From Syllables to Syntax:
Investigating Staged Linguistic Development through Computational Modeling
Kris Jack (kjack@computing.dundee.ac.uk)
Chris Reed (chris@computing.dundee.ac.uk)
Annalu Waller (awaller@computing.dundee.ac.uk)
Division of Applied Computing, University of Dundee,
Dundee, DD1 4HN, Scotland.
Abstract
A new model of early language acquisition is introduced. The
model demonstrates the staged emergence of lexical and
syntactic acquisition. For a period, no linguistic activity is
present. The emergence of first words signals the onset of the
holophrastic stage that continues to mature without syntactic
activity. Syntactic awareness eventually emerges as the result
of multiple lexically-based insights. No mechanistic triggers
are employed throughout development.
Introduction
Children acquire language in stages, first learning words and
later showing sensitivity to their syntactic properties.
Processes that demonstrate distinct behaviors at different
stages of development are difficult to model within a unified
system. As a result, lexical and syntactic processes are often
modeled independently from one another. Bridging the gap
between these models will increase understanding of the
behavioral shift that ushers in syntactic awareness.
Background
Modeling Word-to-meaning Mappings
Children learn the meanings of a small number of words
early in linguistic development. These first words are often
non-formulaic (Wray, 2002). A non-formulaic word
expresses a word-to-meaning relationship that is not a
function of the word's internal parts.
Siskind (1996) investigates word-to-meaning mappings
using cross-situational analysis. Cross-situational analysis
takes advantage of word-meaning co-occurrences to
establish relationships. His simulations show considerable
success, offering a robust solution to the problem under a
variety of circumstances. Steels (2001) considers the
problem of establishing such mappings through language
games. Treating language as a complex adaptive system, he
shows that social pressures to communicate, through games,
encourage the development of a self-organized lexicon.
Lexical acquisition is also studied within a developmental
framework. Regier (2005) shows that interesting lexical
phenomena, such as fast-mapping, can arise without internal
mechanistic changes. Attentional learning plays an
important role in language acquisition.
Modeling the Emergence of Syntax
All natural languages employ syntax. Syntax allows
individuals to both understand and produce novel utterances.
Unlike non-formulaic language, syntactically produced
utterances are a function of their internal parts.
Elman (1993) finds that complex and simple syntactic
structures can be learned by a neural network. If the
network acquires complex structures first then it is incapable
of learning simple structures. He suggests that the input
must be staged with simple structures provided first.
Dominey and Boucher (2005) investigate developmental
phenomena within a grounded robot. Interesting results
arise as grounded <sentence, event> pairs are learned. The
model, however, employs a manual trigger that activates the
syntactic component, an inadequate explanation for the
emergence of syntax. Kirby (2001) considers language
transmission from generation to generation through the
Iterated Learning Model. He demonstrates that transmission
bottlenecks, that determine the amount of linguistic
exposure a learner receives, have an important effect on the
emergence of syntax. The bottleneck can be neither too
narrow nor too wide for syntactic structures to be derived.
Bridging the Gap between Words and Syntax
None of these models show the developmental shift from
lexical to syntactic awareness reflected in child language
development. Jack, Reed and Waller (2004) consider the
transition from the one-word stage to the two-word stage. A
model is trained on <string, meaning> pairs, testing
interpretation of strings at each training epoch. In early
training, a preference for non-formulaic (lexical)
interpretation emerges. As training continues, this
preference fades giving way to formulaic (syntactic)
interpretations. The behavioral change is an emergent
property of the training process and not artificially triggered.
Although a developmental shift is witnessed it appears very
early in the model and the purely lexical period is very short,
unreflective of natural child language development.
Modeling the Developmental Shift
Children do not understand syntactically complex utterances
from birth. First words, produced at around 10-months-old
(Bates & Goodman, 1999), are non-formulaic (Wray, 2002),
with no indication of syntactic properties. By around 18-
months-old, syntactic awareness emerges (MacWhinney &
Bates, 1989). An accurate model of language acquisition
should reflect the development from the holophrastic stage
(non-formulaic) to the early multi-word stage (formulaic),
accounting for the 8 month gap between relative onsets.

The Holophrastic Stage Specification
During the holophrastic stage, the model shows no syntactic
awareness. All successful string-to-meaning mappings are
performed through non-formulaic interpretation i.e. given
the string “all gone”, the appropriate meaning is mapped
directly without reducing the string to its individual parts,
“all” and “gone”.
