5

COMPUTATIONAL AND
ROBOTIC MODELS OF EARLY
LANGUAGE DEVELOPMENT
A review

Pierre-Yves Oudeyer, George Kachergis, and William Schueller

The role of computational and robotic models of

language development1
Language involves a multitude of components interacting in complex ways in parallel and
within several scales of time and space. There is the developmental timescale of the growth of
an individual person, the glossogenetic or cultural timescale of the evolution of cultures, and the
phylogenetic timescale of the evolution of species. Within the developmental timescale, rang-
ing from moment-to-moment learning to years-long processes, multiple brain circuits interact
with a multi-part body situated in a complex environment, and in various kinds of situations
determining a diversity of information flows. Even though each of these levels and compo-
nents needs to be studied independently in order to reduce the complexity of the problem, it
is also fundamental to study their interactions. The sciences of complexity have taught us that
in many natural systems, there are global phenomena that are the result of local interactions
between components, but where individual study of these components would not allow us to
see the global properties of the whole combined system. Many of the properties of language are
probably not directly encoded by any of the components involved, but are the self-organized
outcomes of the interactions of the components.

Language development as a dynamic system

These self-organizational phenomena are often difficult to understand or to foresee intuitively,
and hard to formulate in words.This is why research into language development today leverages
computer science, mathematics, and robotics to provide complementary methods to develop-
mental psychology and neuroscience. These fields provide methods to build operational models
of the interactions between the components involved in language, conceptualized as a complex
dynamic system (Elman, 1995; Steels, 1997; Smith & Thelen, 2003; Beckner et al., 2009;Thomas
et al., 2016; Samuelson et al., 2017; Oudeyer, 2018a). An operational model is one which defines
the set of its assumptions explicitly and formally, and then computes or simulates the conse-
quences of these assumptions, proving that a particular set of conclusions is entailed. Such for-
malization forces researchers to make explicit all mechanisms and details of a theory, providing

76
 Computational and robotic models

opportunities to identify gaps, as well as to do synthetic experimentation with computers and

robots to test their plausibility, to understand better their implications, and to generate predic-
tions or new hypotheses. There are several kinds of models. In particular, from the structural
point of view, one can identify two broad families: analytical mathematical models and algo-
rithmic models of the causal mechanisms of language formation, which include a particularly
important sub-family of models that consider the central role of embodiment, and that will be
called here “robotic” models (as we will see, they do not necessarily entail the use of a physical
robot, but may use a simulated model of the body).

Analytical mathematical models

A first class of models abstract a certain number of variables from the phenomenon of natural
language development and express how these variables are related and how they change over
time in the form of mathematical equations. Usually this comprises systems of coupled dif-
ferential equations and makes use of the framework of dynamic systems theory, and sometimes
statistical physics (e.g., Nowak et al., 2002; Loreto et al., 2012). When these models are suf-
ficiently simple, their behavior can be predicted analytically and formally proved from their
mathematical structure alone. However, the abstractions and assumptions on which such proofs
are based are sometimes far removed from physical, cognitive, and social realities. Moreover, the
formal language of mathematical relations is not always very suitable for explaining processes at
work in nature (or in culture).

Algorithmic models
A second class of models focuses on the processes of language morphogenesis: these models are
formulated in terms of algorithms. These algorithms are themselves expressed, in practice, using
computer programming languages. Using this kind of formal language to describe a natural
or a cultural process has two solid advantages. First, the great expressivity of these languages
lets them formulate highly complex processes concisely (Dowek, 2011). Second, in the case of
phenomena whose behavior is very difficult to predict analytically from a series of equations, it
can be possible to calculate this behavior automatically through simulation. The programs can
be run on computers in what is a simulation of a morphogenetic process, and researchers can
observe how the simulated system behaves with different parameters. In research on language
development, this approach often involves building artificial systems in which individuals (their
bodies, brains, and behavior), their interactions, and their environment are modeled by programs.
A large share of models have emphasized the modeling of the processes of language represen-
tation and learning, and can be distinguished along several dimensions described in the next
three paragraphs: numerical vs symbolic, supervised vs unsupervised vs reinforcement learning,
normative vs heuristic models.

Learning numerical, symbolic, or hybrid representations

Some models represent language knowledge using strings or structured sets of symbols, such
as in models of grammatical learning where sentences are represented by hierarchical trees of
symbolic categories. Other models represent linguistic knowledge as sets of numbers, organized
in vectors, matrices, or neural activities. For example, the discovery of invariants and associa-
tions in a multimodal flow of speech words and images can be achieved using techniques like
non-negative matrix factorization (Driesen et al., 2009; Mangin et al., 2015). Another large

77
 Pierre-Yves Oudeyer et al.

family of models considers numerical connectionist representations, in a family of models called

connectionist modeling (Westermann & Twomey, 2017). Connectionist models rely on simpli-
fied models of neural circuits and their interactions, and most often the structure of the artificial
neural network is specified by the modeler, and the weights of connections among artificial
neurons are adapted during a simulated learning process. For example, some models of speech
sound acquisition use artificial neural networks to simulate the acquisition of sensorimotor
representations enabling an individual to learn to produce the speech sounds of its environment
(Warlaumont et al., 2013). Some models combine connectionist and symbolic representations,
such as models of various forms of syntactic processing where symbolic sequences of letters
or words are fed into a neural network that learns to predict associated symbolic outputs such
as word boundaries, syntactic categories, or semantic parse trees (Elman, 1990; Westermann &
Twomey, 2017; Dupoux, 2018).

Supervised, unsupervised, or reinforcement learning

Models of language acquisition leverage different types of learning processes as classified by
machine learning theory. The first form of learning is supervised learning, where the model
learns by observing inputs (e.g., an image of an object) and associated outputs to be predicted
(e.g., the name of the object), trying to discover the regularities between these inputs and out-
puts to improve its predictions or classifications. Self-supervised learning is a specific form of
supervised learning where the output is not provided by a teacher, but is directly observed in
the environment by the learner (e.g., the input could be one modality of observations, and the
output another modality which the learner learns to predict from the first modality). Unsuper-
vised learning is another form of learning where inputs are not matched by target outputs, but
rather the learning mechanism attempts to discover regularities in the underlying structure of
the input distribution (e.g., clustering words according to the similarity of the contexts in which
they appear). Finally, reinforcement learning is a form of learning where the learner attempts to
learn a (context-dependent) behavioral policy that maximizes a reward expressed in the form of
a scalar feedback. Hence, the learner is not provided examples of correct behaviors, but rather its
behavioral attempts receive a score of fitness that it tries to improve. For example, some models
of speech learning use an intrinsic reward measuring the saliency of sounds to drive the learning
of vocalizations (Warlaumont, 2013), and some other models use an extrinsic reward measur-
ing task completion to drive the acquisition of sentence interpretation in joint task contexts
(Daubigney et al., 2012).

Normative or heuristic models of learning processes

Another dimension of importance in the landscape of models of learning and representa-
tional mechanisms is the normative/heuristic distinction. One approach that has been very
fruitful has been to model the child language learner as a Bayesian and/or rational learner
which optimally infers structures from observed data and from a priori knowledge expressed
in the form of probabilistic priors (Chater & Manning, 2006; Tenenbaum et al., 2011). The
advantage of such an approach, especially when expressed in the Bayesian framework, is that
it forces one to express formally and in a simple mathematical language most of the model
assumptions, and it provides principled mathematical tools to evaluate the goodness of fit of
particular models to account for a body of empirical data. However, such an approach also
has several drawbacks. First, a number of findings show that the human brain, and especially
the child’s brain, may often learn in non-optimal and non-Bayesian manners, using heuristic

78
 Computational and robotic models

inference and cognitive shortcuts, not using all information available, and prone to various
forms of errors (Morewedge & Kahneman, 2010). There are actually a number of arguments
showing the potential evolutionary selection of such heuristics in a rapidly changing envi-
ronment with severe limits on cognitive and metabolic resources (Todd & Gigerenzer, 2000;
Oudeyer, 2018b). Another drawback of normative Bayesian models is that they require pre-
specification of all possible observations, events, and models in order to be able to compute
probabilities. Hence, by construction they do not address the question of how representa-
tions are learned (they only address the question of how certain representations are selected
among existing ones), which is a fundamental question of language development (e.g., how
do phonetic representations form? How are word meaning representations formed? How
are syntactic categories formed? etc). Also, normative Bayesian models become computa-
tionally intractable if one aims to scale to real-world high-dimensional data (Bossaerts &
Murawksi, 2017). For these reasons, another very large family of models relies on heuris-
tic models of learning, ranging from connectionist approaches (Westermann & Mareschal,
2014; Cangelosi & Schlesinger, 2015; Twomey et al., 2016) to heuristic statistical learning
(Saffran et al., 1996a, b; Mangin et al., 2015) or symbolic learning (Mealier et al., 2017;
Spranger & Steels, 2012). The advantage of these models is their very large expressivity,
their capacity to address the problem of representation learning (especially in connectionist
approaches), and their capacity to combine different kinds of learning mechanisms in the
same model. Also, these models have shown to scale better to complex real-world situations,
as shown by their pervasive use in robotic models of language learning grounded in high-
dimensional spaces of perception and action (Cangelosi et al., 2010). A drawback of these
models is that they often contain many free parameters, and there is no principled unified
statistical framework enabling one to compare in an unequivocal manner their goodness of
fit to account for empirical data.
Beyond a view of the computational modeling landscape organized along the formal tech-
nical dimensions we just described, there are two other ways to structure this landscape. First,
it is possible to classify models in terms of which stage(s) of infant language development they
are focusing on. In this chapter, we are focusing on the models of early language development,
ranging from vocal development to the onset of speech as a communication medium and to
early word learning. Many other models in the literature have also focused on later stages of
language development, with a large focus on the development of syntactic capabilities: for these
works, we refer the readers to these excellent reviews: Chater & Manning, 2006; Monaghan &
Christiansen, 2008;Yang, 2011; McCauley & Christiansen, 2014.
Another way to classify models is in terms of the general causal developmental mechanisms
that they focus on. This is the approach we follow in this chapter, where we analyze models in
terms of several of these general causal mechanisms: cross-situational statistical learning, embodi-
ment, social interaction, self-organization of brain-body-environment couplings, intrinsic moti-
vation, and the links between learning and evolution.
Many computational models of language learning focus largely on the learning mecha-
nisms involved in mapping words to their intended referents, referred as the problem of cross-
situational learning (see also Chapter 11, this volume). In other words, the focus is on the mechanisms
used to detect regularities in language data, while simplifying models of the interaction with the
environment, of how data is collected, and how this impacts the properties of data. For example,
many works model the environment as a database of examples which are incrementally and
randomly selected by the learner to train their learning mechanism (e.g., a database associating
words with their potential meanings). We focus on models for learning meaning, but many of
the issues we highlight are relevant for models of syntax learning as well.

