A Neurolinguistic Model of Grammatical
Construction Processing
Peter Ford Dominey
1
, Michel Hoen
1
, and Toshio Inui
2
Abstract
& One of the functions of everyday human language is to com-
municate meaning. Thus, when one hears or reads the sen-
tence, ‘‘John gave a book to Mary,’’ some aspect of an event
concerning the transfer of possession of a book from John to
Mary is (hopefully) transmitted. One theoretical approach to
language referred to as construction grammar emphasizes this
link between sentence structure and meaning in the form of
grammatical constructions. The objective of the current re-
search is to (1) outline a functional description of grammatical
construction processing based on principles of psycholinguis-
tics, (2) develop a model of how these functions can be imple-
mented in human neurophysiology, and then (3) demonstrate
the feasibility of the resulting model in processing languages of
typologically diverse natures, that is, English, French, and Japa-
nese. In this context, particular interest will be directed toward
the processing of novel compositional structure of relative
phrases. The simulation results are discussed in the context of
recent neurophysiological studies of language processing. &
INTRODUCTION
One of the long-term quests of cognitive neuroscience
has been to link functional aspects of language process-
ing to its underlying neurophysiology, that is, to under-
stand how neural mechanisms allow the mapping of the
surface structure of a sentence onto a conceptual repre-
sentation of its meaning. The successful pursuit of this
objective will likely prove to be a multidisciplinary endeav-
or that requires cooperation between theoretical, devel-
opmental, and neurological approaches to the study of
language, as well as contributions from computational
modeling that can eventually validate proposed hypothe-
ses. The current research takes this multidisciplinary
approach within the theoretical framework associated
with grammatical constructions. The essential distinc-
tion within this context is that language is considered to
consist of a structured inventory of mappings between
the surface forms of utterances and meanings, referred
to as grammatical constructions (see Goldberg, 1995).
These mappings vary along a continuum of complexity.
At one end are single words and fixed ‘‘holophrases’’
such as ‘‘Gimme that’’ that are processed as unparsed
‘‘holistic’’ items (see Tomasello, 2003). At the other
extreme are complex abstract argument constructions
that allow the use of sentences like this one. In between
are the workhorses of everyday language, abstract argu-
ment constructions that allow the expression of spatio-
temporal events that are basic to human experience
including active transitive (e.g., John took the car) and
ditransitive (e.g., Mary gave my mom a new recipe)
constructions (Goldberg, 1995).
In this context, the ‘‘usage-based’’ perspective holds
that the infant begins language acquisition by learning
very simple constructions in a progressive development
of processing complexity, with a substantial amount of
early ground that can be covered with relatively mod-
est computational resources (Clark, 2003; Tomasello,
2003). This is in contrast with the ‘‘continuity hypothe-
sis’’ issued from the generative grammar philosophy,
which holds that the full range of syntactic complexity
is available in the form of a universal grammar and is
used at the outset of language learning (see Tomasello,
2000, and comments on the continuity hypothesis de-
bate). In this generative context, the universal grammar
is there from the outset, and the challenge is to under-
stand how it got there. In the usage-based construction
approach, initial processing is simple and becomes
increasingly complex, and the challenge is to explain
the mechanisms that allow full productivity and com-
positionality (composing new constructions from exist-
ing ones). This issue is partially addressed in the current
research.
Theoretical Framework for
Grammatical Constructions
If grammatical constructions are mappings from sen-
tence structure to meaning, then the system must be
capable of (1) identifying the type of grammatical con-
struction for a given sentence and (2) using the identi-
fied construction and its corresponding mapping to
1
CNRS UMR 5015, France,
2
Kyoto University, Japan
D 2006 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 18:12, pp. 2088–2107
extract the meaning of the sentence. Interestingly, this
corresponds to Townsend and Bever’s (2001) two steps
of syntactic processing, that is, (i) parsing of the sen-
tence into phrasal constituents and words (requiring
access to lexical categories and word order analysis) and
(ii) subsequent analysis of phrasal structure and long-
distance dependencies by means of syntactic rules.