The Early Multi-word Stage Specification
During the early multi-word stage, the model shows
syntactic awareness. Some successful string-to-meaning
mappings are performed through formulaic interpretation i.e.
given the string “all gone”, it is reduced to its individual
parts, “all” and “gone”. Non-formulaic language persists.
A model is implemented to investigate this developmental
shift. The remainder of the paper describes this model and
discusses its behavior.
The Model
Training Data
The Miniature Language Acquisition framework (Feldman,
Lakoff, Stolcke, & Weber, 1990) allows language
acquisition to be studied by coupling visual events with
linguistic descriptions. Using this framework, a scene
building game is played. An object appears on a scene and
is described. The object always appears next to another
object. These <event, description> pairs are entered into the
system as training data.
Objects are expressed by a set of feature tuples. A
feature tuple expresses a value and an object identifier.
Values are derived from simulated visual data, consistent
with computer vision technology capabilities. Object
identifiers uniquely identify the object that the value belongs
to. Since there are always two objects in an event, they are
numbered 1, and 2. 1 is the first object in the scene while 2
is the second. Objects vary in shape, color and position.
The object {<red, (1)>, <circle, (1)>} reflects that the first
object in the scene is a red circle. Object identification is
present in infants (Kellman, Gleitman, & Spelke, 1987).
Events are expressed by a set of feature tuples comprising
two objects and the relationship between them. The event
{<red, (1)>, <circle, (1)>, <pink, (2)>, <cross, (2)>,
<above, (0)>, <right, (0)>} reflects that a pink cross
appeared to the upper right of a red circle. Relative
positions are expressed as binary relationships along
horizontal and vertical planes, as suggested by infant
interpretations of spatial locations (Quinn, 2003).
Descriptions are syllable-segmented strings. Descriptions
are not word-segmented as fluent speech contains no known
acoustic analog of the blank spaces in text (Brent & Siskind,
2001). A syllabic base is implemented as infants are likely
to represent sound based on a syllable covariant (Dehaene-
Lambertz & Houston, 1998; Mehler, Dupoux, Nazzi, &
Dehaene-Lambertz, 1996). Word spellings are retained for
readability unless a words share syllables e.g. low occurs in
lower and yellow, producing “low er” and “ ye low”.
Training data are randomly generated. Objects can appear
in 10 colors, 10 shapes, and 8 relative locations of one
another. Any combination of the 3 can be generated,
allowing a total of 80,000 different events. Descriptions are
constructed according to the grammar specification in Figure
1. The grammar is instantiated when producing training
data alone and is not accessible by the model during
learning. The grammar is for reader's convenience alone.
Figure 1: The grammar specification for event descriptions.
Overview
The model is designed to investigate the appearance of
lexical and syntactic processes. A set of training data
(<event, description> pairs) are randomly generated and
input into the system. Each pair is analyzed by the Lexical
Analysis Unit. Lexical items are determined from data
regularities through cross-situational analysis (Siskind,
1996). These items are processed by the Syntactic Analysis
Unit that derives syntactic rules and phrasal categories.
Syntactic rules specify the interaction between phrasal
categories.
The Lexical Analysis Unit
Training data are entered into the model in the form of
<event, description> pairs. Lexical items are derived based
on these data. Given that strings are syllable-based, word
boundaries are not provided and must be derived. In some
cases, these word boundaries overlap, increasing ambiguity.
Meaning 'boundaries' must also be derived since not all
feature tuple sets are singletons e.g. below can be
represented as {<below, (0)>, <even_horizontal, (0)>}. The
model must further derive how these strings and meanings
are related to one another.
Assume that the model contains pair (1). On the entry of
pair (2), the model checks if the pair has been encountered
before. If so, then a count is kept of the number of times
that it has appeared and lexical analysis ends. If it has not
been encountered before, then cross-situational analysis
begins, a process that extracts event and string equalities.
The notion behind this process is that words will co-occur
more often with their referents than with other meanings.
Regularities are extracted across events and descriptions
individually before recombining the results.