79
 Pierre-Yves Oudeyer et al.

Robotic models: embodiment, social interaction, and intrinsically

motivated active learning
However, as we will detail more in the sections below, the real-world environment, and the
way it is perceived and acted upon by an active body, and through situated interaction with
others, contains a lot of structure that can guide learning processes. This is why computational
models of language development have in recent years been hybridized in implementations
that combine the use of computers and robotics. These models, instead of representing the
brain-body-environment system purely as computer algorithms, make use of programs run-
ning on computers that are embedded in physical robots. Here, only the “brain” is formulated
algorithmically, while the body is modeled using mechatronic elements, and the environment
approximates to humans’ real environment (including interaction with social peers though
human-robot interaction). Furthermore, the “brain” in these approaches is not simply viewed
as a passive statistical learning system, but rather an intrinsically motivated and goal-directed
system. This approach is currently at the heart of developmental robotics (Cangelosi & Schles-
inger, 2015), where there is intense activity around the modeling of sensorimotor, cognitive,
and social development (Oudeyer, 2010). Finally, beyond the body and motivational context,
language development has also been analyzed from the perspective of its bi-directional interac-
tion with cultural and phylogenetic evolution (Steels, 2003; Cangelosi et al., 2010).

Symbol grounding problem

Building this kind of robotic model is interesting from several points of view. First, regarding
language, it addresses the symbol grounding problem, in other words the fundamental problem
of how symbols are grounded in the physical world.The problem here, so eloquently formulated
by Stevan Harnad (1990), is understanding how the symbols commonly used for describing and
modeling languages (such as words and grammatical rules) can become meaningful in the physi-
cal and social reality of a real organism. In particular, it involves a capacity to link the abstract
world of symbols to the concrete world of numerical and chemical quantities that are perceived
and manipulated by the brain and the body in context. It is hard to see how models that are
purely algorithmic, implemented entirely on computers that at base represent the world in a
symbolic, discrete way, might usefully inform any questions about symbol grounding. Hybrid
algorithmic-robotic models, on the other hand, which by definition are at the edge between
the symbolic and the physical worlds, are extraordinary tools for studying this problem (Steels &
Kaplan, 2000; Steels, 2012). Here algorithmic models of language are confronted with physi-
cal and social reality, and whether or not there is an effective grounding of symbols – a strong
constraint on the plausibility of these models – can be tested empirically.

Brain-body-environment dynamic interactions

There is another very good reason for using robots. As Esther Thelen and Linda Smith, for
example, have argued in their theory of development (Smith & Thelen, 2003), the formation
of behavioral and cognitive structures results from dynamic interaction among the brain, the
body, and the environment. The body and the environment, whose physical substrate gives
rise to particular properties of structure generation, have a crucial role. Embodiment, that is
to say the material composition and the geometry of a body and its sensorimotor system, can
dramatically simplify the acquisition of certain behaviors. For example, Chen Yu and Linda
Smith showed how geometrical hand-eye relations and the physical manipulation of objects

80
 Computational and robotic models

could create favorable situations for learning the meanings of first words (Yu & Smith, 2012).
Thus, the body carries out physically a type of information processing, sometimes referred to as
“morphological computation” (Pfeifer et al., 2007). In this context, robots make it possible to
model mechatronically – in a straightforward, realistic way – interactions among the brain, the
body, and the environment that would be far too complex or even impossible to model algo-
rithmically. The section on the role of embodiment below provides several examples of robotic
models studying this perspective. In other spheres, many other examples can be found today of
robotic models being used to enhance understanding of animal and human behavior (Oudeyer,
2010), concerning such varied phenomena as navigation and phototropy in insects, control of
locomotion in dolphins, and distinguishing between self and non-self in human infants, but also
the impact of the visual system on the formation of linguistic concepts.

Mechanisms of cross-situational learning

Beyond learning the sounds of a language (e.g., Feldman, Griffiths, Goldwater, & Morgan, 2013),
and segmenting contiguous speech into words (e.g., Monaghan & Christiansen, 2010), infants
face the daunting challenge of mapping words to referents. Each utterance from a caregiver
might have any of a variety of intended meanings that can be difficult to discern in spite of cues
such as gaze and pointing. A useful source of disambiguating information may be had if infants
are able to track which words are frequently heard in conjunction with particular referents. If a
given word and its intended referent co-occur repeatedly across a variety of scenes, and infants
can track this conjunction to some extent, then the word’s meaning may be learned cross-
situationally. Such cross-situational learning is thought to be an important way for children to
learn words with concrete referents (Smith & Yu, 2008; Pinker, 1989). Since most real-world
situations contain many possible word-referent mappings, but time and attention are limited,
learners likely use heuristics or strategies (implicit or explicit; learned or innate) to restrict the
number of meanings they consider. Computational models of cross-situational learning have
sought to discover the representations and mechanisms that people use to track and disambigu-
ate word-referent co-occurrences.
Two dominant modeling approaches have emerged to account for cross-situational word
learning: associative models and hypothesis-testing models. These accounts are discussed below
with details of how they are typically compared to human learning behavior in short-term
experiments. However, models of both types often make simplifying assumptions about the
input (e.g., pre-segmented words and well-defined objects) that are unlikely to be met in the
real-world language environment. Models proposed in developmental robotics use machine
learning models with fewer assumptions about the input, but also have not yet been shown
to scale to real-world input. Independently, a variety of simulation investigations have asked
whether simple learning mechanisms can scale to acquiring a full-sized adult lexicon in a rea-
sonable amount of time. For analytical purposes, these simulations often make unrealistic sim-
plifying assumptions about the independence of learning each word, and about the sampling
process by which words and referents are experienced.
Cross-situational word learning has been studied in infants (Smith & Yu, 2008), children
(Akhtar & Montague, 1999), and adults. While it is assumed that many of the same word-
learning mechanisms are operating in adults, it is important to note that cognitive abilities in
infants and children are still developing. Adults are often studied because their greater attention
span enables the use of more complex, extended designs. In a typical adult study, participants are
asked to learn the referents of roughly 20 nonce words by watching a series of training trials.
On each trial, learners see an array of two to four unfamiliar objects (e.g., sculptures) and hear

81
 Pierre-Yves Oudeyer et al.

• •
Trial 1: Trial 2:
"stigson" "bosa"
"manu" "stigson"

Hypothesis-Testing Model

!
manu Trial 2: bosa

Biased Associative Model •

fX1
Trial 1: stigson manu Trial 2: bosa stigson

Figure 5.1 •
Example of two cross-situational learning trials (top) from which three word-referent pairs
might be learned, along with schematic representations that might be learned by a hypothesis-
testing model (middle; e.g., propose-but-verify) that tracks a single hypothesized referent per
word, and an associative model (bottom; e.g., the familiarity- and uncertainty-biased associa-
tive model) that attends to all co-occurring words and objects to some extent.

a corresponding number of pseudo-words (e.g., stigson, manu). See Figure 5.1 for an example of
two cross-situational learning trials, from which an observer may learn three word-object map-
pings. Although each word refers to a single onscreen object, the referent of each pseudo-word
is ambiguous on a given trial, because the intended referent is not indicated. In a typical learning
scenario, participants might view each of 18 word-object pairings six times as they appear four
at a time across 27 trials, for a total duration of several minutes (Yu & Smith, 2007; Kachergis,
Yu, & Shiffrin, 2013; Suanda & Namy, 2012).

Hypothesis-testing models
The hypothesis-testing theories view word learning as a problem of induction with an enor-
mous hypothesis space that must be reduced by the learner applying a number of language-
specific constraints in order to simplify the problem (Markman, 1992). In this view, infants
generate hypotheses that are consistent with this set of constraints and principles. For example,
the global principle (or bias) of mutual exclusivity (ME) assumes that every object has only one
name (Markman & Wachtel, 1988). At a lower level, the fill-the-lexical-gap bias is proposed to
cause children to want to find a name for an object with no known name (Clark, 1987; Merri-
man & Bowman, 1989). When given a set of familiar and unfamiliar objects, it has been shown
that 28-month-olds assume that a new label maps to an unfamiliar object (Mervis & Bertrand,
1994). Similarly, the principle of contrast states that an infant given a new word will seek to

82
 Computational and robotic models

attach it to an unlabeled object (Clark, 1987). Fill-the-gap, ME, and contrast make many of the
same predictions made by the more general novel name-nameless category principle (N3C),
which states that novel labels map to novel objects (Golinkoff, Mervis, & Hirsh-Pasek, 1994). In
order to be of aid to infant learners, such principles are thought to be either innate or developed
very early in life (Markman, 1992).
The hypothesis-testing approach is used in the formal analysis of language acquisition (Gold,
1967; Pinker, 1979), stemming from inferential methods in the philosophy of science, with
developmental theories built upon this rationale (Carey, 1978; Carey & Bartlett, 1978; Clark,
1987). A shared intuition among these approaches is that the multitude of co-occurrences avail-
able in the visual and auditory environment of the learner is far too complex to be tracked,
stored, and updated (Gleitman, 1990; Gershkoff-Stowe & Hahn, 2007; Medina, Snedeker,
Trueswell, & Gleitman, 2011; Trueswell, Medina, Hafri, & Gleitman, 2013). A model that exem-
plifies the hypothesis-testing approach is the propose-but-verify model (Trueswell et al., 2013),
which assumes that only a single hypothesized referent is stored for each word, and that on
further exposure to the word this proposal is recalled with some probability, and then either
verified if the referent is present – increasing the future probability of recall – or discarded if
the referent is absent (see Figure 5.1). In case the hypothesized referent is absent or fails to be
retrieved, a new hypothesis is chosen from the currently available referents.This model accounts
for adult word-learning behavior in some experimental contexts (Trueswell et al., 2013), but
not in others (Kachergis & Yu, 2017). More sophisticated rule-based models of word learn-
ing scale well in some corpus-based applications (Siskind, 1996), but have not attempted to
account for human behavior in experiments, or in real-world learning environments. A Bayesian
model of cross-situational word learning makes binary word-referent hypotheses according to
the global co-occurrence structure, combined with an a priori preference for a small lexicon
(Frank, Goodman, & Tenenbaum, 2009). This model is able to learn small lexicons from child-
parent interactions, based on transcribed speech and hand-coded representations of the visible
objects. These models showcase the learning power of sparse, hypothesis-based representations
in combination with rules and biases about which hypotheses to form.

Associative learning models

In another view, word learning can be explained as a gradual accumulation in memory of
all experienced co-occurring words and referents, perhaps relying on domain-general associa-
tive mechanisms rather than positing logic-based constraints or language-specific mechanisms
(Smith, 2000; Regier, 2005; Kachergis, 2012). In general, these models accumulate associative
strength between any co-occurring word and object, but the increment of a particular asso-
ciation can be weighted by factors such as prior knowledge (i.e., an already-existing associa-
tion based on past co-occurrence), novelty, or uncertainty (the entropy of a word or referent’s
associations; see Figure 5.1). This online attentional shifting can allow associative models to
show order effects such as highlighting (Kachergis, 2012), bootstrapping of low frequency word
meanings when they appear in the context of known referents (Kachergis,Yu, & Shiffrin, 2016),
and inference-like behavior much like a mutual exclusivity bias (Kachergis,Yu, & Shiffrin, 2012).
These behaviors of the biased associative model emerge from the interaction of competing
attentional biases to 1. strengthen already-strong associations and 2. attend to words or referents
with uncertain (high-entropy) associations (Kachergis et al., 2012). Rather than using arbi-
trary associative strengths, some associative models maintain conditional probability distribu-
tions that are updated as new co-occurrences are observed (Fazly, Alishahi, & Stevenson, 2010).

