib Leiber 10
Regular Trend Trends 10
8lrregular Nest Nester 10
Regular Hobby Hobbys 10
9lrregular Gemach Gemacher 10
Regular Deck Decks 10
10lrregular Mund Miinder 11
Regular Cafe Cafes 11
11lrregular Gespenst Gespsenster 16
Regular Flamingo Flamingos 13
12lrregular Lid Lider 20
Regular Karton Kartons 19

 

6O

 

IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII

lllljlllzlljjlliljjljjljljjjllll