A Collaborative Tool for the Computational Modelling of Child Language Acquisition
Cognitive Aspects of Computational Language Acquisition EACL (2009)
Available from
Kris Jack's profile on Mendeley.
or
Abstract
A large number of computational language learners have been proposed for modelling the process of child language acquisition. Comparing them, however, can be difficult due to the different assumptions that they make, the diverse test results presented, and the different linguistic behaviours investigated. This paper introduces a toolkit that allows different language learners to be trained, tested and ana lysed under standardised conditions. The results can be easily compared with one another and with typical child language development to highlight the relative advantages and disadantages of learners.
Available from
Kris Jack's profile on Mendeley.
Page 1
A Collaborative Tool for the Computational Modelling of Child Language Acquisition
A Collaborative Tool for the Computational Modelling of Child
Language Acquisition
Kris Jack
CEA LIST, Laboratoire d'ingénerie de la connaissance multimédia multilingue
18 route du Panorama, BP6
FontenayauxRoses, F92265 France
mrkrisjack@gmail.com
Abstract
A large number of computational language
learners have been proposed for modelling the
process of child language acquisition. Com
paring them, however, can be difficult due to
the different assumptions that they make, the
diverse test results presented, and the different
linguistic behaviours investigated. This paper
introduces a toolkit that allows different lan
guage learners to be trained, tested and ana
lysed under standardised conditions. The res
ults can be easily compared with one another
and with typical child language development
to highlight the relative advantages and disad
vantages of learners.
1 Introduction
The computational modelling of language acquis
ition can help in understanding the acquisition
process by estimating the problem faced by chil
dren and designing algorithms that solve it in a
similar way as they do. Many such models have
been produced in recent years, tackling various
linguistic behaviours. Like in any relatively new
domain of research, however, the treatment of
one problem often reveals the presence of several
more that in turn require new solutions of their
own. This has led to the design and implementa
tion of numerous learners that differ in either
subtle or fundamental ways. Given such variety,
it is not yet clear which kind of model, or com
bination of models, can best account for the over
all behaviour witnessed during child language de
velopment.
When surveying the computational language
acquisition literature, the relative advantages and
disadvantages of language learners are not al
ways clear. As such, it can be very difficult to
compare different learners with one another. The
main problem is the lack of standardisation in the
field. Language learners are constructed with
different underlying assumptions, largely due to
the lack of consensus in linguistic theory, are
trained using different data (that can vary from
miniature languages to full blown natural lan
guages) and are tested using different testing
measures (some of which include the 'Looks
good to me' approach).
In this paper, a toolkit for investigating the
computational modelling of child language ac
quisition is proposed. The Language Acquisition
Toolkit (LAT) allows researchers to work collab
oratively in solving the modelling task, while ad
dressing the problems introduced. It is an at
tempt to bring the field forward by creating a
standardised way for testing and implementing
language acquisition learners. The issues ad
dressed in this paper are largely driven by engin
eering concerns although the choices that are
made by the modeller will impact not only on
their learner but also on the associated language
theory. The driving motivation behind the LAT
is that the best way to compare different language
learners is to compare the behaviours that they
produce. The closer a learner's behaviour is to
the behaviour witnessed in children, the better the
model.
The LAT is a computational framework that
can train, test and analyse the linguistic perform
ance of a computational language learner in order
to chart developmental linguistic trajectories.
The motivation for the LAT shall first be ex
plored before describing it in detail, discussing its
features and considering future directions.
Language Acquisition
Kris Jack
CEA LIST, Laboratoire d'ingénerie de la connaissance multimédia multilingue
18 route du Panorama, BP6
FontenayauxRoses, F92265 France
mrkrisjack@gmail.com
Abstract
A large number of computational language
learners have been proposed for modelling the
process of child language acquisition. Com
paring them, however, can be difficult due to
the different assumptions that they make, the
diverse test results presented, and the different
linguistic behaviours investigated. This paper
introduces a toolkit that allows different lan
guage learners to be trained, tested and ana
lysed under standardised conditions. The res
ults can be easily compared with one another
and with typical child language development
to highlight the relative advantages and disad
vantages of learners.
1 Introduction
The computational modelling of language acquis
ition can help in understanding the acquisition
process by estimating the problem faced by chil
dren and designing algorithms that solve it in a
similar way as they do. Many such models have
been produced in recent years, tackling various
linguistic behaviours. Like in any relatively new
domain of research, however, the treatment of
one problem often reveals the presence of several
more that in turn require new solutions of their
own. This has led to the design and implementa
tion of numerous learners that differ in either
subtle or fundamental ways. Given such variety,
it is not yet clear which kind of model, or com
bination of models, can best account for the over
all behaviour witnessed during child language de
velopment.
When surveying the computational language
acquisition literature, the relative advantages and
disadvantages of language learners are not al
ways clear. As such, it can be very difficult to
compare different learners with one another. The
main problem is the lack of standardisation in the
field. Language learners are constructed with
different underlying assumptions, largely due to
the lack of consensus in linguistic theory, are
trained using different data (that can vary from
miniature languages to full blown natural lan
guages) and are tested using different testing
measures (some of which include the 'Looks
good to me' approach).
In this paper, a toolkit for investigating the
computational modelling of child language ac
quisition is proposed. The Language Acquisition
Toolkit (LAT) allows researchers to work collab
oratively in solving the modelling task, while ad
dressing the problems introduced. It is an at
tempt to bring the field forward by creating a
standardised way for testing and implementing
language acquisition learners. The issues ad
dressed in this paper are largely driven by engin
eering concerns although the choices that are
made by the modeller will impact not only on
their learner but also on the associated language
theory. The driving motivation behind the LAT
is that the best way to compare different language
learners is to compare the behaviours that they
produce. The closer a learner's behaviour is to
the behaviour witnessed in children, the better the
model.
The LAT is a computational framework that
can train, test and analyse the linguistic perform
ance of a computational language learner in order
to chart developmental linguistic trajectories.
The motivation for the LAT shall first be ex
plored before describing it in detail, discussing its
features and considering future directions.
Page 2
2 Background
The process of modelling child language acquisi
tion is very complex, as many of the first at
tempts confirmed (Feldman et al., 1990; Suppes,
Liang & Bottner, 1991). Rather than modelling
the process in entirety, an undoubtedly daunting
task, modellers took the simplified approach of
focusing upon individual linguistic behaviours,
leading to much research into relatively con
strained problems such as understanding over
and undergeneralisation errors (Plunkett, Sinha,
Moller & Strandsby, 1992), single word learning
(Regier, 2005), syntactic category acquisition
(Redington, Chater & Finch, 1988) and past
tense learning (Rumelhart & Mcclelland, 1986).
While such models have led to valuable insights
in the domain, it can be difficult to see how each
of them is related to one another given the lack of
standardised learning, testing and analysis.
Often, the variety found in computational
models reflects the divisions between linguistic
theories pertaining to child language acquisition
(Kaplan, Oudeyer & Bergen, 2008). Given that
linguists remain divided about how children learn
language, it is not surprising to find a similar di
vision in the computational modelling com
munity as well. One of the fundamental issues
that separates modellers is the kind of data that
the learner learns. This can range from the use of
plain textual data (Elman, 1993), to grounded
sensorbased input (Roy, 2008). Standardising
the type of learning data would thus be useful for
comparing language learners.