83
 Pierre-Yves Oudeyer et al.

The intuition behind associative learning models is that any words heard in a given context will
impact the representation of the referents to some extent. While it may seem too difficult for a
learner to track the associations between a word and not only its intended referent but its many
distractors, the presence of the unintended associations both provides a sense of context (e.g.,
forks and knives often appear together) and serves as noise when learners are trying to retrieve
the correct association. Associative models tend to be able to account for detailed behavioral
effects found in experiments, including interactions of context diversity and word frequency
(Kachergis et al., 2016) as well as response trajectories during word learning (Kachergis & Yu,
2017). Recent efforts to match detailed human learning trajectories across a range of experi-
mental conditions have found that sampling versions of models – both Bayesian (Yurovsky &
Frank, 2015) and associative (Kachergis & Yu, 2017) – best match human behavior by storing
multiple (but not all possible) hypothesized referents for each word. It has been pointed out that
simple hypothesis-testing (e.g., Medina et al., 2011) and simple associative accounts are at the
endpoints of a continuum of sampling models (Yu & Smith, 2012). A growing family of models
combine associative learning with online referent selection to achieve better fits to empirical
data (e.g., McMurray, Horst, & Samuelson, 2012; Kachergis & Yu, 2017).

Developmental robotics models

However, as mentioned earlier, both hypothesis-testing and associative models of cross-
situational word learning usually operate on the pre-segmented words and objects that are
assumed to be easily identified by adult learners in cross-situational learning studies (Yu &
Smith, 2007; Kachergis et al., 2016; Suanda & Namy, 2012). Modeling these short-term experi-
ments has yielded important insights into how people store and update particular word-referent
associations (or hypotheses), but using pre-segmented words and objects that are distinct from
each other and from the background may oversimplify the mapping problem. Real-world learn-
ing environments offer a stream of changing multimodal information, from which learners must
extract regularities at the appropriate level. It may be that learners need to first learn phonemes,
then words, and even word classes before being able to map the appropriate entities to refer-
ents. Similarly, learners may need to learn to segment the visual world into objects, properties,
and actions before being able to map words to these entities. However, learners may be able
to take advantage of regularities in the cross-modal representations of both auditory and visual
information (and other modalities, even). This is the approach taken by developmental robot-
ics researchers, who have investigated cross-situational word learning using machine learning
algorithms that build shared, cross-modal representations of scenes and utterances, from which
words and concepts can be extracted (Mangin, Filliat, ten Bosch, & Oudeyer, 2015; Chen et al.,
2018). Such models suggest how it is possible to simultaneously learn different types of words,
referents, and their associations without assuming separate processes for learning the structures
within each modality (e.g., object detection or word segmentation). Nonetheless, these models
are often tested in artificial experiments with small numbers of objects, utterances, and actions,
and will require more effort to be scaled up to realistic situations and to be compared in detail
to human word learners.

Large-scale simulations
Efforts to understand whether cross-situational learning can realistically learn an adult-sized
vocabulary come from simulation studies investigating simple learning mechanisms. Blythe,
Smith, and Smith (2010) tested how quickly a logic-based fast mapping mechanism (that strictly

84
 Computational and robotic models

rules out any referents not currently present when a word appears – often resulting in one-shot
learning) can be expected to learn the 60,000 words in an adult-sized vocabulary.The simulation
showed that learning time for a full vocabulary using a fast mapping mechanism is well within
reason, with 99% of the words learned by the time 940,000 words have been sampled (i.e., 142
words per day for 18 years). Simulations of a hypothesis-testing “guess-and-test” mechanism had
learning times that were only 50% slower – requiring a still reasonable 214 learning episodes per
day. However, these analytical estimates rely on making a variety of simplifying assumptions that
likely impact the validity of these estimates. The assumptions made by Blythe et al. (2010) and
others (Blythe, Smith, & Smith, 2016;Vogt, 2012) are: 1. a word is only heard when its meaning
is present in the situation, 2. perception of words and situations is errorless, 3. every situation has
the same number of possible referents, 4. each word maps to a single meaning, 5. learners know
the space of meanings that are possible, and 6. words are assumed to be learned independently,
meaning learners are expected not to be using even a mutual exclusivity bias. Many of these
assumptions further simplify the learning problem, although, critically, some are not realistically
plausible (e.g., words are often heard in the absence of the referent). Furthermore, the distribu-
tions of words, referents, and situations are independently randomly sampled in these simula-
tions, rather than reflecting the skewed frequency distributions found in real-world speech and
the nested structure found in natural scenes (Hidaka, Torii, & Kachergis, 2017). Future studies
will need to consider long-term learning in simulations using more realistic learning mecha-
nisms, as well as more realistic distributions of experience, with interdependent word learning.

The statistics of the language environment

The exponential distribution of word frequency – with a few words appearing quite often, and
the majority of words being vanishingly rare – has intrigued researchers for decades, but there
is no consensus on the cause of it (Zipf, 1949; Piantadosi, 2014). The fact that many words are
infrequently heard presents a difficulty for learning, as there will be few opportunities to learn
these words and disambiguate their meanings. If referents also have an exponential frequency
distribution, simulations predict that cross-situational learning will be orders of magnitude
slower (Vogt, 2012). Much recent research has been focused on the distribution and structure
of both the language input that children receive (Jones & Rowland, 2017; Hart & Risley, 1995)
and the visual scenes they encounter. While we often think of visual scenes as offering a mul-
titude of possible referents (Medina et al., 2011), analysis of recordings from head-mounted
cameras on infants during free play has shown that infants often have only a single object
dominating their view (Smith, Yu, & Pereira, 2011). The object filling their visual field is often
being actively manipulated by the child or caregiver, and is likely to be mentioned. This finding
seems to deflate the fabled complexity of the cross-situational learning problem. Importantly,
the studies reviewed so far primarily address learning the semantics of words, largely ignoring
syntax. However, a variety of models have been proposed to address the problem of learning
syntax from experiencing sentences (e.g., Chang, Dell, & Bock, 2006; Freudenthal et al., 2007;
Thomas & Knowland, 2014), as well as simultaneously learning semantics and syntax (e.g., a
Bayesian model: Abend et al., 2017; a connectionist model: Li, Farkas, & MacWhinney, 2004).
Due to space limitations we will not cover syntax learning in-depth, but many of the issues we
identify for cross-situational word-learning models also apply to syntax learning models.
Rather than making assumptions about the learning process, some researchers have opted
to use statistical waiting time models to investigate the impact of factors such as word length,
word class, frequency in child-directed speech, and imageability to predict words’ age of acqui-
sition (Braginsky, Yurovsky, Marchman, & Frank, 2016; Hidaka, 2013; Mollica & Piantadosi,

85
 Pierre-Yves Oudeyer et al.

2017). Fitted to children’s vocabulary growth curves, these statistical models have been used to
estimate the number of exposures and whether the rate of learning changes during develop-
ment (Hidaka, 2013; Mollica & Piantadosi, 2017). Another approach is to use network theoretic
models and semantic relatedness measures to model the growth process of children’s vocabulary
(Hills, Maouene, Riordan, & Smith, 2010). Much research has been devoted to characterizing
individual variability in early word learning, finding that the amount of child-directed speech
children receive correlates with vocabulary size and school readiness (Hart & Risley, 1995; Hut-
tenlocher, Haight, Bryk, Seltzeer, & Lyons, 1991). Recent efforts to create large shared databases
of early word-learning data such as WordBank (Frank, Braginsky,Yurovsky, & Marchman, 2016)
and recordings of child-directed speech in the home such as HomeBank (VanDam et al., 2016)
and CHILDES (MacWhinney, 2000) promise to unveil more about the structure and content
of children’s language input and its effects on vocabulary learning, and to serve as new goalposts
and constraints for modeling efforts.
However, children do not learn only by passively receiving this audiovisual stream of infor-
mation. Rather, they actively use their body to explore language and its referents, and they are
active communicative partners, likely attending to and leveraging a variety of social cues to aid
their learning. The following reviews how computational models have approached these active
dimensions of language development.

Robotic models: the role of embodiment, situatedness, and

social interaction
Child language learners are distinguished from most artificial machine learning systems in that
they are situated in a real-world physical environment with social peers. They are embodied in
a very complex physical organism which filters information from the environment and affords
certain kinds of actions on the environment. Also, children learn language in the context of
achieving social and material goals. This situatedness and embodiment provides multiple forms
of constraints that both define and guide language learning processes. Robotic models have
been developed to study the roles and structures of these constraints. It is important to note that
while robotic models focus on the physical and situatedness properties of the learner, they are
not necessarily implemented with real-world physical robots. Indeed, many robotic models are
entirely implemented in a virtual world with a simulator of the robot body and its physical/
social interaction with the environment.The advantage of using simulated robotic bodies is that
it facilitates systematic experimentation (gathering more statistics, larger exploration of the space
of parameters). The drawback of simulated robotic models is that some aspects of the real world
are difficult to simulate adequately, such as the natural variations (structured noise) in perceptual
channels or social interaction with a (model) caretaker. Let us now give a few representative
examples of (simulated or real-world) robotic models of language development.

Leveraging spontaneous structure from physics and body morphology

Models of speech development have extensively studied the role of physical embodiment, due
to the central role of the vocal tract and auditory system in phonetic and phonological learning
(Stevens, 1972; Schwartz et al., 1997). The vocal tract is one of the most complex organs in the
body, where a large number of muscles are used to continuously deform soft material parts such
as the tongue, lips, or vocal folds (Boersma, 1998). In addition, the produced sounds are perceived
by a complex auditory system with many non-linearities (Schwartz et al., 1997). One scientific
challenge in speech learning is understanding how children learn to produce the speech sounds

86
 Computational and robotic models

of their native language given the high complexity of this sensorimotor system. From the point
of view of control theory, this appears to be a conundrum given the high-dimensionality of the
space and severe limits on time and energy available to the child for trying out vocal tract move-
ments (Bernstein, 1967). So how can children learn canonical speech sounds already by the end
of their first year? Several models have studied the natural dynamics of vocal tract movements,
resulting from both mechanical coupling of movements and neural synergies among articulators.
For example, Kelso (Kelso et al., 1986) took a dynamic systems approach, showing that random
motor commands sent to the vocal tract produced already highly structured movements (hence
speech sounds) due to the spontaneous structure resulting from these coupling dynamics. This
enables one to show how learning speech sounds may amount to tuning some parameters of
these spontaneous structures, which is much easier than learning from scratch and without the
constraints of high-dimensional movements of the articulators.

Early words are grounded in concrete action and

perceptual repertoires
The role of embodiment has also been emphasized in several models of early word learning.
A first obvious reason is that the first words children learn are concrete and directly related to
their bodies, their actions, and their interactions with the environment (Bloom, 1995).Thus, the
representation of the meanings of these first words, especially verbs and nouns, is intrinsically
defined in terms of the action repertoire of children, and the way they perceive visually, hapti-
cally, auditorily, or spatially the objects around them. From this perspective, robotic models of
bodies and their physical perception and action on the environment are a prerequisite for mod-
eling the acquisition of the meaning of these early words, such as shown in models of learning of
the names of shapes and colors (Steels, 2001; Roy, 2005), simple manipulative actions (Cangelosi
et al., 2010), or spatial relationships among objects (Spranger & Steels, 2012).