1. <{<red, (1)>, <circle, (1)>, <pink, (2)>, <cross, (2)>,
<above, (0)>, <right, (0)>},
“a pink cross to the u pper right of the red cir cle”>
2. <{<green, (1)>, <cir cle, (1)>, <red, (2)>, <diamond, (2)
>, <even_vertical, (0)>, <right, (0)>},
“a red dia mond to the right of the green cir cle”>

Event regularities are derived based on feature tuples
equalities. Feature tuple comparisons are value sensitive
and identifier insensitive. That is, the feature tuple <red, (1)
> is equal to any feature tuple with the value red regardless
of identifier value. All feature tuples equalities are extracted
over the two events, producing (3) and (4).
3. {<red, (1)>, <circle, (1)>, <right, (0)>}
4. {<cir cle, (1)>, <red, (2)>, <right, (0)>}
Description comparisons are syllable form sensitive
reflecting infants' sensitivity to syllabic patterns (Houston,
Santelmann, & Jusczyk, 2004). Descriptions are aligned,
(5) and (6), and syllable lists are extracted producing (7) and
(8), both of which are representative of the same string set.
5. “a pink cross to the u pper right of the red cir cle”
6. “a red dia mond to the right of the green cir cle”
7. “a”, “to the”, “right of the”, “red”, and “cir cle”
8. “a”, “red”, “to the”, “right of the”, “cir cle”
Event and description regularities are recombined
producing <{feature tuple}, string> pairs. All combinations
of regularities from the first event and the first description
produce one set of co-occurrences (e.g. <{<red, (1)>,
<circle, (1)>, <right, (0)>}, “a”>), while second event and
second description combinations produce the rest. Each pair
is re-entered into the system and activates the same process
as the original training data.
Cross-situational analysis produces a number of <{feature
tuple}, string> pairs. Often, more than one {feature tuple}
accompanies each string. To avoid ambiguity, each string
must be represented by only one {feature tuple}. Given the
list of {feature tuple}s that a string is related to, the {feature
tuple} with the closest distribution to the string is selected
(times string appears with {feature set} by times string
appears in total, taking the result that is closest to 1.0). In
some cases, a string may be represented by two {feature
tuple}s that are equal. For example, <{<red, (1)>}, “red”>
means that “red” is associated with the redness of object 1
and <{<red, (2)>}, “red”> means that “red” is associated
with the redness of object 2. Feature tuple equality is value
based, regardless of identifier. The relationship is written as
<{<red, (1, 2)>}, “red”> for brevity and represents the
redness of either object.
Each <{feature tuple}, string> pair indicates a syllable
set-to-meaning relationship. If more than one string is
related to the same {feature tuple} then synonymy occurs.
Synonymy is rare in natural language. Children actively
avoid synonymy during language learning, following a
principle of mutual exclusivity (Markman & Wachtel,
1988). The string with the highest probability of being
represented by each unique {feature tuple} is derived. The
most probable <{feature tuple}, string> pairs are stored as
lexical items in the model.
Lexical items are not always representative of adult word-
to-meaning boundaries. Interesting phenomena arise
throughout early training steps. For examples, the word
“red” should be representative of redness in any object. The
model is found to under-generalize words, representing
redness in only one object. Mismatches are also found, such
as <{<circle, (1)>}, “to the”> appear. These phenomena are
indicative of the holophrastic stage in learning, indicating
that children may follow a similar strategy.
The Syntactic Analysis Unit
Non adult-like lexical items can also express syntactic
relationships. Lexical item (9) is clearly a formulaic
function of lexical items (10) and (11). The Syntactic
Analysis Unit is responsible for discovering and encoding
this relationship.
9. <{<red, (1, 2)>, <circle, (1, 2)>}, “red cir cle”>
10. <{<red, (1, 2)>}, “red”>
11. <{<circle, (1, 2)>}, “cir cle”>
Syntactic relationships are discovered within lexical item
triumvirates (as in (9)-(11)). One lexical item, (9), must be
the function of the two others items, (10) and (11). The
lexical items must satisfy both string and {feature tuple}
relationships. Given two strings, the model must produce
the third through string concatenation, i.e. string1+ string2
= string3. Also, given two {feature tuple}s, the model must
produce the third through set union i.e. {feature tuple}1 U
{feature tuple}2 = {feature tuple}3. {Feature tuple}
equality is identifier insensitive, so identifiers need not
match.