Typical computational models are often tested
under different circumstances and using different
techniques. For example, while some papers of
fer a general analysis of the model's behaviour,
others focus on particular features, while some
test language comprehension, others test lan
guage production, and while some consider de
velopmental growth, others consider only the
start and end points of training. Although this is
often justifiable in the context of the research
problem, it makes it difficult to directly compare
two models. It would be useful to put all models
through the same set of rigorous tests in order to
find out how they are similar and how they differ
from one another. Such standardised testing will
often reveal important differences that may have
previously been hidden.
Practically, however, not all models that are
described in the literature are made available for
download. As a result, researchers often have to
spend time recreating models. This assumes, of
course, that the model has been described in
enough detail that it can be faithfully recreated.
Much time could be saved if such models were
available for download, from a common reposit
ory, such as the Weka makes machine learning
algorithms freely available in a software work
bench (http://www.cs.waikato.ac.nz/~ml/).
A good language learner should not just solve
language learning problems, but should do so in a
similar way as is witnessed in children. Based on
psycholinguistic evidence, several linguistic
timetables have been derived containing import
ant linguistic milestones (Brown, 1973; Ingram,
1989; Pinker, 1994; Tomasello, 2005). The char
acter of language development is a significant
feature in child language acquisition and model
lers should be encouraged to model it to better
understand the process. A language learner that
demonstrates a good use of syntax at the same
time as producing its first words is not very real
istic. Instead, there should be a prolonged period
in which words are learned followed by the emer
gence of syntax. Unfortunately, a language mod
el can often produce behaviours at unexpected
times, signalling a problem with the linguistic
theory that it embodies. A standardised approach
to analysing the linguistic development of a lan
guage learner would be an advantage.
3 The Language Acquisition Toolkit
3.1 Introduction
The Language Acquisition Toolkit (LAT) is a
piece of software that allows researchers to de
velop and test computational language learners
within a standardised environment. The LAT's
target users are researchers who have basic skills
in software development and are comfortable us
ing the programming language Java. It assumes
that the language learner operates under the re
strictions imposed in the miniature language
paradigm (Feldman et al., 1990). The LAT can
be obtained from www.langac.com and is avail
able under a GNU public license meaning that
the code can be reproduced and modified without
obtaining permission.
The LAT is an attempt to standardise the train
ing, testing and analysing of language learners
within an open an accessible environment (Figure
1). In training, the language learner observes a
The process of modelling child language acquisi
tion is very complex, as many of the first at
tempts confirmed (Feldman et al., 1990; Suppes,
Liang & Bottner, 1991). Rather than modelling
the process in entirety, an undoubtedly daunting
task, modellers took the simplified approach of
focusing upon individual linguistic behaviours,
leading to much research into relatively con
strained problems such as understanding over
and undergeneralisation errors (Plunkett, Sinha,
Moller & Strandsby, 1992), single word learning
(Regier, 2005), syntactic category acquisition
(Redington, Chater & Finch, 1988) and past
tense learning (Rumelhart & Mcclelland, 1986).
While such models have led to valuable insights
in the domain, it can be difficult to see how each
of them is related to one another given the lack of
standardised learning, testing and analysis.
Often, the variety found in computational
models reflects the divisions between linguistic
theories pertaining to child language acquisition
(Kaplan, Oudeyer & Bergen, 2008). Given that
linguists remain divided about how children learn
language, it is not surprising to find a similar di
vision in the computational modelling com
munity as well. One of the fundamental issues
that separates modellers is the kind of data that
the learner learns. This can range from the use of
plain textual data (Elman, 1993), to grounded
sensorbased input (Roy, 2008). Standardising
the type of learning data would thus be useful for
comparing language learners.
Typical computational models are often tested
under different circumstances and using different
techniques. For example, while some papers of
fer a general analysis of the model's behaviour,
others focus on particular features, while some
test language comprehension, others test lan
guage production, and while some consider de
velopmental growth, others consider only the
start and end points of training. Although this is
often justifiable in the context of the research
problem, it makes it difficult to directly compare
two models. It would be useful to put all models
through the same set of rigorous tests in order to
find out how they are similar and how they differ
from one another. Such standardised testing will
often reveal important differences that may have
previously been hidden.
Practically, however, not all models that are
described in the literature are made available for
download. As a result, researchers often have to
spend time recreating models. This assumes, of
course, that the model has been described in
enough detail that it can be faithfully recreated.
Much time could be saved if such models were
available for download, from a common reposit
ory, such as the Weka makes machine learning
algorithms freely available in a software work
bench (http://www.cs.waikato.ac.nz/~ml/).
A good language learner should not just solve
language learning problems, but should do so in a
similar way as is witnessed in children. Based on
psycholinguistic evidence, several linguistic
timetables have been derived containing import
ant linguistic milestones (Brown, 1973; Ingram,
1989; Pinker, 1994; Tomasello, 2005). The char
acter of language development is a significant
feature in child language acquisition and model
lers should be encouraged to model it to better
understand the process. A language learner that
demonstrates a good use of syntax at the same
time as producing its first words is not very real
istic. Instead, there should be a prolonged period
in which words are learned followed by the emer
gence of syntax. Unfortunately, a language mod
el can often produce behaviours at unexpected
times, signalling a problem with the linguistic
theory that it embodies. A standardised approach
to analysing the linguistic development of a lan
guage learner would be an advantage.
3 The Language Acquisition Toolkit
3.1 Introduction
The Language Acquisition Toolkit (LAT) is a
piece of software that allows researchers to de
velop and test computational language learners
within a standardised environment. The LAT's
target users are researchers who have basic skills
in software development and are comfortable us
ing the programming language Java. It assumes
that the language learner operates under the re
strictions imposed in the miniature language
paradigm (Feldman et al., 1990). The LAT can
be obtained from www.langac.com and is avail
able under a GNU public license meaning that
the code can be reproduced and modified without
obtaining permission.
The LAT is an attempt to standardise the train
ing, testing and analysing of language learners
within an open an accessible environment (Figure
1). In training, the language learner observes a
Page 3
simulated world in which actionbased events oc
cur. Both simulated descriptions and visual data
are sent to the language learner for analysis. The
LAT then tests both the language learner's com
prehension and production capacities. Compre
hension is tested by sending a description to the
language learner and scoring the visual data that
are produced. Similarly, production is tested by
sending visual data to language learner and scor
ing the descriptions produced. The LAT then
analyses the results obtained from testing and de
velops data describing the learner's development.
Figure 1: LAT Overview. A language learner is
placed in the LAT's simulated world where it
learns from simulated audio and visual data. The
LAT tests the learner and the results are used to
produce data describing its development.
3.2 Training
The LAT can be configured to train different
language learners by generating a simulated en
vironment in which actionbased events occur.
The simulated environment operates within the
miniature language acquisition paradigm (Feld
man et al., 1990), a simplified simulation of the
realworld. A simulation is employed rather than
grounding the model in the realworld in order to
better control the number and type of problems
that are being investigated in a single experiment.
While the miniature language paradigm imposes
a number of constraints, the proposed simulation
contains enough complexity to justify its use.
The learner is trained by watching an event
that is simulated in the blocks world in which a
number of geometric objects can be found.
When an event occurs, a symbolic representation
of the description and visual data are generated.
More concretely, an event is the pairing of a sim
ulated description and a action, e=〈d , a〉 .