Social scaffolding and imitation

A second straightforward role of embodiment and situatedness, studied in robotic models, relates
to social interaction and the non-verbal cues in the gaze, gestures, and discourse used by lan-
guage partners (Rohlfing et al., 2016). These non-verbal cues enable joint attention, and more
generally social guidance. Infants have been shown to attend to a variety of social and discourse
cues, which may greatly simplify both speech segmentation and word learning. For example, the
speech directed at infants by caregivers (e.g., motherese: child- or infant-directed speech) is char-
acterized by intonation and prosodic cues, and has been shown to aid segmentation (Thiessen,
Hill, & Saffran, 2005). It has been shown that adults are better able to segment artificial languages
when given prosodic cues, which seem to serve as a filter on the learned transitional probabilities
(Shukla, Nespor, & Mehler, 2007).
Beyond speech segmentation, social cues can reasonably be expected to impact word learn-
ing as well. Experiments have found that infants can follow a speaker’s gaze to infer what they
are referring to (Baldwin, 1993). Episodes of joint attention between an infant and caregiver,
as when playing with toys together, show characteristics such as shorter sentences from the
caregiver and more utterances from both in the dyad (Tomasello, 1988). Caregivers referring
to objects that were already focused on by the infant were correlated with a larger vocabulary,
while children of caregivers who attempted to redirect their attention had smaller vocabularies.
Other research seeking to characterize natural interactions between caregivers and children
has found that child-directed speech is quite repetitive, with repetitions of phrases, not just single

87
 Pierre-Yves Oudeyer et al.

words (Snow, 1972). An analysis of parent-child interactions while playing with toys showed
the informativeness of a variety of social cues relating to the hands and eyes of the speaker, as
well as to the continuity of discourse about particular referents (Frank, Tenenbaum, & Fernald,
2013). No single cue served as a perfect filter for the cross-situational learning of words, but in
combination these cues much reduce the ambiguity of intended meanings. Hearing an utter-
ance, infants may jointly consider the uncertainty about a speaker’s intended meaning as well as
uncertainty about the meaning of each word. This framing of the problem as one of communi-
cative inference is the basis for a model that simultaneously learns intended word-referent map-
pings as well as the relative value of social cues in making such inferences (Johnson, Demuth, &
Frank, 2012).
Early robotic models of language acquisition (Steels & Kaplan, 2000) compared the quality of
learning input (level of ambiguity between utterances and perceived scene) provided to a robot
learner in situations where 1. the human is socially and physical engaged in the interaction, syn-
chronizing pointing gestures towards referents while monitoring the gaze of the robot to ensure
gestures and referents are attended to at the right moments; 2. the human is semi-engaged, only
using utterances but not using actions to drive the learner’s attention; and 3. the human is not
socially and physically engaged, only describing the scene with utterances, independently of
what the robot is currently looking at. This kind of model allows us to quantify the additional
learning efficiency resulting from these various levels of engagement, leveraging embodiment
and situatedness. Other projects specifically designed robot learners capable of moving not only
to act upon objects but to communicate with social peers and realize joint attention (Scassellati,
1999). Some recent lines of work have used robotic models of embodied social language learn-
ers to also study how humans naturally teach language and how they use social cues to provide
feedback, for example using motherese (infant-directed speech) or motionese to demonstrate
simplified and highly informative learning examples (Vollmer & Schillingmann, 2017).
Developmental robotics models have further studied various links between sensorimotor
learning and social language learning. For example, a model based on intrinsically motivated
learning for efficient coding via active perception learns to copy goals, rather than the specific
motor movement, allowing it to learn simple behaviors such as gaze-following (Triesch, 2013).
The model begins by observing a tutor’s behavior and models the sensory consequences of the
behavior. Next, the model acts and receives a reinforcement signal from within that encodes
how well its sensations are matched by the sensory model. The model’s behavior is adapted to
make the sensory consequences of its actions better match the sensory model learned from
watching the tutor’s actions.
Cederborg and Oudeyer (2013) introduced a model for learning to acquire multiple skills by
observing a tutor’s ambiguous demonstrations. The model integrates concepts and techniques
from earlier cross-situational learning models, as well as models of motor learning by demonstra-
tions that treat meanings as complex sensorimotor policies with coordinate systems that must be
inferred. A contribution of Cederborg and Oudeyer is that the model learns both linguistic and
non-linguistic skills in a single process, without specifying a linguistic channel to the model. The
proof-of-concept demonstrates the viability of this approach, and future investigations will be
needed to determine how well it scales, and how well it matches human developmental trajectories.

Multimodal regularities
Other models have studied subtler, but equally fundamental, roles of embodiment. In addi-
tion to a flow of passively perceived utterances and visual scenes, embodiment and situatedness
provide the learner with the opportunity to also observe concurrently a flow of actions and

88
 Computational and robotic models

effects on the scene (including proprioception). This additional flow of information, enabled by
embodiment and consisting in an action-oriented modality, contains structure which can often
facilitate statistical inference of ambiguous structures and associations in the linguistic domain
(Rohlfing et al., 2006). For example, Mangin et al. (2015) show how invariants (e.g., words) in
low-level unsegmented speech streams, as well as their combinatorial structure and associations
with objects and actions, can be learned jointly with invariants and structure in low-level flows
of images and action movements using multimodal cross-situational learning methods. Like in
other related models (e.g., Cangelosi et al., 2016; Sugita & Tani, 2005; Mohammad et al., 2009),
such correlated flows of linguistic and sensorimotor information enable inference of general
structures of sentences and generalization, i.e., understanding the meaning of new sentences
whose precise word sequence was not encountered during training. Another example of the
facilitating role of sensorimotor information flows in language learning is the embodied model
of linguistic number counting presented in De La Cruz and colleagues (2014). Here, a neural
network model is used to account for how children might learn to count linguistically by pro-
nouncing the numbers in sequence, and how this might bootstrap internal representations of
numbers that link the names of numbers to meaningful underlying number representations. The
model compares a situation where the neural network is only observing sequences of linguistic
names, and a situation where the network is also concurrently observing the proprioceptive
information of finger counting actions: experiments have shown that observing propriocep-
tive finger information improves both the accuracy of counting and the quality of the acquired
internal representations of numbers. Interestingly, the mediating effect of sensorimotor repre-
sentations for language learning has also been used to model surprising effects of posture during
word-learning experiments (Morse et al., 2015), reproducing the observations of Samuelson
et al. (2011) that the inference of word meanings referring to objects can be significantly influ-
enced by the posture children have when they hear these novel words.

Curiosity, intrinsic motivation, and active learning

in language development
Learning in general, and language learning in particular, is not simply a passive information-
processing cognitive phenomenon in children. It is also a motivated and active process of inquiry
into the external world. Indeed, to learn speech and language, children spontaneously explore
how their vocal tract can produce a variety of sounds, and how these sounds produce social
effects in their peers, and they systematically point to objects or ask questions about all kinds of
things to get linguistic information. Doing so, they collect the large databases of learning exam-
ples that are needed to learn language (Oudeyer & Smith, 2016), but this is very costly in brain
time and energy: exploration, trial and error, and conversations take a lot of time; and process-
ing the initially unknown linguistic data requires significant cognitive effort. This is particularly
mysterious in early speech development, when infants have not yet understood what language
is, or how it can be used as a tool to fulfill external goals (e.g., asking a caretaker to bring an
object). Nonetheless, infants systematically spend a lot of energy exploring what sounds they
can make with their vocal tract (Oller, 2000), not knowing that it may be useful for the yet-
to-be-discovered language tool. So one fundamental question is: why do they spend so much
energy exploring and learning language? What are the proximal mechanisms that push them to
be interested in language learning at the developmental level? Several theoretical perspectives
can be taken (Oudeyer, 2006). One consists of speculating that evolution might have selected
language-specific motivational circuits that push infants to explore language. However, no pre-
cise naturalized formulation of this approach has been articulated so far, in particular in terms of

89
 Pierre-Yves Oudeyer et al.

how these language-specific motivational circuits might be developmentally implemented and

how they can address the variety of exploratory language activities. Another perspective is that
of social learning, and imitation in particular.While imitation learning indeed plays a major role
in language learning (Bloom, Hood, & Lightbown, 1974; Kuhl, 2000), and has been the focus of
several computational models (Cederborg & Oudeyer, 2013), it cannot easily account for the full
range of spontaneous language exploration shown by infants: adult speakers do not themselves
show such systematic exploration (e.g., they do not spontaneously babble all kinds of sounds),
and thus imitating the motivated exploration process is not a possibility. Also, while adults pro-
vide feedback and encouragement, the general difficulty of using this mechanism to drive the
interest of children towards various kinds of activities is not fully compatible with the universal
and systematic spontaneous interest shown by children in language exploration.

Early language acquisition driven by curiosity

Another hypothesis is that children are equipped with general intrinsic motivation systems
that push them to explore their body, their actions, and their vocal tract and language, through
curiosity-driven learning (Oudeyer & Kaplan, 2006; see also Chapter 6, this volume). This
hypothesis has been fleshed out in the past decade through a series of computational models
of curiosity-driven learning in developmental robotics (Oudeyer & Kaplan, 2006; Oudeyer &
Smith, 2016; Forestier & Oudeyer, 2017). Considering child development in general, psycholo-
gists proposed early on that humans and other animals may be equipped with neural intrinsic
motivation circuits that push them to explore activities and stimuli for their own sake (Berlyne,
1960), as opposed to maximizing an external reward such as food or social feedback.These moti-
vational neural circuits were proposed to rely on intrinsic rewards measuring the interestingness
of stimuli or activities in terms of quantities such as novelty, surprise, cognitive dissonance, or
intermediate complexity (Berlyne, 1960; White, 1959; Festinger, 1957). Within the past decade,
research on computational modeling of these processes of spontaneous exploration driven by
forms of curiosity has provided several formal frameworks used to understand these mechanisms
(Gottlieb et al., 2013; Friston et al., 2017; Oudeyer, 2018b), and to experiment with in robotic
models of sensorimotor learning (Baldassarre & Mirolli, 2013). Within this line of research,
some models have considered the hypothesis that intrinsic rewards in humans are provided
by sensorimotor activities which provide progress in learning predictive models of the world
(Kaplan & Oudeyer, 2007), and have showed how it could enable robot learners to spontane-
ously discover skills of increasing complexity (Oudeyer et al., 2007; Oudeyer & Smith, 2016). In
the Playground Experiment (Oudeyer & Kaplan, 2006), the sensorimotor space included both
leg/head movements and (simulated) vocal tract movements, as well as perception of object
movements, visual saliency, and sounds. The environment contained both physical objects with
affordances and also an “adult” peer robot contingently imitating the vocalization of the learning
robot. This experiment showed that the same general curiosity-driven learning system leads the
robot learner to engage in orderly exploration and learn how to act upon objects with its legs/
head, and how to provoke vocal responses from the peer robot by producing vocalizations, i.e.,
to enter spontaneously into a primitive form of speech interaction.
Using a similar model, but focusing on the study of the explored vocal sounds, Moulin-Frier
et al. (2014) conducted experiments where a robot explored the control of a realistic model of
the vocal tract in interaction with vocal peers through a drive to maximize learning progress.
This model relied on a physical model of the vocal tract, its motor control, and the auditory
system. The experiments showed how such a mechanism can explain the adaptive transition
from vocal self-exploration with little influence from the speech environment, to a later stage

90
 Computational and robotic models

where vocal exploration becomes influenced by the vocalizations of peers. Within the initial
self-exploration phase, a sequence of vocal production stages self-organizes, and shares proper-
ties with infant data (Oller, 2000): the vocal learner first discovers how to control phonation,
then vocal variations of unarticulated sounds, and finally articulated proto-syllables. As the vocal
learner becomes more proficient at producing complex sounds, the imitating vocalizations of
the teacher provide high learning progress, resulting in the well-known infant shift from vocal
self-exploration to vocal imitation (Oller, 2000).