Rules capture these relationships. They relate Phrasal
Categories (PCs) to one another by the application of
Transformations (Ts). Each new term is defined before the
rule is presented.
Rules are expressed in the form PC1 = PC2(T1) PC3(T2),
where PC1 is produced by combining the results of PC2,
being transformed by T1, and P3, being transformed by T2.
Phrasal Categories are expressed as the pairing of a set
of strings and a list of feature tuple identifiers, <{string},
(identifier)>. PCs are created to support rule relationships.
There are two kinds of PCs; parent and child. Given the rule
PC1 = PC2(T1) PC3(T2), PC1 is a root, while PCs 2 and 3
are children. Root PCs acquire lexical item 1's data and
identifier end points from Ts 1 and 2. Child PCs are
populated with strings from the original lexical items that
they are derived and the appropriate T start point.
Transformations are expressed as a set of feature tuple
identifier pairs, {feature tuple identifier pair}. Feature tuple
identifier pairs define the mapping from a start point to an
end point, in transforming feature tuple identifiers, <start
identifier, end identifier>.
The Syntactic Analysis unit produces rule (12) from
lexical items (9)-(11).
12. PC1 = PC2(T1) PC3(T2), where
PC1 = <{“red cir cle”}, ((1, 2), (1, 2))>,
PC2 = <{“red”}, ((1, 2))>,
PC3 = <{“cir cle”}, ((1, 2))>,
T1 = {<(1, 2), (1, 2)>} and T2 = {<(1, 2), (1, 2)>}.
Rule (12) expresses a functional path to derive lexical
item (9), using items (10) and (11). It specifies the mapping

from the meaning of items (10) and (11) to producing item
(9). Rule (12) shows how to generate a {feature tuple} that
represents the string “red cir cle”. First, the model searches
for lexical items that represent the child PCs. Lexical items
for “red” and “cir cle” are found; <{<red, (1, 2)>}, “red”>
and <{<circle, (1, 2)>}, “cir cle”> respectively. Each
lexical item is transformed based on its PC's T. The lexical
item for “red” is transformed by T1 and “cir cle” by T2. In
this case <{<red, (1, 2)>}, “red”> becomes <{<red, (1, 2)
>}, “red”> (no change) and <{<circle, (1, 2)>}, “cir cle”>
becomes <{<circle, (1, 2)>}, “cir cle”> (no change). The
results are joined together through set union producing <
{<red, (1, 2)>, <circle, (1, 2)>}, “red cir cle”>.
The Syntactic Analysis Unit analyzes every combination
of lexical item triumvirates and produces a rule for each
group that expresses a syntactic relationship. Rules can
express similar relationships. Rules (13)-(15) all express the
same relationship. Rule (13) is the short-hand version of
rule (12) for improved readability. They state, that “red cir
cle”, “blue cir cle” and “pink dia mond” can each be
produced by applying the same transformation rules to their
children. A transformation rule must have the same start
point and end point to be considered equal.
13. {“red”}((1, 2) -> (1, 2)), {“cir cle”}((1, 2) -> (1, 2))
14. {“blue”}((1, 2) -> (1, 2)), {“cir cle”}((1, 2) -> (1, 2))
15. {“pink”}((1, 2) -> (1, 2)), {“dia mond”}((1, 2) -> (1, 2))
When rules are found to express the same relationship,
they are merged together. Merging rules (13)-(15) produces
(16). (16) has the generative capacity to produce 6 different
strings; “red cir cle”, “blue cir cle”, “pink cir cle”, “red dia
mond”, “blue dia mond”, and “pink dia mond”.
16. {“red”, “blue”, “pink”}((1, 2) -> (1, 2)), {“cir cle”, “dia
mond”}((1, 2) -> (1, 2))
Rule (16) captures the English grammar rule, NP = Adj.
N, where the 'adjective' set contains “red”, “blue”, and
“pink” and the noun set contains “cir cle” and “dia mond”.
The rule states, among other combinations, that when the
string “red” directly precedes the string “dia mond”, a red
diamond is being indicated. To emphasize, the rule does not
just indicate that there is redness in the scene, nor that there
is diamond in the scene, but that there is an object in the
scene that shares both the properties red and diamond.
This result is particularly interesting. From syllable
segmented strings combined with feature based meanings,
English-like grammar rules are derived. Each rule defines a
mapping based not only on individual lexical items, but
groups of lexical items, or PCs, producing syntactic units.