Events are represented following evidence from
child studies. First, it is assumed that the learner
can establish a triadic relationship between an
object, a speaker and themselves in order to asso
ciate a description with an action. This kind of
relationship is typically called jointattention and
does not appear in children until around 12
monthsold (Tomasello, 1995). As such, the
symbolic content present in descriptions and ac
tions are limited to those found in child literature
during the first year of life.
An infant's acoustic sensitivity is so attuned
that from fourdaysold she demonstrates the
ability to differentiate between native and non
native speech (Mehler et al., 1988). Such dis
crimination lies in rhythmic properties that differ
over language groups (DehaeneLambertz &
Houston, 1998; Mehler, Dupoux, Nazzi & De
haeneLambertz, 1996) and is likely to be syl
lablebased since infants detect change in syllable
quantity, but not in phoneme quantity over
samples of speech (BijeljacBabic, Bertoncini &
Mehler, 1993). Infants also detect vowel change,
a syllable covariant, more readily than consonant
change (Bertoncini et al., 1988), further support
ing a syllabic base. A description is thus repres
ented as a nonzero length ordered list of syl
lables in the LAT. Word segmentations are not
included as there is no acoustic equivalent of the
blank space in written language.
In terms of visual sensitivity, infants can
identify objects through retinal and object dis
placement during motion from four monthsold
(Kellman, Gleitman & Spelke, 1987), and make
relative spacial distinctions between left and
right, and above and below, from three to ten
months old (Quinn & Schyns, 2003). Infants can
also make use of shape and colour to differentiate
between objects in the first year of life (Landau,
Smith & Jones, 1988). The LAT thus describes
the physical properties of objects that inhabit the
blocks world (e.g. shape, colour, size and posi
tion), referred to as features. An action is defined
as a nonzero length ordered list of feature sets,
where each feature set is associated with a unique
time interval. A set of features describes all ob
jects that can be seen in an event. Note that ac
tions in this terminology do not relate to actions
in terms of verbs in natural language, but to a list
of descriptions of scenes. Properties such as
Analysing
Training
comprehension
Testing
+
production
Language Learner
Developmental
Data
cur. Both simulated descriptions and visual data
are sent to the language learner for analysis. The
LAT then tests both the language learner's com
prehension and production capacities. Compre
hension is tested by sending a description to the
language learner and scoring the visual data that
are produced. Similarly, production is tested by
sending visual data to language learner and scor
ing the descriptions produced. The LAT then
analyses the results obtained from testing and de
velops data describing the learner's development.
Figure 1: LAT Overview. A language learner is
placed in the LAT's simulated world where it
learns from simulated audio and visual data. The
LAT tests the learner and the results are used to
produce data describing its development.
3.2 Training
The LAT can be configured to train different
language learners by generating a simulated en
vironment in which actionbased events occur.
The simulated environment operates within the
miniature language acquisition paradigm (Feld
man et al., 1990), a simplified simulation of the
realworld. A simulation is employed rather than
grounding the model in the realworld in order to
better control the number and type of problems
that are being investigated in a single experiment.
While the miniature language paradigm imposes
a number of constraints, the proposed simulation
contains enough complexity to justify its use.
The learner is trained by watching an event
that is simulated in the blocks world in which a
number of geometric objects can be found.
When an event occurs, a symbolic representation
of the description and visual data are generated.
More concretely, an event is the pairing of a sim
ulated description and a action, e=〈d , a〉 .
Events are represented following evidence from
child studies. First, it is assumed that the learner
can establish a triadic relationship between an
object, a speaker and themselves in order to asso
ciate a description with an action. This kind of
relationship is typically called jointattention and
does not appear in children until around 12
monthsold (Tomasello, 1995). As such, the
symbolic content present in descriptions and ac
tions are limited to those found in child literature
during the first year of life.
An infant's acoustic sensitivity is so attuned
that from fourdaysold she demonstrates the
ability to differentiate between native and non
native speech (Mehler et al., 1988). Such dis
crimination lies in rhythmic properties that differ
over language groups (DehaeneLambertz &
Houston, 1998; Mehler, Dupoux, Nazzi & De
haeneLambertz, 1996) and is likely to be syl
lablebased since infants detect change in syllable
quantity, but not in phoneme quantity over
samples of speech (BijeljacBabic, Bertoncini &
Mehler, 1993). Infants also detect vowel change,
a syllable covariant, more readily than consonant
change (Bertoncini et al., 1988), further support
ing a syllabic base. A description is thus repres
ented as a nonzero length ordered list of syl
lables in the LAT. Word segmentations are not
included as there is no acoustic equivalent of the
blank space in written language.
In terms of visual sensitivity, infants can
identify objects through retinal and object dis
placement during motion from four monthsold
(Kellman, Gleitman & Spelke, 1987), and make
relative spacial distinctions between left and
right, and above and below, from three to ten
months old (Quinn & Schyns, 2003). Infants can
also make use of shape and colour to differentiate
between objects in the first year of life (Landau,
Smith & Jones, 1988). The LAT thus describes
the physical properties of objects that inhabit the
blocks world (e.g. shape, colour, size and posi
tion), referred to as features. An action is defined
as a nonzero length ordered list of feature sets,
where each feature set is associated with a unique
time interval. A set of features describes all ob
jects that can be seen in an event. Note that ac
tions in this terminology do not relate to actions
in terms of verbs in natural language, but to a list
of descriptions of scenes. Properties such as
Analysing
Training
comprehension
Testing
+
production
Language Learner
Developmental
Data
Page 4
push and pull are thus not explicitly represented
as symbolic features.
Two types of events can occur in the blocks
world: actionbased; and descriptive. In the case
of an actionbased event, an object performs an
action while in the case of a descriptive event,
objects do not change. As a result, actionbased
events contain different feature sets, giving the
impression of change, while descriptive events
contain the same feature sets, indicating no
change. The description in an actionbased event
describes the action while the description in a de
scriptive event describes an object in the static
scene. Objects can perform several actions in
cluding moving, flashing, growing, shrinking, ap
pearing, disappearing, destroying another object,
hitting another object, pushing another object and
pulling another object.
The LAT randomly generate events that can be
used as training data. It can create objects, make
them perform actions, and describe the events by
instantiating appropriate grammar fragments. To
encourage the use of standardised sets of training
data, a number of sets of data have been ran
domly generated that each contain 10,000 events.
These data have been generated from different
parameters (e.g. amount of noise, probability that
an object will perform an action in an event,
probabilities for each action to occur, number of
time intervals for an event, number of
colours/shapes/sizes/actions possible) with differ
ent language properties (e.g. recursion present/
not present, number of rules, language in use).
To provide concrete examples of typical LAT
training data, one data set, called the Appearance
data set will be presented in detail. The appear
ance data set is inspired from a study with real
participants. Participants sat in front of a com
puter screen that initially showed a blank white
screen. They were asked to describe all changes
that were made to the screen in enough detail that
a stranger could recreate the scene using only
their descriptions. By pressing a key on the key
board, a new geometric object appeared on the
screen and the change was described by the parti
cipant. While the addition of an object to a scene
appears to be a trivial change, participants pro
duced complex linguistic descriptions that re
vealed a deep knowledge of their language. For
example, descriptions such as “a blue circle ap
peared to the upper right of the green square at
the bottom” and “a red circle appeared between
the four squares making the shape of a cross”
(Jack, 2005).