Discovering the linguistic function of speech utterances

More recently, Forestier and Oudeyer (2017) extended these models in experiments showing
not only how curiosity-driven exploration could lead a learner to explore its vocal tract, but
also how it could lead to learning how to use these speech sounds to manipulate simulated
social peers, e.g., getting them to bring an interesting object which is out of physical reach. Fur-
thermore, the developmental trajectories generated in these simulations share several qualitative
properties with infant development, e.g., overlapping waves phenomena in tool use develop-
ment (Siegler, 1996), showing the plausibility of the hypothesis that curiosity may play a crucial
role in the onset of language development.

Active learning generates an ordered curriculum for learning

Finally, this family of models also enables study of the links between curiosity and the impact
of curriculum learning in language acquisition. As reviewed in Smith et al. (2018), the input
statistics children are exposed to during their first years evolve with time along a learning
curriculum that controls the growth of complexity in the perceived situations. This has been
argued to be key in enabling children to learn efficiently and quickly the concepts that are
needed for language grounding. The actions and attentional decisions made by infants play a
key role in structuring this input, and models of curiosity-driven learning have shown how
mechanisms searching for niches of learning progress push the learner to focus on activities or
stimuli of gradually increasing complexity (Kaplan & Oudeyer, 2007). In other words, curiosity
can be viewed as a mechanism for actively controlling the growth of complexity in language
formation (Schueller & Oudeyer, 2016). Indeed, an empirical study of adult learners shows that
giving them active control over which objects they will see named in the next trial signifi-
cantly improves their word-learning performance, as compared to passively viewing a randomly
ordered selection (Kachergis et al., 2013). In simulations of learning a full-sized vocabulary with
Zipf-distributed word and referent frequency distributions, such self-directed control over the
to-be-named referents was necessary in order to make learning possible on a realistic timescale
(Hidaka et al., 2017).

Cultural evolution: how languages become learnable

Not only is language acquired during one’s lifetime, but it also evolves over time. Children learn
how to walk in more or less the same way today as they did a few centuries ago, but what they
learn about the lexicon and the grammar of the same language (e.g., English) is significantly
different. Pronunciation may have changed; a lot of new words may have appeared, while some
others are not used any more; and grammar structures such as case-marking may have disap-
peared, for example. This type of evolution happens too fast to be accounted for by biological
evolution, with some phenomena even taking place in a matter of weeks (e.g., invention of

91
 Pierre-Yves Oudeyer et al.

new words within a population). Biological evolution has its roots in genetic material being
copied and passed on from body to body, whereas cultural evolution takes the form of ideas,
words, or conceptual structures being passed on from brain to brain. The cultural evolution of
language derives from the repeated interactions between individuals of the population using a
language, with everyone learning from and adapting to their interlocutor.These interactions can
be between people of the same generation, designated as horizontal transmission, or between
different generations, what is called vertical transmission (see Figure 5.2).

Cultural evolution can shape languages

The question is, in the wide landscape of possible configurations and structures, are they all
equivalent and only the result of a slow drift of languages? Or, like biological evolution, do
some of these structures have an advantage, in particular concerning language acquisition? If
we look, for example, at the distribution of vowel systems (De Boer, 2001; Oudeyer, 2018a),
word order (Ferrer-i-Cancho, 2016), word frequency (Zipf, 1949), or even lexicon size for the
domains of color (Kay, Berlin, Maffi, Merrifield, & Cook, 2009) or kinship (Kemp & Regier,
2012), some structures are far more frequent than others in natural languages, which supports
the latter hypothesis. Two main pressures are exerted on language during this evolutionary pro-
cess: expressivity (how well a language fits the context of its usage) and learnability (how well a
language can be transmitted to and acquired by new learners). A third pressure, closely related to
the previous ones, is the facility to reproduce the language in terms of both precision and energy
needs, highly dependent on the characteristics of the articulo-auditory system of the agents.

Language Games and horizontal transmission

Language Games are a set of computational models in which a population of individuals have
pairwise interactions while trying to build or learn a common communication system (Steels,
2001; Loreto, Baronchelli, Mukherjee, Puglisi, & Tria, 2011). For each interaction in a typi-
cal Language Game, the corresponding pair of simulated agents – randomly selected from the
population – are assigned roles: one is the speaker, uttering a word to refer to a selected meaning
or scene, and the other is the hearer, trying to guess what the speaker was referring to. Within a

learns learns
Iterated
Learning
Random ~ ~\ te~es ~ \ te~es Structured
Language,
Vertical Language ~ ~ ~
easily learnable
transmission 1st 2nd
generation generation

Language (• *Q* *
** *
Games Common language,
No
** * *°*
self-organized,
Horizontal Language
learnable patterns
transmission Random pairwise interactions:
invention or agreement on words

Figure 5.2 Illustration of the two main classes of computational models of language cultural evolution,
and how these processes impact language learnability: Iterated Learning and Language Games.

92
 Computational and robotic models

certain number of interactions, which typically depends on the size of the population, all agents
agree on a common language and succeed in communicating efficiently. In other words, while
the rules underlying interactions remain simple, a communication system can self-organize. For
many of the models, convergence towards a shared vocabulary has been not only observed in
simulation but also proven mathematically (De Vylder, 2007; Baronchelli, Felici, Loreto, Cagli-
oti, & Steels, 2006). Moreover, the resulting linguistic structures can show interesting properties,
like categories that are well fitted to both the environment and the sensorimotor system of the
users. For example, in De Boer (2001) and Oudeyer (2018a), a population of simulated indi-
viduals commonly acquire a vowel system. The resulting vowels are always well distributed over
the continuous space of possible vowels, and are therefore easily learnable for a new individual
that would join the population.Vowels are also selected in a way that minimizes the articulatory
energy needed to produce them. A third evolutionary pressure is resistance to noise: because
of the non-linearity of the articulatory system, some configurations may be more unstable and
sensitive to small variations in the motor commands of the articulatory system. The more stable
ones are selected during the evolutionary process. Lending credence to these results, the statisti-
cal distribution of the number of vowels over numerous simulations resembles the distribution
found in natural languages. Another example is the collective negotiation of names for colors,
modeled in different ways (e.g., Steels & Belpaeme, 2005; Puglisi, Baronchelli, & Loreto, 2008).
In particular, the model used by Baronchelli, Gong, Puglisi, and Loreto (2010) arrives at a dis-
tribution of color categories that is adapted to both the human eye and the frequency of colors
in the environment. This model also fits real data, with the average number of color categories
produced by the model matching what is observed in the World Color Survey (Kay et al., 2009).
Language Games have been used to model many other parts of language, including spatial rep-
resentation (Spranger, 2012) and grammatical structures (Van Trijp, 2012), and many times the
simulated agents are made to interact using real robotic bodies (Spranger, 2012; Steels, 2001).

Iterated Learning and vertical transmission

Another type of model, called Iterated Learning, focuses on language transmission between
generations (Kirby, Griffiths, & Smith, 2014). Here, simulated agents also interact in pairs but in
chains: each of them represents a generation, and interacts only with the previous generation
and the following one. A first random model of language is generated, and a set of examples of
usage of this language are shown to the first generation. Not all possible objects are found in the
examples: learning individuals have to generalize from a reduced set only. This is an important
part of the models, called the transmission bottleneck, which leads to modifications of the language
in each generation. The first generation later uses what they have learned about the language
to generate a new set of examples that is used to teach the language to the second generation.
The generalization process relies on built-in cognitive biases of the agents: they tend to prefer
compressible languages.
This process is allowed to proceed for a chosen number of generations, and the language of
the final generation is observed. The exact definition of Iterated Learning sums it up quite well:
“Iterated learning is the process by which a behavior arises in one individual through induction
on the basis of observations of behavior in another individual who acquired that behavior in the
same way” (Kirby et al., 2014, p. 108). Typically, the objects that individuals have to refer to are
combinatorial: they have, for example, a shape and a texture, and all combinations of possible
shapes and textures can be found. The final languages range from holistic, with a distinct word
for every possible object, to completely structured, with a word for each shape and a word for
each texture, the name of the object being a combination of the two. In Kirby,Tamariz, Cornish,

93
 Pierre-Yves Oudeyer et al.

and Smith (2015), it has been shown that with both pressures of expressivity and learnability,
structured languages are selected. With only one of the two pressures, languages tend to either
be holistic or degenerate (with one single word for everything). Even if the starting language is
random, the preferred structure is selected and shaped over generations. This illustrates another
mechanism of cultural evolution: some patterns are favored and progressively selected because
of cognitive biases, and because of these very biases are easier to acquire by new learners having
them as well.
Those models do not pretend to describe the full process of language evolution, as they
each focus on some specific aspects of language evolution. Therefore, they do not represent real
language evolution as a whole. However, by studying them we understand that simple mecha-
nisms are enough to observe formation and self-organization of languages. Specific patterns
and structures emerge and can be selected, which in turn facilitate language acquisition. This
provides a theoretical perspective from which one can interpret the relative ease with which
children acquire language.

Conclusion
Modeling the development and learning of language has inspired researchers in computer sci-
ence, psychology, and robotics to adopt diverse approaches to the many challenges involved. We
have sought to highlight the main modeling approaches along with the behaviors and empiri-
cal data they seek to explain, while also outlining the remaining gaps between these accounts,
where future research must be aimed.
For example, cognitive models of cross-situational word learning carried out in the psychol-
ogy lab typically assume that words and referents are trivially identified and segmented, and
that words always appear with their intended referents – assumptions which are often violated
in real-world scenes. While more complex developmental robotics models rarely make these
assumptions, both cognitive and robotics models are typically only applied to matching human
behavior in small-scale learning scenarios, involving short utterances, a few objects at a time, and
a total vocabulary of tens of words. In contrast, other studies use mathematical analysis and simu-
lations of learning a full-sized vocabulary, but often make oversimplifying assumptions about the
distribution of words, referents, and even the cross-situational learning mechanism, while only
attempting to match gross overall human learning rates. Future studies will need to investigate
how well robotics models combine with cognitive models to account for both detailed short-
term human learning behavior of vocabulary and also long-term learning in real-world scenes
with full-fledged language and grammatical structures.
Another open dimension of research concerns computational modeling of the discovery of
speech as a linguistic tool to communicate with others about referents and to achieve joint tasks.
Indeed, most existing computational models (there are few exceptions) have so far relied on cogni-
tive architecture models where language is implicitly assumed to be a system of labels associated
with communicative referents. However, for early developing infants, speech sounds (like gestures)
are initially part of a rich, unorganized, and continuous flow of multimodal information: the spe-
cial communicative status of these sounds (or gestures) is only progressively discovered. This also
highlights the need to develop further computational theories of the ways language development is
embedded within the broader picture of sensorimotor, cognitive, and social development.