These lexical items are established by drawing word and
meaning boundaries. The PCs are established by drawing
lexical item boundaries. The fixing of these lexical item
boundaries allows the model to treat different words in a
similar way and, ultimately, produce novel relationships
such “red dia mond” in the previous example. Furthermore,
the lexical item boundaries change the model's perception of
lexical status. While lexical analysis produced items such as
“red cir cle”, syntactic analysis draws a boundary through
the string and its related meaning, allowing it to be
deconstructed and reconstructed with the application of
other items. PC role (parent or child) and membership,
therefore, is a better indicator of lexical status than the
lexical items themselves.
Comprehension
The model is tested for evidence of language acquisition
through comprehension tasks. Given a string, the model
must derive a {feature tuple}. Following the example from
the last section, assume that the system contains rule (16)
and has never encountered the string “red dia mond” in
training. The steps involved in comprehending the string
“red dia mond” provide an interesting source for discussion.
PC membership offers a better indication of lexical status
than lexical items. The model searches for the string in all
PCs. If the string appears in a PC then its lexical item
representation is retrieved. If the string does not appear in a
PC then the comprehension process continues regardless.
The model has never encountered the string, so it does not
reside in any PCs.
The model contains rules that specify how to produce
meanings for a number of strings. These rules take two
substrings as input. Using these rules, the string to parse is
dissected into two parts. Any string that contains more than
one syllable can be dissected. The string “red dia mond” is
dissected, by syllable boundaries, producing the pairs
<“red”, “dia mond”> and <“red dia”, “mond”>. Each string
is recursively processed by the comprehension algorithm
detailed in this section. Taking <“red”, “dia mond”> first,
the string “red” is processed revealing that it appears in PC1
and is associated with lexical item <{<red, (1, 2)>}, “red”>.
With similar success, “dia mond” is found to be a member
of PC2 with associated lexical item <{<diamond, (1, 2)>},
“dia mond”>. The string “dia mond” is further dissected
and processed in the same recursive function. Neither “dia”
nor “mond” appear in PCs. With results for “red” (appears
in PC1) and “dia mond” (appears in PC2), the model
searches for a rule that can combine members of these
categories, discovering rule (16). The rule is instantiated to
yield <{<red, (1, 2)>, <diamond, (1, 2)>}, “red dia mond”>.
A possible meaning for the entire string “red dia mond” is,
therefore, {<red, (1, 2)>, <diamond, (1, 2)>}. The
comprehension algorithm searches for additional results
using the alternative dissection, <“red dia”, “mond”>. No
further results are derived. The string “red dia mond” is
correctly identified as {<red, (1, 2)>, <diamond, (1, 2)>}.
In some cases, more than one meaning is derived for a
single string. Each string can map to a non-formulaic result,
through no use of rules, as well as formulaic results, through
the use of rules. Comprehension reintroduces a form of
homonymy into the model. “The red cross” can refer to the
Red Cross Foundation and “the red square” the square in
Moscow just as likely as their geometrically shaped
counterparts found in this study. As long as multiple
meanings provide plausible interpretations for strings, they
are useful. String interpretation should reduce the semantic
burden in communication, not produce a single,
unambiguous interpretation.

As training data are added to the system, lexical items,
rules, and PCs are derived. PCs often include lexical items
that express English like PCs, found in (17)-(19). PC
membership grows as more training data are added. At
times, more than one PC appears to express the same string
set membership, but at different stages of development. For
example, (17) represents the full set of colors available to
the system, while (18) and (19) express subsets of (17).
17. <{“red”, “blue”, “pink”, “green”, “white”, “black”, “ye
low”, “gray”, “lime”, “pur ple”}, ((1, 2))>
18. <{“red”, “white”, “black”, “lime”}, ((1, 2))>
19. <{“ye low”, “gray”, “pur ple”}, ((1, 2))>
During comprehension, PCs are substitutable for one
another if they appear to express the same string member
set, but at different stages of development. (17)-(19) are all
considered substitutable for one another. Given the string
“white”, PCs (17)-(19) are all representative; (17) and (18)
as “white” is a member of their string sets and (19) as it is a
subset of (17).