Given the complexity of the language pro
duced, a simplified version the task was construc
ted in which only the appearance of one object
next to another object was considered. By re
stricting the context, there is less demand for a
computational language learner to have a rich se
mantic representation of scenes. This served as a
reasonable starting point from which to conduct
the investigation. The actions in the Appearance
data set were constructed by randomly generating
one object and placing it in the middle of a 3x3
grid scene and then adding a second object,
which was also randomly generated, in a differ
ent position. Eight colours and shapes were used.
Each action was also accompanied by an appro
priate description that was generated using a
grammar fragment (Figure 2).
E NP→ 1 PAR 2
PAR1 NP→ 1 PART PAR2 REL NP→ 2
RELT REL Det→ 2 REL REL→ 1 | REL2
REL1 → a bove | be low |
to the REL4
REL2 → REL3 REL4
REL3 to the low er→ | to
the u pper
REL4 → left of | right of
NP1 Det→ 1 Nbar NP2 Det→ 2 Nbar
Nbar SHP COL→
Det1 a→ Det2 the→
COL black | blue | grey→
| green | pink | black | red
| ye low
SHP cir cle | cross | dia→
mond | heart | rec tang gle
| star |square | tri ang gle
Figure 2: Miniature Language from Appearance
Data Set. All strings are syllable segmented rather
than word segmented.
Events from this data set have actions that are
described using a 2frame time interval, where
the first set of features describes the state of the
scene before the action occurs and the second set
of features describes the scene after the action oc
curs (Figure 3). Note that it is assumed that the
learner can identify concepts such as colour,
shape and position and that such symbolic in
formation is associated with a particular object.
The notion of objecthood, where the first object
in the scene is O1 and the second object is O2, is
carried across time intervals with O1 being recog
as symbolic features.
Two types of events can occur in the blocks
world: actionbased; and descriptive. In the case
of an actionbased event, an object performs an
action while in the case of a descriptive event,
objects do not change. As a result, actionbased
events contain different feature sets, giving the
impression of change, while descriptive events
contain the same feature sets, indicating no
change. The description in an actionbased event
describes the action while the description in a de
scriptive event describes an object in the static
scene. Objects can perform several actions in
cluding moving, flashing, growing, shrinking, ap
pearing, disappearing, destroying another object,
hitting another object, pushing another object and
pulling another object.
The LAT randomly generate events that can be
used as training data. It can create objects, make
them perform actions, and describe the events by
instantiating appropriate grammar fragments. To
encourage the use of standardised sets of training
data, a number of sets of data have been ran
domly generated that each contain 10,000 events.
These data have been generated from different
parameters (e.g. amount of noise, probability that
an object will perform an action in an event,
probabilities for each action to occur, number of
time intervals for an event, number of
colours/shapes/sizes/actions possible) with differ
ent language properties (e.g. recursion present/
not present, number of rules, language in use).
To provide concrete examples of typical LAT
training data, one data set, called the Appearance
data set will be presented in detail. The appear
ance data set is inspired from a study with real
participants. Participants sat in front of a com
puter screen that initially showed a blank white
screen. They were asked to describe all changes
that were made to the screen in enough detail that
a stranger could recreate the scene using only
their descriptions. By pressing a key on the key
board, a new geometric object appeared on the
screen and the change was described by the parti
cipant. While the addition of an object to a scene
appears to be a trivial change, participants pro
duced complex linguistic descriptions that re
vealed a deep knowledge of their language. For
example, descriptions such as “a blue circle ap
peared to the upper right of the green square at
the bottom” and “a red circle appeared between
the four squares making the shape of a cross”
(Jack, 2005).
Given the complexity of the language pro
duced, a simplified version the task was construc
ted in which only the appearance of one object
next to another object was considered. By re
stricting the context, there is less demand for a
computational language learner to have a rich se
mantic representation of scenes. This served as a
reasonable starting point from which to conduct
the investigation. The actions in the Appearance
data set were constructed by randomly generating
one object and placing it in the middle of a 3x3
grid scene and then adding a second object,
which was also randomly generated, in a differ
ent position. Eight colours and shapes were used.
Each action was also accompanied by an appro
priate description that was generated using a
grammar fragment (Figure 2).
E NP→ 1 PAR 2
PAR1 NP→ 1 PART PAR2 REL NP→ 2
RELT REL Det→ 2 REL REL→ 1 | REL2
REL1 → a bove | be low |
to the REL4
REL2 → REL3 REL4
REL3 to the low er→ | to
the u pper
REL4 → left of | right of
NP1 Det→ 1 Nbar NP2 Det→ 2 Nbar
Nbar SHP COL→
Det1 a→ Det2 the→
COL black | blue | grey→
| green | pink | black | red
| ye low
SHP cir cle | cross | dia→
mond | heart | rec tang gle
| star |square | tri ang gle
Figure 2: Miniature Language from Appearance
Data Set. All strings are syllable segmented rather
than word segmented.
Events from this data set have actions that are
described using a 2frame time interval, where
the first set of features describes the state of the
scene before the action occurs and the second set
of features describes the scene after the action oc
curs (Figure 3). Note that it is assumed that the
learner can identify concepts such as colour,
shape and position and that such symbolic in
formation is associated with a particular object.
The notion of objecthood, where the first object
in the scene is O1 and the second object is O2, is
carried across time intervals with O1 being recog
Page 5
nised as the same object before and after an ac
tion occurs.
Before action (t=1) After action (t=2)
O1: square
O1: blue
O1: x2
O1: y2
O1: square
O1: blue
O1: x2
O1: y2
O2: circle
O2: yellow
O2: x3
O2: y3
a ye low cir cle to the u pper right of the blue
square
Figure 3: Sample Event from Appearance Data
Set. Two time frames represented graphically and
as feature sets. The accompanying syllable seg
mented description of the event is also shown.
The remaining data sets contain more complex
events in which more actions and richer mini
ature languages are employed. Actions are ran
domly generated, with respect to the constraints
imposed on the data set (e.g. number of colours,
shapes, and actions) and appropriate descriptions
are generated. These descriptions are produced
by following a heuristic that minimises the num
ber of syllables that can appear in a single de
scription. This reduces the production of unnat
ural sentences. For example, take the case where
an object appears in a scene amongst 10 other ob
jects. A description could be generated to de
scribe the action with respect to one other object,
two other objects or as many as 10 other objects.
While such descriptions are all valid, many of
them would sound unnatural if employed. The
algorithm selects descriptions by favouring those
that have fewer syllables. A parser is then em
ployed that eliminates invalid descriptions that
can be misinterpreted. By making a parsimoni
ous use of syllables, more natural descriptions
tend to be produced. More abstract language can
also be found such as the use of the word 'bully
ing' to describe pushing, pulling and hitting.
3.3 Testing
The LAT monitors the linguistic development
of a language leaner by testing its comprehension
and production capacities. The learner's compre
hension and production are tested at every round
of training.
For each set of training data, there is an associ
ated set of testing data, ensuring a standardised
test procedure for language learners. Test data is
produced using grammar rules for producing de
scriptions and heuristics for producing actions.
The tests are constructed to reflect the properties
found in the training data's miniature language.