Figure 1A illustrates an example of an active transi-
tive sentence and its mapping onto a representation of
meaning. The generalized representation of the corre-
sponding active transitive construction is depicted in
Figure 1B. The ‘‘slots’’ depicted by the NPs and V can
be instantiated by different nouns and verbs in order to
generate an open set of new sentences. For each sentence
corresponding to this construction type, the mapping of
sentence to meaning is provided by the construction.
Figure 1C illustrates a more complex construction that
contains an embedded relative phrase. In this context, a
central issue in construction grammar will concern how
the potential diversity of constructions are identified.
Part of the response to this lies in the specific ways in
which function and content words are structurally orga-
nized in distinct sentence types. Function words (also
referred to as closed class words because of their limited
number in any given language), including determiners,
prepositions, and auxiliary verbs, play a role in defining
the grammatical structure of a sentence (i.e., in specify-
ing who did what to whom) although they carry little
semantic content. Content words (also referred to as
open class words because of their essentially unlimited
number) play a more central role in contributing pieces
of meaning that are inserted into the grammatical
structure of the sentence. Thus, returning to our exam-
ples in Figure 1, the thematic roles for the content
words are determined by their relative position in the
sentences with respect to the other content words and
with respect to the function words. Although this is the
case in English, Bates, McNew, MacWhinney, Devescovi,
and Smith (1982) and MacWhinney (1982) have made
the case more generally, stating that across human lan-
guages, the grammatical structure of sentences is spec-
ified by a combination of cues including word order,
grammatical function words (and or grammatical mark-
ers attached to the word roots), and prosodic structure.
The general idea here is that these constructions are
templates into which a variety of open class elements
(nouns, verbs, etc.) can be inserted in order to express
novel meanings. Part of the definition of a construction
is the mapping between slots or variables in the tem-
plate and the corresponding semantic roles in the
meaning, as illustrated in Figure 1. In this context, a
substantial part of the language faculty corresponds to a
structured set of such sentence-to-meaning mappings,
and these mappings are stored and retrieved based on
the patterns of structural markers (i.e., word order and
function word patterns) unique to each grammatical
construction type. Several major issues can be raised
with respect to this characterization. The issues that we
address in the current research are as follows: (1) Can
this theoretical characterization be mapped onto human
functional neuroanatomy in a meaningful manner and
(2) can the resulting system be demonstrated to account
for a restricted subset of human language phenomena
in a meaningful manner?
Implementing Grammatical Constructions
in a Neurocomputational Model
Interestingly, this characterization of grammatical con-
structions can be reformulated into a type of sequence
learning problem, if we consider a sentence as a se-
quence of words. Determining the construction type for
a given sentence consists in analyzing or recoding the
sentence as a sequence of open class and closed class
elements, and then performing sequence recognition on
this recoded sequence. Dominey (1995) and Dominey,
Arbib, and Joseph (1995) have demonstrated how such
sequence recognition can be performed by a recurrent
prefrontal cortical network (a ‘‘temporal recurrent net-
work’’ [TRN]) that encodes sequential structure (see
also Dominey 1998a, b). Then corticostriatal connec-
tions allow the association of different categories of
sequences represented in the recurrent network with
different behavioral responses.
Once the construction type has thus been identified,
the corresponding mapping of open class elements
onto their semantic roles must be retrieved and per-
formed. This mapping corresponds to what we have
called ‘‘abstract structure’’ processing (Dominey, Lelekov,
Ventre-Dominey, & Jeannerod, 1998). In this context, the
reordering in the form–meaning mapping in Figure 1B
can be characterized as the abstract structure ABCD-
Figure 1. Grammatical construction overview. (A) Specific example
of a sentence-to-meaning mapping. (B) Generalized representation
of the construction. (C) Sentence-to-meaning mapping for sentence
with relativized phrase. (D) Compositional sentence-to-meaning
mapping for sentence with relative phrase in which the relative
phrase has been extracted.
Dominey, Hoen, and Inui 2089
BACD where A-D represent variable slots. In order to
accommodate such abstract structures, rather than rep-
resenting sequences of distinct elements, we modified
the recurrent network model to represent sequences of
variables corresponding to prefrontal working memory
elements (Dominey et al., 1998).