Note
1 Some material from this section was adapted from Oudeyer (2018a, CC-BY).

94
 Computational and robotic models

References
Abend, O., Kwiatkowski, T., Smith, N. J., Goldwater, S., & Steedman, M. (2017). Bootstrapping language
acquisition. Cognition, 164, 116–143. doi:10.1016/j.cognition.2017.02.009
Akhtar, N., & Montague, L. (1999). Early lexical acquisition: The role of cross – situational learning. First
Language, 19, 34–358. doi:10.1177/014272379901905703
Aslin, R., Saffran, J., & Newport, E. (1998). Computation of conditional probability statistics by 8-month-old
infants. Psychological Science, 1–4. doi:10.1111/1467-9280.00063
Baldassarre, G., & Mirolli, M. (2013). Intrinsically motivated learning in natural and artificial systems. Berlin:
Springer.
Baldwin, D. A. (1993). Infants’ ability to consult the speaker for clues to word reference. Journal of Child
Language, 20(2), 395–418. doi:10.1017/S0305000900008345
Baronchelli, A., Felici, M., Loreto, V., Caglioti, E., & Steels, L. (2006). Sharp transition towards shared
vocabularies in multi-agent systems. Journal of Statistical Mechanics: Theory and Experiment, 2006(6).
doi:10.1088/1742-5468/2006/06/P06014
Baronchelli, A., Gong, T., Puglisi, A., & Loreto, V. (2010). Modeling the emergence of universality in
color naming patterns. Proceedings of the National Academy of Sciences, 107(6), 2403–2407. doi:10.1073/
pnas.0908533107
Beckner, C., Blythe, R., Bybee, J., Christiansen, M. H., Croft, W., Ellis, N. C., . . . Schoenemann, T.
(2009). Language is a complex adaptive system: Position paper. Language Learning, 59(S1), 1–26.
doi:10.1111/j.1467-9922.2009.00533.x
Berlyne, D. (1960). Conflict, arousal and curiosity. New York, NY: McGraw-Hill.
Bernstein, N. (1967). The coordination and regulation of movements. Oxford, NY: Pergamon.
Bloom, L. (1995). The transition from infancy to language: Acquiring the power of expression. Cambridge, UK:
Cambridge University Press.
Bloom, L., Hood, L., & Lightbown, P. (1974). Imitation in language development: If, when, and why. Cogni-
tive Psychology, 6(3), 380–420. doi:10.1016/0010-0285(74)90018-8
Blythe, R. A., Smith, A. D. M., & Smith, K. (2016).Word learning under infinite uncertainty. Cognition, 151,
18–27. doi:10.1016/j.cognition.2016.02.017
Blythe, R. A., Smith, K., & Smith, A. D. M. (2010). Learning times for large lexicons through cross-
situational learning. Cognitive Science, 34(4), 620–642. doi:10.1111/j.1551-6709.2009.01089.x
Boersma, P. (1998). Functional phonology: Formalizing the interactions between articulatory and perceptual drives.
The Hague: Holland Academic Graphics.
Bossaerts, P., & Murawski, C. (2017). Computational complexity and human decision-making. Trends in
Cognitive Sciences, 21(12), 917–929. doi: 10.1016/j.tics.2017.09.005
Braginsky, M., Yurovsky, D., Marchman, V. A., & Frank, M. C. (2016). From uh-oh to tomorrow: Predicting age
of acquisition for early words across languages. Proceedings of the 38th Annual Conference of the Cognitive
Science Society.
Cangelosi, A., Morse, A., Di Nuovo, A., Rucinski, M., Stramandinoli, F., Marocco, M., . . . Fischer, K. (2016).
Embodied language and number learning in developmental robots. In M. H. Fischer & Y. Coello (Eds.),
Foundations of embodied cognition. Oxon: Routledge.
Cangelosi, A., & Schlesinger, M. (2015). From babies to robots:The contribution of developmental robotics
to developmental psychology. Child Development Perspectives, 12(3), 183–188. doi:10.1111/cdep.12282
Cangelosi, A., Metta, G., Sagerer, G., Nolfi, S., Nehaniv, C., Fischer, K., Zeschel, A. (2010). Integration of
action and language knowledge: A roadmap for developmental robotics. IEEE Transactions on Autono-
mous Mental Development, 2(3), 167–195. doi:10.1109/TAMD.2010.2053034
Carey, S. (1978). The child as word learner. In M. Halle, J. Bresnan, & G. A. Miller (Ed.), Linguistic theory and
psychological reality. Cambridge, MA: MIT Press.
Carey, S., & Bartlett, E. (1978). Acquiring a single new word. Papers and Report on Child Language Develop-
ment, 15, 17–29.
Cederborg, T., & Oudeyer, P. Y. (2013). From language to motor gavagai: Unified imitation learning of
multiple linguistic and nonlinguistic sensorimotor skills. IEEE Transactions on Autonomous Mental Devel-
opment, 5(3), 222–239. doi:10.1109/TAMD.2013.2279277
Chang, F., Dell, G. S., & Bock, K. (2006). Becoming syntactic. Psychological Review, 113(2), 234–272.
doi:10.1037/0033-295X.113.2.234
Chater, N., & Manning, C. D. (2006). Probabilistic models of language processing and acquisition. Trends in
Cognitive Sciences, 10(7), 335–344. doi:10.1016/j.tics.2006.05.006

95
 Pierre-Yves Oudeyer et al.

Chen, Y., Bordes, J. B., & Filliat, D. (2018). Comparison studies on active cross-situational object-word
learning using Non-Negative Matrix Factorization and Latent Dirichlet Allocation. IEEE Transactions
on Cognitive and Developmental Systems, 10(4), 1023–1034.
Clark, E.V. (1987).The principle of contrast: A constraint on language acquisition. In B. MacWhinney (Ed.),
Mechanisms of Language Acquisition. Hillsdale, NJ: Lawrence Erlbaum.
Daubigney, L., Geist, M., Chandramohan, S., & Pietquin, O. (2012). A comprehensive reinforcement learn-
ing framework for dialogue management optimization. IEEE Journal of Selected Topics in Signal Processing,
6(8), 891–902. doi:10.1109/JSTSP.2012.2229257
De Boer, B. (2001). The origins of vowel systems (Vol. 1). Oxford: Oxford University Press.
De La Cruz,V. M., Di Nuovo, A., Di Nuovo, S., & Cangelosi, A. (2014). Making fingers and words count in
a cognitive robot. Frontiers in Behavioral Neuroscience, 8, 13. doi:10.3389/fnbeh.2014.00013
De Vylder, B. (2007). The evolution of conventions in multi-agent systems (Unpublished doctoral dissertation),
Vrije Universiteit Brussel, Brussels.
Dowek, G. (2011). Une Deuxième Révolution Galiléenne? Retrieved from https://who.rocq.inria.fr/Gilles.
Dowek/galilee.pdf
Driesen, J., Ten Bosch, L., & Van Hamme, H. (2009). Adaptive non-negative matrix factorization in a computa-
tional model of language acquisition. Proceedings of INTERSPEECH 2009, 10th Annual Conference of
the International Speech Communication Association, Brighton, UK.
Dupoux, E. (2018). Cognitive science in the era of artificial intelligence: A roadmap for reverse-engineering
the infant language-learner. Cognition, 173, 34–59. doi:10.1016/j.cognition.2017.11.008
Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2), 179–211. doi:10.1207/
s15516709cog1402_1
Elman, J. L. (1995). Language as a dynamical system. In R. F. Port & T. van Gelder (Eds.), Mind as motion:
Explorations in the dynamics of cognition (pp. 195–223). Cambridge, MA: MIT Press.
Fazly, A., Alishahi, A., & Stevenson, S. (2010). A probabilistic computational model of cross-situational word
learning. Cognitive Science, 34(6), 1017–1063. doi:10.1111/j.1551-6709.2010.01104.x
Feldman, N. H., Griffiths, T. L., Goldwater, S., & Morgan, J. L. (2013). A role for the developing lexicon in
phonetic category acquisition. Psychological Review, 120(4), doi:10.1037/a0034245.
Ferrer-i-Cancho, R. (2016). Kauffman’s adjacent possible in word order evolution. In S. G. Roberts, C. Cuskley, L.
McCrohon, L. Barceló-Coblijn, O. Fehér & T.Verhoef (eds.), The Evolution of Language: Proceedings
of the 11th International Conference (EVOLANG11).
Festinger, L. (1957). A theory of cognitive dissonance. Evanston, IL: Row, Peterson and Company.
Forestier, S., & Oudeyer, P-Y. (2017). A unified model of speech and tool use early development. Proceedings of
the 39th Annual Meeting of the Cognitive Science Society.
Frank, M. C., Braginsky, M.,Yurovsky, D., & Marchman,V.A. (2016).Wordbank:An open repository for devel-
opmental vocabulary data. Journal of Child Language, 44(3), 677–694. doi:10.1017/S0305000916000209
Frank, M. C., Goodman, N. D., & Tenenbaum, J. B. (2009). Using speakers’ referential intentions to
model early cross-situational word learning. Psychological Science, 20(5), 578–585. doi:10.1111/
j.1467-9280.2009.02335.x
Frank, M. C.,Tenenbaum, J. B., & Fernald,A. (2013). Social and discourse contributions to the determination
of reference in cross-situational word learning. Language Learning and Development, 9, 1–24. doi:10.1080/
15475441.2012.707101
Freudenthal, D., Pine, J. M., Aguado-Orea, J., & Gobet, F. (2007). Modeling the developmental patterning
of finiteness marking in English, Dutch, German, and Spanish using MOSAIC. Cognitive Science, 31(2),
311–341. doi:10.1080/15326900701221454
Friston, K. J., Lin, M., Frith, C. D., Pezzulo, G., Hobson, J. A., & Ondobaka, S. (2017). Active inference,
curiosity and insight. Neural Computation, 29(10), 2633–2683. doi:10.1162/neco_a_00999
Gershkoff-Stowe, L., & Hahn, E. R. (2007). Fast mapping skills in the developing lexicon. Journal of Speech,
Language, and Hearing Research, 50, 682–697. doi:10.1044/1092-4388(2007/048)
Gleitman, L. (1990). The structural sources of word meaning. Language Acquisition, 1, 3–55. doi:10.1207/
s15327817la0101_2
Gold, E. M. (1967). Language identification in the limit. Information and Control, 16, 447–474. doi:10.1016/
S0019-9958(67)91165-5
Golinkoff, R. M., Mervis, C. B., & Hirsh-Pasek, K. (1994). Early object labels: The case for a devel-
opmental lexical principles framework. Journal of Child Language, 21(1), 125–155. doi:10.1017/
S0305000900008692