As such, the learner is only tested on the kind of
descriptions and actions that it has the opportun
ity to learn through observing events. Con
cretely, a testing set is a set of events where each
event relates one or more descriptions to one or
more actions. The set of testing data associated
with the Appearance training data set can be used
to test the learner's vocabulary, certain multi
word combinations and full sentences. Using the
terminology found in Appearance's grammar
fragment (Figure 2), the LAT tests for the com
prehension of shapes (SHP), colours (COL), ob
jects (Nbar), indefinite objects (NP1), definite ob
jects (NP2) and events (E).
In testing the learner's comprehension, the
LAT sends a description as input and receives a
set of actions as output. The output is automatic
ally scored by comparing it with the expected
output that is associated with the description.
Actions are compared based on the feature values
that are relevant to the given description. Given
the description “a ye low cir cle to the u pper
right of the blue square” (Figure 3), the colours,
shapes and relative positions of the objects are
relevant whereas their exact positions are not.
The LAT equally accepts a yellow circle that ap
pears higher or further right than its idealised po
sition with respect to the blue square, as long as
the relative positions remain correct.
Borrowing from research in child language ac
quisition studies, four kinds of incorrect re
sponses are identified: overextended; underex
tended; mismatched; or incorrect. For example,
the meaning of the description “square” is under
extended if the learner only uses it to refer to red
squares, blue squares and green squares, but not
to squares of other colours. Similarly, the mean
ing of the description “red square” is overexten
ded if it refers to red squares, blue squares and
red circles. A mismatch is found if the descrip
tion “square” is used to refer to objects other than
squares, for examples circles and triangles, but
never to squares themselves. Results that deviate
from these cases are simply considered incorrect.
3
2
1
1 2 3
3
2
1
1 2 3
tion occurs.
Before action (t=1) After action (t=2)
O1: square
O1: blue
O1: x2
O1: y2
O1: square
O1: blue
O1: x2
O1: y2
O2: circle
O2: yellow
O2: x3
O2: y3
a ye low cir cle to the u pper right of the blue
square
Figure 3: Sample Event from Appearance Data
Set. Two time frames represented graphically and
as feature sets. The accompanying syllable seg
mented description of the event is also shown.
The remaining data sets contain more complex
events in which more actions and richer mini
ature languages are employed. Actions are ran
domly generated, with respect to the constraints
imposed on the data set (e.g. number of colours,
shapes, and actions) and appropriate descriptions
are generated. These descriptions are produced
by following a heuristic that minimises the num
ber of syllables that can appear in a single de
scription. This reduces the production of unnat
ural sentences. For example, take the case where
an object appears in a scene amongst 10 other ob
jects. A description could be generated to de
scribe the action with respect to one other object,
two other objects or as many as 10 other objects.
While such descriptions are all valid, many of
them would sound unnatural if employed. The
algorithm selects descriptions by favouring those
that have fewer syllables. A parser is then em
ployed that eliminates invalid descriptions that
can be misinterpreted. By making a parsimoni
ous use of syllables, more natural descriptions
tend to be produced. More abstract language can
also be found such as the use of the word 'bully
ing' to describe pushing, pulling and hitting.
3.3 Testing
The LAT monitors the linguistic development
of a language leaner by testing its comprehension
and production capacities. The learner's compre
hension and production are tested at every round
of training.
For each set of training data, there is an associ
ated set of testing data, ensuring a standardised
test procedure for language learners. Test data is
produced using grammar rules for producing de
scriptions and heuristics for producing actions.
The tests are constructed to reflect the properties
found in the training data's miniature language.
As such, the learner is only tested on the kind of
descriptions and actions that it has the opportun
ity to learn through observing events. Con
cretely, a testing set is a set of events where each
event relates one or more descriptions to one or
more actions. The set of testing data associated
with the Appearance training data set can be used
to test the learner's vocabulary, certain multi
word combinations and full sentences. Using the
terminology found in Appearance's grammar
fragment (Figure 2), the LAT tests for the com
prehension of shapes (SHP), colours (COL), ob
jects (Nbar), indefinite objects (NP1), definite ob
jects (NP2) and events (E).
In testing the learner's comprehension, the
LAT sends a description as input and receives a
set of actions as output. The output is automatic
ally scored by comparing it with the expected
output that is associated with the description.
Actions are compared based on the feature values
that are relevant to the given description. Given
the description “a ye low cir cle to the u pper
right of the blue square” (Figure 3), the colours,
shapes and relative positions of the objects are
relevant whereas their exact positions are not.
The LAT equally accepts a yellow circle that ap
pears higher or further right than its idealised po
sition with respect to the blue square, as long as
the relative positions remain correct.
Borrowing from research in child language ac
quisition studies, four kinds of incorrect re
sponses are identified: overextended; underex
tended; mismatched; or incorrect. For example,
the meaning of the description “square” is under
extended if the learner only uses it to refer to red
squares, blue squares and green squares, but not
to squares of other colours. Similarly, the mean
ing of the description “red square” is overexten
ded if it refers to red squares, blue squares and
red circles. A mismatch is found if the descrip
tion “square” is used to refer to objects other than
squares, for examples circles and triangles, but
never to squares themselves. Results that deviate
from these cases are simply considered incorrect.
3
2
1
1 2 3
3
2
1
1 2 3
Page 6
The LAT can score both single words and
phrases based on these categories.
In addition, the output produced by the learner
can also be described using the standard informa
tion retrieval measures of precision, recall, and
the emeasure which is a weighted combination
of the two former values (van Rijsbergen, 1979).
The process of testing the learner's production
is similar to that of testing comprehension.
Rather than the LAT sending a description as in
put, however, it sends an action. The learner then
produces a set of descriptions as output. Results
from production are scored using the same prin
ciples as applied during comprehension. That is,
the learner's output is compared to the expected
output and it is scored as either correct, overex
tended, underextended, mismatched or incorrect.
3.4 Analysing
Both the comprehension and production res
ults that are produced from testing are used to
evaluate the learner's linguistic stage of develop
ment. Several types of analysis have been de
signed to ease learner comparisons: roundbased;
trialbased; and learnerbased. Roundbased ana
lyses analyse the results produced from a single
round of testing. Trialbased analyses take
roundbased statistics and compare them with
previous rounds in order to find behavioural
trends in the data. Finally, learnerbased ana
lyses compare trialbased data for several trials in
order to extract general behavioural trends. By
performing analyses at all three levels of detail, a
more complete account of the learner's behaviour
is produced.
The LAT is currently able to perform a num
ber of roundbased analyses that are often found
in the literature: summary of test results in terms
of correct results and errors; chart the linguistic
generativity of the learner; and present evidence
of syntactic activity.
Roundbased analyses produce results that are
then used to determine the model's stage of lin
guistic development using data from child lan
guage studies: prelinguistic; holophrastic; early
multiword; late multiword; and abstract stages.
A number of trialbased analyses are per
formed using these data, in order to identify par
ticular linguistic behaviours: linguistic develop
ment; vocabulary acquisition; comprehension/
production imbalance. With the creation of a lin
guistic development timetable, all data can also
be presented in terms of stages. For example, the
number of words that are correctly comprehen
ded and the rate of vocabulary acquisition can be
shown by stage.
Modelbased analyses can performed when the
results from several trials are available. Each of
the results, such as the rate of vocabulary acquisi
tion during a stage, are compared across trials to
identify general behavioural trends.