PC substitutions allow abstract categories such as
adjectives to form faster. During training, it is common for
PCs like (17)-(19) to form. Each of these PCs are created
through the derivation of different rules but all appear to
suggest the inclusion of an adjective. Abstract categories
such as noun, adjective and verb are not necessarily present
in young language learners. Studies show that children
acquire language in an item-based, piecemeal fashion
(Tomasello, 2000). Verb analysis, in particular, shows an
uneven usage. For example, a child may only use the word
“cut” according to the sentence frame “cut ___”, while
“draw” may be used in a variety of manners “draw ___”,
“draw ___ on ___”, “draw ___ for ___”, and “___ draw on
___”. This suggests that the abstract category of verb is not
yet in place, since the verbs are employed with different
constraints. This model reflects a similar 'verb island'
formation but with adjectives and nouns. PC substitutions
allow the islands to be connected relatively easily. Future
studies will focus on varying the levels of PC
substitutability.
Model Behavior
The model is tested to investigate the emergence of the
holophrastic and early multi-word stages. The first correct
non-formulaic (non rule-based) and formulaic (rule-based)
interpretations signal the beginning of the holophrastic and
early multi-word stages respectively.
The Developmental Shift
The model is trained with 10 sets of 30 randomly generated
<event, description> pairs. Results presented are an average
over the 10 sets. At each epoch, the model is tested for
interpretation of 120 strings (10 colors, 10 shapes, 100 color
shape combinations). Each string interpretation yields a set
of possible meanings. Correct meanings are charted in
Figure 2 dependent upon how they are derived (non-
formulaically, or formulaically).
Developmental Shift
0
5
10
15
20
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
No. <event, description>s entered
No
. co
rre
ct p
ars
es
Non-Formulaic
Formulaic
Figure 2: Number of correct non-formulaic and
formulaic interpretations
For 3 epochs, there are no successful string
interpretations. That is, a period of pre-linguistic activity, or
linguistic inactivity characterizes the early training epochs.
The first correct interpretation emerges at epoch 4 and is
non-formulaic. This is the model's first word. It signals the
onset of the holophrastic stage. Being non-formulaic, the
word-to-meaning mapping is representative of first words in
child language development. In one set of data, the model's
first word is “pen ta gon”, appropriately associated with
{pentagon, (1, 2)}. For 10 epochs, lexical insights emerge
with an increasing volume of correct non-formulaic string
interpretations. These strings consistently represent single
words. At epoch 14, the first non-formulaic word
combination is accurately interpreted. That is, it is
interpreted without the use of rules, but as a single unit.
This non-formulaic interpretation of a word combination
spurs syntactic activity. The first formulaic interpretation is
successfully derived at epoch 14, signaling the onset of the
early multi-word stage. The emergence of syntax following
a period of lexical activity is consistent with child language
development.
This result demonstrates two emergent properties in the
model; lexical and syntactic awareness. From the outset, the
model shows no lexical or syntactic awareness. After a
short period of inactivity, lexical awareness emerges,
evidenced by the acquisition of first words. The
holophrastic stage continues unperturbed for a lengthy
period before syntactic awareness emerges. Given a larger
and more varied training set, that is representative of child
linguistic exposure, the period is predicted to lengthen.
Lexical and Syntactic Expressivity
The model is trained with 10 sets of 65 randomly generated
<event, description> pairs. Results presented are an average
over the 10 sets. At each epoch, the model is tested for non-
formulaic interpretation of 20 strings (10 colors, 10 shapes),
and formulaic interpretation of 100 strings (color shape
combinations). Each string interpretation yields a set of
possible meanings. Correct meanings are charted in Figure
3 dependent upon how they are derived (non-formulaically,
or formulaically).

Model Expressivity
0
20
40
60
80
100
0 5 10 15 20 25 30 35 40 45 50 55 60 65
No. <event, description>s entered
%
cor
rec
t p
ars
es
Non-Formulaic
Formulaic
Figure 3: Percentage of correct formulaic and non-
formulaic interpretations
The distinction between non-formulaic and formulaic
language is clear. The former makes no use of rules while
the latter does make use of rules. Formulaic language is
most expressive when rules are applicable to large sets of
data i.e. phrasal category string membership is high. This
model identifies a formulaic relationship at epoch 14. The
relationship is representative of the English grammar rule
NP = Adj. N. On establishing this formulaic expression, the
PCs representing adjectives and nouns constrain rule
expressivity. A correlation between the percentage of
lexical items acquired and the expressivity of the formulaic
expression exists. PC membership swells as subset and
superset relationships are derived, allowing abstract
categories to form.