96
 Computational and robotic models

Gottlieb, J., Oudeyer, P. Y., Lopes, M., & Baranes, A. (2013). Information-seeking, curiosity, and attention:
Computational and neural mechanisms. Trends in Cognitive Sciences, 17(11), 585–593. doi:10.1016/j.
tics.2013.09.001
Hart, B., & Risley, T. R. (1995). Meaningful differences in the everyday experience of young American children.
Baltimore, MD: Paul H. Brookes Publishing.
Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1–3), 335–346. doi:
10.1016/0167-2789(90)90087-6
Hidaka, S. (2013). A computational model associating learning process, word attributes, and age of acquisi-
tion. PLoS One, 8(11), e76242. doi:10.1371/journal.pone.0076242
Hidaka, S.,Torii,T., & Kachergis, G. (2017). Quantifying the impact of active choice in word learning. Proceedings
of the 38th Annual Conference of the Cognitive Science Society.
Hills, T. T., Maouene, J., Riordan, B., & Smith, L. B. (2010). The associative structure of language: Contex-
tual diversity in early word learning. Journal of Memory and Language, 63(3), 259–273. doi:10.1016/j.
jml.2010.06.002
Huttenlocher, J., Haight,W., Bryk, A., Seltzeer, M., & Lyons,T. (1991). Early vocabulary growth: Relation to
language input and gender. Developmental Psychology, 27(2), 236–248. doi:10.1037/0012-1649.27.2.236
Jones, G., & Rowland, C. F. (2017). Diversity not quantity in caregiver speech: Using computational mod-
eling to isolate the effects of the quantity and the diversity of the input on vocabulary growth. Cognitive
Psychology, 98, 1–21. doi:10.1016/j.cogpsych.2017.07.002
Johnson, M., Demuth, K., & Frank, M. C. (2012). Exploiting social information in grounded language
learning via grammatical reduction. In Proceedings of the 50th annual meeting of the association for computa-
tional linguistics (pp. 883–891). Jeju Island, Korea: Association for Computational Linguistics.
Kachergis, G. (2012). Learning nouns with domain-general associative learning mechanisms. In N. Miyake,
D. Peebles, & R. P. Cooper (Ed.), Proceedings of the 34th annual conference of the cognitive science society
(pp. 533–538). Austin, TX: Cognitive Science Society.
Kachergis, G., & Yu, C. (2017). Observing and modeling developing knowledge and uncertainty during
cross-situational word learning. IEEE Transactions on Cognitive and Developmental Systems. doi:10.1109/
TCDS.2017.2735540
Kachergis, G.,Yu, C., & Shiffrin, R. M. (2012).An associative model of adaptive inference for learning word –
referent mappings. Psychonomic Bulletin and Review, 19(2), 317–324. doi:10.3758/s13423-011-0194-6
Kachergis, G.,Yu, C., & Shiffrin, R. M. (2013). Actively learning object names across ambiguous situations.
Topics in Cognitive Science, 5(1), 200–213. doi:10.1111/tops.12008
Kachergis, G.,Yu, C., & Shiffrin, R. M. (2016). A bootstrapping model of frequency and contextual diversity
effects in word learning. Cognitive Science. doi:10.1111/cogs.12353
Kaplan, F., & Oudeyer, P.Y. (2007). In search of the neural circuits of intrinsic motivation. Frontiers in Neu-
roscience, 1, 17.
Kay, P., Berlin, B., Maffi, L., Merrifield,W. R., & Cook, R. (2009). The world color survey. Stanford, CA: CSLI
Publications. doi:10.1016/B978-008044612-7/50064-0
Kelso, J. A. S., Saltzman, E. L., & Tuller, B. (1986). The dynamical perspective on speech production: Data
and theory. Journal of Phonetics, 14, 29–59.
Kemp, C., & Regier, T. (2012). Kinship categories across languages reflect general communicative princi-
ples. Science, 336(6084), 1049–1054. doi:10.1126/science.1218811
Kirby, S., Griffiths, T., & Smith, K. (2014). Iterated learning and the evolution of language. Current Opinion
in Neurobiology, 28, 108–114. doi:10.1016/j.conb.2014.07.014
Kirby, S., Tamariz, M., Cornish, H., & Smith, K. (2015). Compression and communication in the cultural
evolution of linguistic structure. Cognition, 141, 87–102. doi:10.1016/j.cognition.2015.03.016
Kuhl, P. K. (2000). A new view of language acquisition. Proceedings of the National Academy of Sciences, 97(22),
11850–11857. doi:10.1073/pnas.97.22.11850
Li, P., Farkas, I., & MacWhinney, B. (2004). Early lexical development in a self-organizing neural network.
Neural Networks, 17(8–9), 1345–1362. doi:10.1016/j.neunet.2004.07.004
Loreto,V., Baronchelli, A., Mukherjee, A., Puglisi, A., & Tria, F. (2011). Statistical physics of language dynam-
ics. Journal of Statistical Mechanics: Theory and Experiment, 2011(4). doi:10.1088/1742-5468/2011/04/
P04006
Loreto,V., Mukherjee, A., & Tria, F. (2012). On the origin of the hierarchy of color names. Proceedings of the
National Academy of Sciences, 109(18), 6819. doi:10.1073/pnas.1113347109
MacWhinney, B. (2000). The CHILDES project:Tools for analyzing talk. Hillsdale, NJ: Lawrence Erlbaum.

97
 Pierre-Yves Oudeyer et al.

Mangin, O., Filliat, D., ten Bosch, L., & Oudeyer, P-Y. (2015). MCA-NMF: Multimodal concept acquisition
with non-negative matrix factorization. PLoS One, 10(10), 1–35. doi:10.1371/journal.pone.0140732
Markman, E. M. (1992). Constraints on word learning: Speculations about their nature, origins and domain
specificity. In M. R. Gunnar & M. P. Maratsos (Ed.), Modularity and constraints in language and cognition:
The Minnesota symposium on child psychology (pp. 59–101). Hillsdale, NJ: Lawrence Erlbaum.
Markman, E. M., & Wachtel, G. F. (1988). Children’s use of mutual exclusivity to constrain the meanings of
words. Cognitive Psychology, 20, 121–157. doi:10.1016/0010-0285(88)90017-5
McCauley, S. M., & Christiansen, M. H. (2014). Prospects for usage-based computational models of gram-
matical development: Argument structure and semantic roles. Wiley Interdisciplinary Reviews: Cognitive
Science, 5(4), 489–499. doi:10.1002/wcs.1295
McMurray, B., Horst, J. S., & Samuelson, L. K. (2012). Word learning emerges from the interaction of
online referent selection and slow associative learning. Psychological Review, 119(4), 83877. doi:10.1037/
a0029872
Mealier, A. L., Pointeau, G., Mirliaz, S., Ogawa, K., Finlayson, M., & Dominey, P. F. (2017). Narrative con-
structions for the organization of self experience: Proof of concept via embodied robotics. Frontiers in
Psychology, 8. doi:10.3389/fpsyg.2017.01331
Medina, T. N., Snedeker, J., Trueswell, J. C., & Gleitman, L. R. (2011). How words can and cannot be
learned by observation. Proceedings of the National Academy of Sciences, 108, 9014–9019. doi:10.1073/
pnas.1105040108
Merriman, W. E., & Bowman, L. L. (1989). The mutual exclusivity bias in children’s word learning. Mono-
graphs of the Society for Research in Child Development, 54(3–4), 1–129. doi:10.2307/1166130
Mervis, C. B., & Bertrand, J. (1994). Acquisition of the Novel Name – Nameless Category (N3C) principle.
Child Development, 65, 1646–1662. doi:10.1111/j.1467-8624.1994.tb00840.x
Mohammad, Y. F. O., Nishida, T., & Okada, S. (2009). Unsupervised simultaneous learning of ges-
tures, actions and their associations for human-robot interaction. IROS, 2537–2544. doi:10.1109/
IROS.2009.5353987
Mollica, F., & Piantadosi, S. T. (2017). How data drive early word learning: A cross-linguistic waiting time
analysis. Open Mind: Discoveries in Cognitive Science. doi:10.1162/opmi_a_00006
Monaghan, P., & Christiansen, M. H. (2008). Integration of multiple probabilistic cues in syntax acquisi-
tion. In H. Berends (Ed.), Trends in corpus research: Finding structure in data (pp. 139–163). Amsterdam,The
Netherlands: John Benjamins Publishing Company.
Monaghan, P., & Christiansen, M. H. (2010).Words in puddles of sound: Modelling psycholinguistic effects
in speech segmentation. Journal of Child Language, 37(3), 545–564.
Morewedge, C. K., & Kahneman, D. (2010). Associative processes in intuitive judgment. Trends in Cognitive
Sciences, 14, 435–440. doi:10.1016/j.tics.2010.07.004
Morse, A. F., Benitez,V. L., Belpaeme,T., Cangelosi, A., & Smith, L. B. (2015). Posture affects how robots and
infants map words to objects. PloS One, 10(3), e0116012. doi:10.1371/journal.pone.0116012
Moulin-Frier, C., Nguyen, S. M., & Oudeyer, P. Y. (2014). Self-organization of early vocal development
in infants and machines: The role of intrinsic motivation. Frontiers in Psychology, 4. doi:10.3389/
fpsyg.2013.01006
Nowak, M. A., Komarova, N. L., & Niyogi, P. (2002). Computational and evolutionary aspects of language.
Nature, 417, 611–617. doi:10.1038/nature00771
Oller, D. K. (2000). The emergence of the speech capacity. Mahwah, NJ: Lawrence Erlbaum Associates.
Oudeyer, P-Y. (2010). On the impact of robotics in behavioral and cognitive sciences: From insect naviga-
tion to human cognitive development. IEEE Transactions on Autonomous Mental Development, 2(1), 2–16.
doi:10.1109/TAMD.2009.2039057
Oudeyer, P-Y. (2018a). Self-organization in the evolution of speech (2nd ed.). Oxford: Oxford University Press.
Oudeyer, P-Y. (2018b). Computational theories of curiosity-driven learning. In G. Gordon (Ed.), The New
Science of Curiosity, Hauppauge, NY: Nova Science Publishers.
Oudeyer, P-Y., & Kaplan, F. (2006). Discovering communication. Connection Science, 18(2), 189–206.
doi:10.1080/09540090600768567
Oudeyer, P-Y., Kaplan, F., & Hafner,V.V. (2007). Intrinsic motivation systems for autonomous mental develop-
ment. IEEE Transactions on Evolutionary Computation, 11(2), 265–286. doi:10.1109/TEVC.2006.890271
Oudeyer, P-Y., & Smith, L. B. (2016). How evolution may work through curiosity-driven developmental
process. Topics in Cognitive Science, 8(2), 492–502. doi:10.1111/tops.12196
Pfeifer, R., Lungarella, M., & Iida, F. (2007). Self-organization, embodiment, and biologically inspired
robotics. Science, 318(5853), 1088–1093. doi:10.1126/science.1145803