The LAT thus offers a standardised platform
for training, testing and analysing language mod
els. The results from all analyses can be auto
matically compared to determine the differences
between learners and which learner best fits child
language data.
4 Discussion
The LAT is a freely available tool that offers a
standardised environment from which language
modellers can develop their language learners. It
is an attempt to advance the domain by offering a
platform where common goals can be focussed
upon in a collaborative environment. It aims to
standardise the training, testing and analysing of
language learners by understanding the needs of
language modellers through collaboration.
By using the LAT, the language modeller ac
cepts the need to work with standardised training
data. Such standardisation is widespread in com
putational linguistics. For example, in the field
of automatic text classification, there are several
databases of preclassified documents (e.g. Reu
ters21578, Reuters Corpus Volume 1 and 20
Newsgroups) that researchers can use to evaluate
different algorithms and to compare their results.
The LAT offers different sets of training data that
are constrained by principles of the miniature
language paradigm. In using such data, the mod
elling task differs from the task that a child faces
in a number of ways. In particular, the learning
problem is simplified in that the realworld con
tains many more objects and that natural lan
guage has far more linguistic structures and
words than the language fragments. It is for
these reasons, however, that such a paradigm is
attractive. Many language learning problems can
be effectively investigated by first simplifying
the problem and then developing solutions.
When such problems in the miniature language
paradigm have been adequately solved, it is en
visaged that the LAT can be grounded in a real
phrases based on these categories.
In addition, the output produced by the learner
can also be described using the standard informa
tion retrieval measures of precision, recall, and
the emeasure which is a weighted combination
of the two former values (van Rijsbergen, 1979).
The process of testing the learner's production
is similar to that of testing comprehension.
Rather than the LAT sending a description as in
put, however, it sends an action. The learner then
produces a set of descriptions as output. Results
from production are scored using the same prin
ciples as applied during comprehension. That is,
the learner's output is compared to the expected
output and it is scored as either correct, overex
tended, underextended, mismatched or incorrect.
3.4 Analysing
Both the comprehension and production res
ults that are produced from testing are used to
evaluate the learner's linguistic stage of develop
ment. Several types of analysis have been de
signed to ease learner comparisons: roundbased;
trialbased; and learnerbased. Roundbased ana
lyses analyse the results produced from a single
round of testing. Trialbased analyses take
roundbased statistics and compare them with
previous rounds in order to find behavioural
trends in the data. Finally, learnerbased ana
lyses compare trialbased data for several trials in
order to extract general behavioural trends. By
performing analyses at all three levels of detail, a
more complete account of the learner's behaviour
is produced.
The LAT is currently able to perform a num
ber of roundbased analyses that are often found
in the literature: summary of test results in terms
of correct results and errors; chart the linguistic
generativity of the learner; and present evidence
of syntactic activity.
Roundbased analyses produce results that are
then used to determine the model's stage of lin
guistic development using data from child lan
guage studies: prelinguistic; holophrastic; early
multiword; late multiword; and abstract stages.
A number of trialbased analyses are per
formed using these data, in order to identify par
ticular linguistic behaviours: linguistic develop
ment; vocabulary acquisition; comprehension/
production imbalance. With the creation of a lin
guistic development timetable, all data can also
be presented in terms of stages. For example, the
number of words that are correctly comprehen
ded and the rate of vocabulary acquisition can be
shown by stage.
Modelbased analyses can performed when the
results from several trials are available. Each of
the results, such as the rate of vocabulary acquisi
tion during a stage, are compared across trials to
identify general behavioural trends.
The LAT thus offers a standardised platform
for training, testing and analysing language mod
els. The results from all analyses can be auto
matically compared to determine the differences
between learners and which learner best fits child
language data.
4 Discussion
The LAT is a freely available tool that offers a
standardised environment from which language
modellers can develop their language learners. It
is an attempt to advance the domain by offering a
platform where common goals can be focussed
upon in a collaborative environment. It aims to
standardise the training, testing and analysing of
language learners by understanding the needs of
language modellers through collaboration.
By using the LAT, the language modeller ac
cepts the need to work with standardised training
data. Such standardisation is widespread in com
putational linguistics. For example, in the field
of automatic text classification, there are several
databases of preclassified documents (e.g. Reu
ters21578, Reuters Corpus Volume 1 and 20
Newsgroups) that researchers can use to evaluate
different algorithms and to compare their results.
The LAT offers different sets of training data that
are constrained by principles of the miniature
language paradigm. In using such data, the mod
elling task differs from the task that a child faces
in a number of ways. In particular, the learning
problem is simplified in that the realworld con
tains many more objects and that natural lan
guage has far more linguistic structures and
words than the language fragments. It is for
these reasons, however, that such a paradigm is
attractive. Many language learning problems can
be effectively investigated by first simplifying
the problem and then developing solutions.
When such problems in the miniature language
paradigm have been adequately solved, it is en
visaged that the LAT can be grounded in a real
Page 7
environment where vast volumes of data are
available for processing.
The results from learning can then be tested
using a standardised set of tests. The learner is
treated as a black box, meaning that the LAT
evaluates its output alone without entering into
its inner workings. This helps to keep the LAT's
functionality independent from the learner by fo
cussing on the way in which it behaves rather
than how it produces particular behaviours, simil
ar to the relationship found between the linguist
and child in the real world. By testing both com
prehension and production on a large set of de
scriptions and actions, a complete picture of the
learner's linguistic state can be derived. The
LAT also checks for language errors such as
overextensions, underextensions and mis
matches. Individual results are made available to
the researcher in a tabular format as well as
providing overall recall, precision and emeasure
scores.
By standardising the test results, different lan
guage learners can be easily compared with one
another. The LAT can analyse these results to
discover behavioural trends in the data with can
be used in further comparisons. It is also inter
esting to note that the LAT makes an attempt to
compare the behaviour produced by a language
learner with that of children. Inspired by child
language development timetables, a set of mile
stones has been derived that are used to charac
terise the learner's behaviour in terms of stages.
The LAT attempts to encourage researchers to
consider the developmental behaviour of their
language learners over time.
It is important to note that the LAT is a work
in progress. This disclaimer is likely to remain
true for many years. Developing a gold standard
is a difficult task and one that risks to evolve over
time. The LAT should be regarded as a proposal
for standardisation. Being a collaborative pro
ject, any contributor can challenge this proposal
by offering their own solutions. Contributors are
encouraged to create their own data and al
gorithms and to upload them to the LAT. A gold
standard can only emerge from the selections that
are made by other modellers, who vote by using
certain data and algorithms in their own model
ling tasks. In this sense, the proposed instanti
ation of the LAT described in this article is less
important than the idea behind the LAT itself.
5 Future Considerations
In designing the LAT, it quickly became clear
that the task was not straightforward. Designing
a tool that can make useful and standardised
comparisons between language learners is a com
plex task. A balancing act between not excluding
certain types of learners and creating a con
strained, manageable environment is not without
its difficulties. As such, it is worth considering
future developments for the LAT. While still in a
preliminary state of development, it is hoped that
a collaborative approach to the task will allow it
to be steered in the directions that are best adap
ted to its potential users. A number of these dir
ections are now considered.