Concretely, from a developmental perspective, we
demonstrated that the resulting abstract recurrent net-
work (ARN) could simulate human infant performance
in distinguishing between abstract structures such as
ABB versus AAB (Dominey & Ramus, 2000) as described
by Marcus et al. (1999). In the grammatical construction
analog, these abstract structures correspond to the
mapping of word order in the sentence onto semantic
arguments in the meaning as illustrated in Figure 1. We
thus demonstrated how the dual TRN/ARN system could
be used for sentence comprehension. The sequence of
closed class elements defining the construction was
processed by the TRN (corresponding to a recurrent
corticocortical network). Then, via modifiable cortico-
striatal synapses, the resulting pattern of cortical activity
recovered the corresponding abstract structure for re-
ordering the open class elements into the appropriate
semantic argument order by the ARN (Dominey, Hoen,
Blanc, & Lelekov-Boissard, 2003; Dominey, 2002).
The model essentially predicted a common neuro-
physiological basis for abstract mappings of the form
BHM-HMB and form-to-meaning mappings such as ‘‘The
ball was hit by Mary’’: Hit(Mary, ball). Concretely, we
predicted that brain lesions in the left perisylvian cortex
that produce syntactic comprehension deficits would
produce correlated impairments in abstract sequence
processing. The first test of this prediction was thus to
compare performance in aphasic patients on these two
types of behavior. We observed a highly significant
correlation between performance on syntactic compre-
hension and abstract structure processing in aphasic
patients, as well as in schizophrenic patients (Dominey
et al., 2003; Lelekov et al., 2000). We reasoned that if this
correlation was due to a shared brain mechanism, then
training in one of the tasks should transfer to improve-
ment on the other. Based on this prediction, we subse-
quently demonstrated that reeducation with specific
nonlinguistic abstract structures transferred to improved
comprehension for the analogous sentence types in
a group of aphasic patients (Hoen et al., 2003). In or-
der to begin to characterize the underlying shared neu-
ral mechanisms, Hoen and Dominey (2000) measured
brain activity with event-related potentials for abstract
sequences in which the mapping was specified by a
‘‘function symbol’’ analogous to function words in sen-
tences. Thus, in the sequences ABCxBAC and ABCzABC,
the two function symbols x and z indicate two distinct
structure mappings. We thus observed a left anterior
negativity (LAN) in response to the function symbols,
analogous to the LAN observed in response to gram-
matical function words during sentence processing
(Hoen & Dominey, 2000), again suggesting a shared
neural mechanism. In order to neuroanatomically local-
ize this shared mechanism, we performed a related set
of brain imagery experiments comparing sentence and
sequence processing using functional magnetic reso-
nance imaging (fMRI). We observed that a common cor-
tical network including Brodmann’s area (BA) 44 was
involved in the processing of sentences and abstract
structure in nonlinguistic sequences, whereas BA 45
was exclusively activated in sentence processing, corre-
sponding to insertion of lexical semantic content into
the transformation mechanism (Hoen, Pachot-Clouard,
Segebarth, & Dominey, 2006).
This computational neuroscience approach allowed
the projection of the construction grammar framework
onto the corticostriatal system with two essential prop-
erties: first, construction identification by corticostriatal
sequence recognition, and second, structure mapping
based on the manipulation of sequences of frontal cor-
tical working memory elements (Dominey & Hoen,
2006). Given the initial validation of this model in the
neurophysiological and neuropsychological studies cited
above, we can now proceed with a more detailed study
of how this can contribute to an understanding of the
possible implementation of grammatical constructions,
as illustrated in Figure 2.
As the sentence is processed word by word, a process
of lexical categorization identifies open and closed class
words. This is not unrealistic, as newborns can perform
this categorization (Shi, Werker, & Morgan, 1999), and
several neural network studies have demonstrated lexi-
cal categorization of this type based on prosodic cues
(Blanc, Dodane, & Dominey, 2003; Shi et al., 1999). In
the current implementation, only nouns and verbs are
recognized as open class words, with the modification of
these by adjectives and adverbs left for now as a future
issue. The meanings of the open class words are re-
trieved from the lexicon (not addressed here, but see
Dominey & Boucher, 2005; Roy, 2002; Dominey, 2000;
Siskind, 1996) and these referent meanings are stored in
a working memory called the PredictedReferentsArray.