98
 Computational and robotic models

Piantadosi, S. T. (2014). Zipf ’s word frequency law in natural language: A critical review and future direc-
tions. Psychonomic Bulletin and Review, 21, 1112–1130. doi:10.3758/s13423-014-0585-6
Pinker, S. (1979). Formal models of language learning. Cognition, 1, 217–283. doi:10.1016/0010-
0277(79)90001-5
Pinker, S. (1989). Learnability and cognition:The acquisition of argument structure. Cambridge, MA: MIT Press.
Puglisi, A., Baronchelli, A., & Loreto, V. (2008). Cultural route to the emergence of linguistic categories.
Proceedings of the National Academy of Sciences, 105(23), 7936–7940. doi:10.1073/pnas.0802485105
Regier, T. (2005). The emergence of words: Attentional learning in form and meaning. Cognitive Science,
29(6), 47. doi:10.1207/s15516709cog0000_31
Rohlfing, K. J., Fritsch, J., Wrede, B., & Jungmann, T. (2006). How can multimodal cues from child-
directed interaction reduce learning complexity in robots? Advanced Robotics, 20(10), 1183–1199.
doi:10.1163/156855306778522532
Rohlfing, K. J., Wrede, B., Vollmer, A. L., & Oudeyer, P. Y. (2016). An alternative to mapping a word onto
a concept in language acquisition: Pragmatic frames. Frontiers in Psychology, 7, 470. doi:10.3389/
fpsyg.2016.00470
Roy, D. (2005). Grounding words in perception and action: Computational insights. Trends in Cognitive Sci-
ences, 9(8), 389–396. doi:10.1016/j.tics.2005.06.013
Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996a). Statistical learning by 8-month-old infants. Science,
274(5294), 1926–1928. doi:10.1126/science.274.5294.1926
Saffran, J. R., Newport, E., & Aslin, R. (1996b). Word segmentation: The role of distributional cues. Journal
of Memory and Language, 35(4), 606–621. doi:10.1006/jmla.1996.0032
Samuelson, L. K., Kucker, S. C., & Spencer, J. P. (2017). Moving word learning to a novel space: A dynamic
systems view of referent selection and retention. Cognitive Science, 41(S1), 52–72. doi:10.1111/cogs.12369
Samuelson, L. K., Smith, L. B., Perry, L. K., & Spencer, J. P. (2011). Grounding word learning in space. PLoS
One, 6(12), e28095. doi:10.1371/journal.pone.0028095
Scassellati, B. (1999). Imitation and mechanisms of joint attention: A developmental structure for building
social skills on a humanoid robot. In Computation for metaphors, analogy, and agents (pp. 176–195). Berlin:
Springer.
Schueller, W., & Oudeyer, P. Y. (2016). Active control of complexity growth in naming games: Hearer’s choice. In
S. G. Roberts, C. Cuskley, L. McCrohon, L. Barceló-Coblijn, O.Feher &Verhoef T. (Eds.) The Evolution of
Language: Proceedings of the 11th International Conference (EVOLANG11). doi:10.17617/2.2248195
Schwartz, J. L., Boe, L. J.,Vallée, N., & Abry, C. (1997). The dispersion-focalization theory of vowel systems.
Journal of Phonetics, 25, 255–286. doi:10.1121/1.4786487
Shukla, M., Nespor, M., & Mehler, J. (2007). An interaction between prosody and statistics in the segmenta-
tion of fluent speech. Cognitive Psychology, 54(1), 1–32. doi:10.1016/j.cogpsych.2006.04.002
Siegler, R. S. (1996). Emerging minds: The process of change in children’s thinking. Oxford: Oxford University
Press.
Siskind, J. M. (1996). A computational study of cross-situational techniques for learning word-to-meaning
mappings. Cognition, 61, 39–91. doi:10.1016/S0010-0277(96)00728-7
Smith, L. B. (2000). How to learn words: An associative crane. In R. Golinkoff & K. Hirsh-Pasek (Éds.),
Breaking the word learning barrier (pp. 51–80). Oxford: Oxford University Press.
Smith, L. B., Jayaraman, S., Clerkin, E., & Yu, C. (2018). The developing infant creates a curriculum for
statistical learning. Trends in Cognitive Sciences, 22(4), 325–336. doi:10.1016/j.tics.2018.02.004
Smith, L. B., & Thelen, E. (2003). Development as a dynamic system. Trends in Cognitive Sciences, 7(8),
343–348. doi:10.1016/S1364-6613(03)00156-6
Smith, L. B., & Yu, C. (2008). Infants rapidly learn word-referent mappings via cross-situational statistics.
Cognition, 106, 1558–1568. doi:10.1016/j.cognition.2007.06.010
Smith, L. B.,Yu, C., & Pereira, A. F. (2011). Not your mother’s view: The dynamics of toddler visual experi-
ence. Developmental Science, 14(1), 9–17. doi:10.1111/j.1467-7687.2009.00947.x
Snow, C. E. (1972). Mothers’ speech to children learning language. Child Development, 43(2), 549–565.
doi:10.2307/1127555
Spranger, M. (2012). The co-evolution of basic spatial terms and categories. In L. Steels (Ed.), Experiments
in cultural language evolution (pp. 111–143). Amsterdam, The Netherlands: John Benjamins Publishing
Company.
Spranger, M., & Steels, L. (2012). Emergent functional grammar for space. In L. Steels (Ed.), Experiments
in cultural language evolution (pp. 207–232). Amsterdam, The Netherlands: John Benjamins Publishing
Company.

99
 Pierre-Yves Oudeyer et al.

Steels, L. (1997). The synthetic modeling of language origins. Evolution of Communication, 1(1), 1–35.
Steels, L. (2001). Language games for autonomous robots. IEEE Intelligent Systems, 16(5), 16–22.
doi:10.1109/MIS.2001.956077
Steels, L. (2003). Evolving grounded communication for robots. Trends in Cognitive Sciences. 7(7) 308–312.
doi:10.1016/S1364-6613(03)00129-3
Steels, L. (Ed.). (2012). Experiments in Cultural Language Evolution. Amsterdam, The Netherlands: John Ben-
jamins Publishing Company.
Steels, L., & Belpaeme, T. (2005). Coordinating perceptually grounded categories through language: A case
study for colour. Behavioral and Brain Sciences, 28(4), 469–488. doi:10.1017/S0140525X05250082
Steels, L., & Kaplan, F. (2000). AIBO’s first words: The social learning of language and meaning. Evolution of
Communication, 4(1), 3–32. doi:10.1075/eoc.4.1.03ste
Stevens, K. N. (1972). The quantal nature of speech: Evidence from articulatory-acoustic data. In D. David
(Eds.), Human communication: A unified view (pp. 51–66). New York, NY: McGraw-Hill.
Suanda, S. H., & Namy, L. L. (2012). Detailed behavioral analysis as a window into cross-situational word
learning. Cognitive Science, 36(3), 545–559. doi:10.1111/j.1551-6709.2011.01218.x
Sugita,Y., & Tani, J. (2005). Learning semantic combinatoriality from the interaction between linguistic and
behavioral processes. Adaptive Behavior, 13(1), 33–52. doi:10.1177/105971230501300102
Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman, N. D. (2011). How to grow a mind: Statistics,
structure, and abstraction. Science, 331(6022), 1279–1285. doi:10.1126/science.1192788
Thiessen, E. D., Hill, E. A., & Saffran, J. R. (2005). Infant directed speech facilitates word segmentation.
Infancy, 7, 49–67. doi:10.1207/s15327078in0701_5
Thomas, M. S., Forrester, N. A., & Ronald, A. (2016). Multiscale modeling of gene – behavior associa-
tions in an artificial neural network model of cognitive development. Cognitive Science, 40(1), 51–99.
doi:10.1111/cogs.12230
Thomas, M. S., & Knowland, V. C. (2014). Modeling mechanisms of persisting and resolving delay
in language development. Journal of Speech, Language, and Hearing Research, 57(2), 467–483.
doi:10.1044/2013_JSLHR-L-12-0254
Todd, P. M., & Gigerenzer, G. (2000). Précis of simple heuristics that make us smart. Behavioral and Brain
Sciences, 23(5), 727–741.
Tomasello, M. (1988). The role of joint attentional processes in early language development. Language Sci-
ences, 10, 69–88. doi:10.1016/0388-0001(88)90006-X
Triesch, J. (2013). Imitation learning based on an intrinsic motivation mechanism for efficient coding.
Frontiers in Psychology, 4, 800. doi:10.3389/fpsyg.2013.00800
Trueswell, J. C., Medina, T. N., Hafri, A., & Gleitman, L. R. (2013). Propose but verify: Fast map-
ping meets cross-situational word learning. Cognitive Psychology, 66(1), 126–156. doi:10.1016/j.
cogpsych.2012.10.001
Twomey, K. E., Morse, A. F., Cangelosi, A., & Horst, J. S. (2016). Children’s referent selection and word
learning: Insights from a developmental robotic system. Interaction Studies: Social Behaviour and Commu-
nication in Biological and Artificial Systems, 17(1), 93–119. doi:10.1075/is.17.1.05two
VanDam, M.,Warlaumont, A. S., Bergelson, E., Cristia, A., Soderstrom, M., De Palma, P., & MacWhinney, B.
(2016). HomeBank: An online repository of daylong child-centered audio recordings. Seminars in Speech
and Language, 37(2), 128–142. doi:10.1055/s-0036-1580745
Van Trijp, R. (2012). The evolution of case systems for marking event structure. In L. Steels (Ed.), Experi-
ments in cultural language evolution (pp. 169–205). Amsterdam, The Netherlands: John Benjamins.
Vogt, P. P. (2012). Exploring the robustness of cross-situational learning under zipfian distributions. Cognitive
Science, 36(4), 726–739. doi:10.1111/j.1551-6709.2011.1226.x
Vollmer, A. L., & Schillingmann, L. (2017). On studying human teaching behavior with robots: A review.
Review of Philosophy and Psychology, 1–41. doi:10.1007/s13164-017-0353-4
Warlaumont, A. S. (2013). Salience-based reinforcement of a spiking neural network leads to increased syllable pro-
duction. Development and Learning and Epigenetic Robotics (ICDL), IEEE Third Joint International
Conference on (pp. 1–7). IEEE.
Warlaumont, A. S., Westermann, G., Buder, E. H., & Oller, D. K. (2013). Prespeech motor learning in a
neural network using reinforcement. Neural Networks, 38, 64–75. doi:10.1016/j.neunet.2012.11.012
Westermann, G., & Mareschal, D. (2014). From perceptual to language-mediated categorization. Philosophi-
cal Transactions of the Royal Society B: Biological Sciences, 369(1634). doi:10.1098/rstb.2012.0391
Westermann, G., & Twomey, K. E. (2017). Computational models of word learning. In G. Westermann &
N. Mani (Eds.), Early word learning (pp. 138–154). Oxon: Routledge.

100
 Computational and robotic models

White, R. (1959). Motivation reconsidered: The concept of competence. Psychological Review, 66, 297–333.
doi:10.1037/h0040934
Yang, C. (2011). Computational models of syntactic acquisition. Wiley Interdisciplinary Reviews: Cognitive
Science, 3(2), 205–213. doi:10.1002/wcs.1154
Yu, C., & Smith, L. B. (2007). Rapid word learning under uncertainty via cross-situational statistics. Psycho-
logical Science, 18, 414–420. doi:10.1111/j.1467-9280.2007.01915.x
Yu, C., & Smith, L. B. (2012). Modeling cross-situational word-referent learning: Prior questions. Psychologi-
cal Review, 119(1), 21–39. doi:10.1037/a0026182
Yurovsky, D., & Frank, M. C. (2015). An integrative account of constraints on cross-situational word learn-
ing. Cognition, 145, 53–62. doi:10.1016/j.cognition.2015.07.013
Zipf, G. (1949). Human behavior and the principle of least effort. Cambridge, MA: Addison-Wesley.

101