This result demonstrates that the expressive power of
syntactic rules is correlated with the number of lexical items
correctly identified in the system. As lexical membership
increases, PC string membership expands, and rules become
more expressive. This finding is consistent with child
language acquisition. As phrasal categories form, they
become increasingly abstract and employed by a number of
rules. Given more strict PC connectivity constraints,
Tomasello's (2000) verb island effect is predicted.
Conclusion
This model demonstrates two behavioral shifts that are
present in child language development. First, syllable
combinations are recognized as expressions of word-to-
meaning mappings signaling the onset of lexical activity.
This behavior persists in the absence of syntactic awareness.
Second, word combinations are recognized as expressions of
syntactic relationships. Syntax emerges and becomes
increasingly expressive over time.
Acknowledgments
The first author is sponsored by a studentship from the
EPSRC, UK.
References
Bates, E., & Goodman, J. C. (1999). On the emergence of
grammar from the lexicon. In B. MacWhinney (Ed.), The
emergence of language.
Brent, M. R., & Siskind, J. M. (2001). The role of exposure
to isolated words in early vocabulary development.
Cognition, 81, 33-44.
Dehaene-Lambertz, G., & Houston, D. (1998). Faster
orientation latency toward native language in two-month-
old infants. Language and Speech, 41, 21-43.
Dominey, P. F., & Boucher, J.-D. (2005). Developmental
stages of perception and language acquisition in a
perceptually grounded robot. Cognitive Systems
Research, 6(3), 243-259.
Elman, J. L. (1993). Learning and development in neural
networks: The importance of starting small. Cognition, 48
(1), 71-99.
Feldman, J. A., Lakoff, G., Stolcke, A., & Weber, S. H.
(1990). Miniature language acquisition: A touchstone for
cognitive science. Proc. 12th Ann. Conf. Of CogSci Soc.
Houston, D. M., Santelmann, L. M., & Jusczyk, P. W.
(2004). English-learning infants’ segmentation of
trisyllabic words from fluent speech. Language and
Cognitive Processes, 19(1), 97-136.
Jack, K., Reed, C., & Waller, A. (2004). A computational
model of emergent simple syntax: Supporting the natural
transition from the one-word stage to the two-word stage.
Coling, 20th Conf on Comp. Ling., Geneva, Switzerland.
Kellman, P. J., Gleitman, H., & Spelke, E. S. (1987). Object
and observer motion in the perception of objects by
infants. Journal of Experimental Psychology - Human
Perception and Performance, 13(4), 586-593.
Kirby, S. (2001). Spontaneous evolution of linguistic
structure: An iterated learning model of the emergence of
regularity and irregularity. IEEE Transactions on
Evolutionary Computation, 5(2), 102-110.
MacWhinney, B., & Bates. (1989). The crosslinguisitic
study of sentence processing. New York: Cambridge
University Press.
Markman, E. M., & Wachtel, G. F. (1988). Children's use of
mutual exclusivity to constrain the meanings of words.
Cognitive Psychology, 20(2), 121-157.
Mehler, J., Dupoux, T., Nazzi, T., & Dehaene-Lambertz, G.
(1996). Coping with linguistic diversity: The infant's
viewpoint. In J. L. Morgan & K. Demuth (Eds.), Signal to
syntax. Mahwah, N.J.: Lawrence Erlbaum.
Quinn, P. C. (2003). Concepts are not just for objects:
Categorization of spatial relation information by infants.
In D. H. Rakison & L. M. Oakes (Eds.), Early category
and concept development: Oxford University Press.
Regier, T. (2005). The emergence of words: Attentional
learning in form and meaning. Cognitive Science, 29(5).
Siskind, J. M. (1996). A computational study of cross-
situational techniques for learning word-to-meaning
mappings. Cognition, 61(1-2), 1-38.
Steels, L. (2001). Language games for autonomous robots.
IEEE Intelligent systems, 16-22.
Tomasello, M. (2000). The item-based nature of children’s
early syntactic development. Trends in Cognitive
Sciences, 4(4), 156-163.
Wray, A. (2002). Formulaic language and the lexicon.
Cambridge: Cambridge University Press.