The next crucial step is the mapping of these referent
meanings onto the appropriate components of the
meaning structure. In Figure 2, this corresponds to
the mapping from the PredictedReferentsArray onto
the meaning coded in the SceneEventArray. As seen in
Figure 2A and B, this mapping varies depending on the
construction type. Thus, the system must be able to
store and retrieve different FormToMeaning mappings
appropriate for different sentence types, corresponding
to distinct grammatical constructions.
During the lexical categorization process, the struc-
ture of the sentence is recoded based on the local
structure of open and closed class words, in order to
yield a ConstructionIndex that will be unique to each
construction type, corresponding to the cue ensemble
of Bates et al. (1982) in the general case. The closed
2090 Journal of Cognitive Neuroscience Volume 18, Number 12
class words are explicitly represented in the Construc-
tionIndex, whereas open class words are represented as
slots that can take open class words as arguments. The
ConstructionIndex is thus a global representation of the
sentence structure. Again, the requirement is that every
different grammatical construction type should yield a
unique ConstructionIndex. This ConstructionIndex can
then be used as an index into an associative memory to
store and retrieve the correct FormToMeaning mapping.
We have suggested that this mechanism relies on
recurrent cortical networks and corticostriatal process-
ing (Dominey et al., 2003). In particular, we propose that
the formation of the ConstructionIndex as a neural
pattern of activity will rely on sequence processing in
recurrent cortical networks, and that the retrieval of the
FormToMeaning component will rely on a corticostriatal
associative memory. Finally, the mapping from form to
meaning will take place in the frontal cortical region
including BAs 44, 46, and 6. This corresponds to the
SceneEventArray, consistent with observations that
event meaning is represented in this BA 44 pars oper-
cularis region, when events are being visually observed
(Buccino et al., 2004), and when their descriptions are
listened to (Tettamanti et al., 2005). Hoen et al. (2006)
provided strong evidence that for both the processing of
grammatical and nonlinguistic structure processing
rules, this frontal cortical region including BAs 44, 46,
and 6 was activated (see Figure 3), indicating its role in
linguistic and nonlinguistic structural mapping.
The proposed role of basal ganglia in rule storage
and retrieval is somewhat related to the procedural
component of Ullman’s (2001, 2004, 2006) grammar pro-
cessing model in which grammatical rules are encoded
in specific (but potentially domain independent) chan-
nels of the corticostriatal system. Longworth, Keenan,
Barker, Marslen-Wilson, and Tyler (2005) indicate a
more restricted, non-language-specific role of the stria-
tum in language in the selection of the appropriate
mapping in the late integration processes of language
comprehension. Neuropsychological evidence for the
role of the striatum in such rule extraction has been
provided in patients with Huntington’s disease (a form
of striatal dysfunction) that were impaired in rule appli-
cation in three domains: morphology, syntax, and arith-
metic (Teichmann et al., 2005). These data are thus
consistent with the hypothesis that the striatum is
Figure 2. Structure-mapping
architecture. (A) Passive
sentence processing: Step 1,
lexical categorization—
open and closed class words
directed to OpenClassArray
and ConstructionIndex,
respectively. Step 2, open
class words in OpenClassArray
are translated to their
referent meanings via the
WordToReferent mapping.
Insertion of this referent
semantic content into
the Predicted Referents
Array (PRA) is realized in
pars triangularis BA 45.
Step 3, PRA elements are
mapped onto their roles
in the SceneEventArray
by the FormToMeaning
mapping specific to each
sentence type. Step 4,
This mapping is retrieved
from ConstructionInventory
(a corticostriatal associative
memory) via the Construction-
Index (a corticocortical
recurrent network) that
encodes the closed and
open class word patterns
that characterize each
grammatical construction type.
The structure mapping process
is associated with activation of
pars opercularis BA 44. In the
current implementation, neural network associative memory for the ConstructionInventory is replaced by a procedural lookup table. (B) Active
sentence. Note difference in ConstructionIndex and in FormToMeaning.