The miniature language paradigm is at the
heart of the LAT. This language can be extended
to include more complex linguistic constructions
and a larger vocabulary. It is suggested that a
systematic approach is followed in which the
learning task is made progressively complex by
adding linguistic features that tend to be wit
nessed in children during development. It seems
reasonable to follow a longitudinal approach to
development. Contributors are also encouraged
to create and submit new training data sets in or
der to explore how complex a miniature language
can become.
The type of information that is available to the
learner could also be changed. At present, the
descriptions lack acoustic information such as
tone. Such data is indispensable in investigating
certain languages such as Mandarin and Swahili.
Similarly, the symbolic representations of visual
objects can be refined to better represent reality.
Colours can be represented by RGB values rather
than linguisticallyrelated symbols, as it is un
likely that children start with such predefined se
mantic categories from the outset of learning.
It is also worthwhile considering more com
plex testing and analysis algorithms. It is likely
that they will be developed in step with new lin
guistic phenomena that are investigated, building
a useful catalogue of tools. In addition, it may be
useful to develop learnerdependant analysis
tools in order to demonstrate how the inner work
ings are related to the outward behaviour.
Finally, it is hoped that the LAT will become a
useful resource not just for modellers who are
comfortable with coding but also nonprogram
mers. They should be able to implement and ex
periment with different kinds of models with the
available for processing.
The results from learning can then be tested
using a standardised set of tests. The learner is
treated as a black box, meaning that the LAT
evaluates its output alone without entering into
its inner workings. This helps to keep the LAT's
functionality independent from the learner by fo
cussing on the way in which it behaves rather
than how it produces particular behaviours, simil
ar to the relationship found between the linguist
and child in the real world. By testing both com
prehension and production on a large set of de
scriptions and actions, a complete picture of the
learner's linguistic state can be derived. The
LAT also checks for language errors such as
overextensions, underextensions and mis
matches. Individual results are made available to
the researcher in a tabular format as well as
providing overall recall, precision and emeasure
scores.
By standardising the test results, different lan
guage learners can be easily compared with one
another. The LAT can analyse these results to
discover behavioural trends in the data with can
be used in further comparisons. It is also inter
esting to note that the LAT makes an attempt to
compare the behaviour produced by a language
learner with that of children. Inspired by child
language development timetables, a set of mile
stones has been derived that are used to charac
terise the learner's behaviour in terms of stages.
The LAT attempts to encourage researchers to
consider the developmental behaviour of their
language learners over time.
It is important to note that the LAT is a work
in progress. This disclaimer is likely to remain
true for many years. Developing a gold standard
is a difficult task and one that risks to evolve over
time. The LAT should be regarded as a proposal
for standardisation. Being a collaborative pro
ject, any contributor can challenge this proposal
by offering their own solutions. Contributors are
encouraged to create their own data and al
gorithms and to upload them to the LAT. A gold
standard can only emerge from the selections that
are made by other modellers, who vote by using
certain data and algorithms in their own model
ling tasks. In this sense, the proposed instanti
ation of the LAT described in this article is less
important than the idea behind the LAT itself.
5 Future Considerations
In designing the LAT, it quickly became clear
that the task was not straightforward. Designing
a tool that can make useful and standardised
comparisons between language learners is a com
plex task. A balancing act between not excluding
certain types of learners and creating a con
strained, manageable environment is not without
its difficulties. As such, it is worth considering
future developments for the LAT. While still in a
preliminary state of development, it is hoped that
a collaborative approach to the task will allow it
to be steered in the directions that are best adap
ted to its potential users. A number of these dir
ections are now considered.
The miniature language paradigm is at the
heart of the LAT. This language can be extended
to include more complex linguistic constructions
and a larger vocabulary. It is suggested that a
systematic approach is followed in which the
learning task is made progressively complex by
adding linguistic features that tend to be wit
nessed in children during development. It seems
reasonable to follow a longitudinal approach to
development. Contributors are also encouraged
to create and submit new training data sets in or
der to explore how complex a miniature language
can become.
The type of information that is available to the
learner could also be changed. At present, the
descriptions lack acoustic information such as
tone. Such data is indispensable in investigating
certain languages such as Mandarin and Swahili.
Similarly, the symbolic representations of visual
objects can be refined to better represent reality.
Colours can be represented by RGB values rather
than linguisticallyrelated symbols, as it is un
likely that children start with such predefined se
mantic categories from the outset of learning.
It is also worthwhile considering more com
plex testing and analysis algorithms. It is likely
that they will be developed in step with new lin
guistic phenomena that are investigated, building
a useful catalogue of tools. In addition, it may be
useful to develop learnerdependant analysis
tools in order to demonstrate how the inner work
ings are related to the outward behaviour.
Finally, it is hoped that the LAT will become a
useful resource not just for modellers who are
comfortable with coding but also nonprogram
mers. They should be able to implement and ex
periment with different kinds of models with the
Page 8
flexibility of looking at different aspects of ac
quisition under different settings and with differ
ent types of data. They can then inform language
modellers directly about how particular language
models perform well and poorly in certain cases.
The collaborative aspect of the LAT encourages
not just programmers to share their code, but for
everyone to share their ideas.
6 Conclusion
This article proposes a tool that facilitates the
consolidation of research into the computational
modelling of child language acquisition under the
miniature language paradigm. The workshop is
being used to launch a first version of the LAT,
that is hoped to help language modellers and
child language experts to communicate and share
their knowledge.
7 Acknowledgements
This research was supported by the JeanLuc
Lagardère Foundation (http://www.fondation
jeanluclagardere.com).
Many thanks to the anonymous reviewers for
their constructive comments.
References
Bertoncini, J., BijeljacBabic, R., Jusczyk, P., Kennedy, L.
& Mehler, J. (1988). An investigation of Young infants'
perceptual representations of speech sounds. Journal of
experimental psychology, 117, pp. 2133.
BijeljacBabic, R., Bertoncini, J. & Mehler, J. (1993). How
Do 4DayOld Infants Categorize Multisyllabic Utter
ances?. Developmental Psychology, 29, pp. 711721.
Brown, R. (1973). A First Language: The Early Stages.
Harvard University Press.
DehaeneLambertz, G. & Houston, D. (1998). Faster Ori
entation Latencies Toward Native Language in Two
MonthOld Infants. Language and Speech, 41, pp.
2143.
Elman, J. (1993). Learning and development in neural net
works: the importance of starting small. Cognition, 48,
pp. 7199.
Feldman, J., Lakoff, G., Stolcke, A. & Weber, S. (1990,).
Miniature Language Acquisition: A Touchstone for
Cognitive Science.
Ingram, D. (1989). First Language Acquisition: Method,
Description and Explanation. Cambridge University
Press.
Jack, K. (2005,). Introducing a Scene Building Game to
Model Early First Language Acquisition, CLUK,
Manchester, England.
Kaplan, F., Oudeyer, P. & Bergen, B. (2008). Computa
tional models in the debate over language learnability.
Infant and Child Development, 17, pp. 5580.
Kellman, P., Gleitman, H. & Spelke, E. (1987). Object and
observer motion in the perception of objects by infants.
Journal of experimental psychology. Human perception
and performance, 13, pp. 586593.
Landau, B., Smith, L. & Jones, S. (1988). The importance
of shape in early lexical learning. Cognitive Develop
ment, 3, pp. 299321.
Mehler, J., Dupoux, T., Nazzi, T. & DehaeneLambertz, G.
(1996). Coping with linguistic diversity: The infant's
viewpoint. Lawrence Erlbaum.
Mehler, J., Jusczyk, P., Lambertz, G., Halsted, N., Berton
cini, J. & AmielTison, C. (1988). A precursor of lan
guage acquisition in young infants. Cognition, 29, pp.
143178.
Pinker, S. (1994). The Language Instinct. William Mor
row.
Plunkett, K., Sinha, C., Moller, M. & Strandsby, O. (1992).
Symbol grounding or the emergence of symbols?
Vocabulary growth in children and a connectionist net.
Connection Science, 4, pp. 293312.
Quinn, P. & Schyns, P. (2003). What goes up may come
down: perceptual process and knowledge access in the
organization of complex visual patterns by young in
fants. Cognitive Science, 27, pp. 923935.
Redington, M., Chater, N. & Finch, S. (1988). Distribu
tional Information: A powerful cue for acquiring syn
tactic categories. Cognitive Science, 22, pp. 425469.
Regier, T. (2005). The emergence of words : Attentional
learning in form and meaning. Cognitive Science, 29,
pp. 819865.
Roy, D. (2008). A mechanistic model of three facets of
meaning. In M. D. Vega, G. Glennberg & G. Graesser
(Eds.), Symbols and Embodiment. Oxford University
Press.
Rumelhart, D. & Mcclelland, J. (1986). On learning the
past tenses of English verbs. In D. Rumelhart & J. Mc
clelland (Eds.), Parallel distributed processing: explor
ations in the microstructure of cognition. MIT Press.
pp. 216271.
Suppes, P., Liang, L. & Bottner, M. (1991). Complexity Is
sues in Robotic Machine Learning of Natural Language.
In L. Lam & V. Naroditsky (Eds.), Modeling Complex
Phenomena. Springer Verlag.
Tomasello, M. (1995). Joint attention as social cognition.
In C. Moore & P. J. Dunham (Eds.), Joint attention: Its
origins and role in development. Erlbaum.
Tomasello, M. (2005). Constructing a Language : A Us
ageBased Theory of Language Acquisition. Harvard
University Press.
van Rijsbergen, C. J. (1979). Information Retrieval. Lon
don, Butterworths.
quisition under different settings and with differ
ent types of data. They can then inform language
modellers directly about how particular language
models perform well and poorly in certain cases.
The collaborative aspect of the LAT encourages
not just programmers to share their code, but for
everyone to share their ideas.
6 Conclusion
This article proposes a tool that facilitates the
consolidation of research into the computational
modelling of child language acquisition under the
miniature language paradigm. The workshop is
being used to launch a first version of the LAT,
that is hoped to help language modellers and
child language experts to communicate and share
their knowledge.
7 Acknowledgements
This research was supported by the JeanLuc
Lagardère Foundation (http://www.fondation
jeanluclagardere.com).
Many thanks to the anonymous reviewers for
their constructive comments.
References
Bertoncini, J., BijeljacBabic, R., Jusczyk, P., Kennedy, L.
& Mehler, J. (1988). An investigation of Young infants'
perceptual representations of speech sounds. Journal of
experimental psychology, 117, pp. 2133.
BijeljacBabic, R., Bertoncini, J. & Mehler, J. (1993). How
Do 4DayOld Infants Categorize Multisyllabic Utter
ances?. Developmental Psychology, 29, pp. 711721.
Brown, R. (1973). A First Language: The Early Stages.
Harvard University Press.
DehaeneLambertz, G. & Houston, D. (1998). Faster Ori
entation Latencies Toward Native Language in Two
MonthOld Infants. Language and Speech, 41, pp.
2143.
Elman, J. (1993). Learning and development in neural net
works: the importance of starting small. Cognition, 48,
pp. 7199.
Feldman, J., Lakoff, G., Stolcke, A. & Weber, S. (1990,).
Miniature Language Acquisition: A Touchstone for
Cognitive Science.
Ingram, D. (1989). First Language Acquisition: Method,
Description and Explanation. Cambridge University
Press.
Jack, K. (2005,). Introducing a Scene Building Game to
Model Early First Language Acquisition, CLUK,
Manchester, England.
Kaplan, F., Oudeyer, P. & Bergen, B. (2008). Computa
tional models in the debate over language learnability.
Infant and Child Development, 17, pp. 5580.
Kellman, P., Gleitman, H. & Spelke, E. (1987). Object and
observer motion in the perception of objects by infants.
Journal of experimental psychology. Human perception
and performance, 13, pp. 586593.
Landau, B., Smith, L. & Jones, S. (1988). The importance
of shape in early lexical learning. Cognitive Develop
ment, 3, pp. 299321.
Mehler, J., Dupoux, T., Nazzi, T. & DehaeneLambertz, G.
(1996). Coping with linguistic diversity: The infant's
viewpoint. Lawrence Erlbaum.
Mehler, J., Jusczyk, P., Lambertz, G., Halsted, N., Berton
cini, J. & AmielTison, C. (1988). A precursor of lan
guage acquisition in young infants. Cognition, 29, pp.
143178.
Pinker, S. (1994). The Language Instinct. William Mor
row.
Plunkett, K., Sinha, C., Moller, M. & Strandsby, O. (1992).
Symbol grounding or the emergence of symbols?
Vocabulary growth in children and a connectionist net.
Connection Science, 4, pp. 293312.
Quinn, P. & Schyns, P. (2003). What goes up may come
down: perceptual process and knowledge access in the
organization of complex visual patterns by young in
fants. Cognitive Science, 27, pp. 923935.
Redington, M., Chater, N. & Finch, S. (1988). Distribu
tional Information: A powerful cue for acquiring syn
tactic categories. Cognitive Science, 22, pp. 425469.
Regier, T. (2005). The emergence of words : Attentional
learning in form and meaning. Cognitive Science, 29,
pp. 819865.
Roy, D. (2008). A mechanistic model of three facets of
meaning. In M. D. Vega, G. Glennberg & G. Graesser
(Eds.), Symbols and Embodiment. Oxford University
Press.
Rumelhart, D. & Mcclelland, J. (1986). On learning the
past tenses of English verbs. In D. Rumelhart & J. Mc
clelland (Eds.), Parallel distributed processing: explor
ations in the microstructure of cognition. MIT Press.
pp. 216271.
Suppes, P., Liang, L. & Bottner, M. (1991). Complexity Is
sues in Robotic Machine Learning of Natural Language.
In L. Lam & V. Naroditsky (Eds.), Modeling Complex
Phenomena. Springer Verlag.
Tomasello, M. (1995). Joint attention as social cognition.
In C. Moore & P. J. Dunham (Eds.), Joint attention: Its
origins and role in development. Erlbaum.
Tomasello, M. (2005). Constructing a Language : A Us
ageBased Theory of Language Acquisition. Harvard
University Press.
van Rijsbergen, C. J. (1979). Information Retrieval. Lon
don, Butterworths.
Sign up today - FREE
Mendeley saves you time finding and organizing research. Learn more
- All your research in one place
- Add and import papers easily
- Access it anywhere, anytime
Start using Mendeley in seconds!
Readership Statistics
5 Readers on Mendeley
by Discipline
60% Linguistics
by Academic Status
40% Ph.D. Student
20% Student (Master)
20% Researcher (at a non-Academic Institution)
by Country
40% Germany
20% United Kingdom
20% Malta