Dominey, Hoen, and Inui 2091
involved in the computational application of rules in-
cluding our proposed mappings.
In contrast, the integration of lexicosemantic con-
tent into this structure processing machinery, filling of
the PredictedReferentsArray, corresponds to a more
language-specific ventral stream mechanism that culmi-
nates in the pars triangularis (BA 45) of the ventral pre-
motor area, consistent with the declarative component
of Ullman’s (2004) model. In this context, Hoen et al.
(2006) observed that when the processing of sentences
and nonlinguistic sequences was compared, BA 45 was
activated exclusively by the sentence processing.
Goals of the Current Study
We have previously demonstrated that the model in
Figure 2 can learn a variety of constructions that can
then be used in different language interaction contexts
(Dominey & Boucher, 2005a, b; Dominey, 2000). How-
ever, from a functional perspective, a model of gram-
matical construction processing should additionally
address two distinct challenges of (1) cross-linguistic
validity and (2) compositionality, which we will address
in the current study.
Although the examples used above have been pre-
sented in English, the proposed model of grammatical
construction processing should accommodate different
types of languages. Thus, the first objective will be to
demonstrate that the proposed system is capable of
learning grammatical constructions in English, French,
and Japanese. Whereas English and French are relatively
similar in their linguistic structure, Japanese is some-
what different in that it allows more freedom in word
order, with information about thematic roles encoded
in case markers.
Regarding the challenge of compositionality, as pre-
sented above, it appears that for each different type of
sentence, the system must have a distinct construction.
Considering Figure 1C, this indicates that every time a
noun is expanded into a relative phrase, a new construc-
tion will be required. This is undesirable for two reasons:
First it imposes a large number of similar constructions to
be stored, and second, it means that if a sentence occurs
with a relative phrase in a new location, the model will
fail to understand that sentence without first having a
<sentence, meaning> pair from which it can learn the
mapping. Alternatively, the model could process con-
structions within constructions in a compositional man-
ner. Thus, the sentence ‘‘The dog that chased the cat bit
Mary’’ could be decomposed into its constituent phrases
‘‘The dog chased the cat’’ and ‘‘The dog bit Mary.’’ The
goal is to determine whether the model as presented
above can accommodate this kind of compositionality.
During the processing of multiple sentences with this
type of embedded relative clause, it will repeatedly
occur that the pattern ‘‘agent that action object’’ will
occur, and will map onto the meaning component
action(agent, object). The same kind of pattern finding
that allows the association of a ConstructionIndex with
the corresponding FormToMeaning mapping could also
work on such patterns that reoccur within sentences.
That is, rather than looking for patterns at the level of
the entire sentence or ConstructionIndex, the model
could apply exactly the same mechanisms to identify
subpatterns within the sentence. This would allow the
system to generalize over these occurrences such that
when this pattern occurs in a new location in a sentence,
it can be extracted. This will decompose the sentence
into the principal and the relative phrases, with the out-
come that both will more likely correspond to existing
constructions (see Figure 1D). Clearly, this will not work
in all cases of new relative clauses, but more importantly,
it will provide a compositional capability for the system
to accommodate a subclass of the possible new sentence
types. This type of segmentation approach has been
explored by Miikkulainen (1996), discussed below.
METHODS
In the first set of experiments (Experiment 1A–C) for each
of the three languages English, Japanese, and French,
respectively, we will expose the model to a set of <sen-
tence, meaning> pairs in a training phase. We then
validate that the model has learned these sentences
by presenting the sentences alone and comparing the
Figure 3. Comparison of brain activation in sentence processing
and nonlinguistic sequence mapping tasks. Subjects read visually
presented sentences and nonlinguistic sequences presented one
word/element at a time, and performed grammaticality/correctness
judgments after each, responding by button press. Red areas
indicate regions activated by both tasks, including a prefrontal
network that involves BAs 44, 46, 9, and 6. Green areas indicate
cortical regions activated exclusively in the sentence processing
task, including BAs 21, 22, 47, and 45. From Hoen et al. (2006)
with permission.
2092 Journal of Cognitive Neuroscience Volume 18, Number 12
