ArticlePDF Available

Guidelines for Designing Social Robots as Second Language Tutors

Authors:
  • Ghent University/University of Plymouth

Abstract and Figures

In recent years, it has been suggested that social robots have potential as tutors and educators for both children and adults. While robots have been shown to be effective in teaching knowledge and skill-based topics, we wish to explore how social robots can be used to tutor a second language to young children. As language learning relies on situated, grounded and social learning, in which interaction and repeated practice are central, social robots hold promise as educational tools for supporting second language learning. This paper surveys the developmental psychology of second language learning and suggests an agenda to study how core concepts of second language learning can be taught by a social robot. It suggests guidelines for designing robot tutors based on observations of second language learning in human–human scenarios, various technical aspects and early studies regarding the effectiveness of social robots as second language tutors.
This content is subject to copyright. Terms and conditions apply.
International Journal of Social Robotics (2018) 10:325–341
https://doi.org/10.1007/s12369-018-0467-6
Guidelines for Designing Social Robots as Second Language Tutors
Tony Belpaeme1,2 ·Paul Vogt3·Rianne van den Berghe5·Kirsten Bergmann4·Tilbe Göksun6·
Mirjam de Haas3·Junko Kanero6·James Kennedy1·Aylin C. Küntay6·Ora Oudgenoeg-Paz5·
Fotios Papadopoulos1·Thorsten Schodde4·Josje Verhagen5·Christopher D. Wallbridge1·Bram Willemsen3·
Jan de Wit3·Vasfiye Geçkin6·Laura Hoffmann4·Stefan Kopp4·Emiel Krahmer3·Ezgi Mamus6·
Jean-Marc Montanier7·Cansu Oranç6·Amit Kumar Pandey7
Accepted: 11 January 2018 / Published online: 25 January 2018
© The Author(s) 2018. This article is an open access publication
Abstract
In recent years, it has been suggested that social robots have potential as tutors and educators for both children and adults.
While robots have been shown to be effective in teaching knowledge and skill-based topics, we wish to explore how social
robots can be used to tutor a second language to young children. As language learning relies on situated, grounded and social
learning, in which interaction and repeated practice are central, social robots hold promise as educational tools for supporting
second language learning. This paper surveys the developmental psychology of second language learning and suggests an
agenda to study how core concepts of second language learning can be taught by a social robot. It suggests guidelines for
designing robot tutors based on observations of second language learning in human–human scenarios, various technical
aspects and early studies regarding the effectiveness of social robots as second language tutors.
Keywords Social robot ·Second language learning ·Robot tutor ·Human–robot interaction
1 Introduction
One of the goals of Human–Robot Interaction (HRI) is to
research and develop autonomous social robots as tutors
that are able to support children learning new skills effec-
tively through repeated interactions. To achieve this, the
interactions between child and robot should be pleasant,
challenging, and pedagogically sound. Interactions need to
BPaul Vogt
p.a.vogt@uvt.nl
1Centre for Robotics and Neural Systems, Plymouth
University, Plymouth, UK
2IDLab – imec, Ghent University, Ghent, Belgium
3Tilburg Center for Cognition and Communication, Tilburg
University, Tilburg, The Netherlands
4Cluster of Excellence Cognitive Interaction Technology,
Bielefeld University, Bielefeld, Germany
5Department of Special Education: Cognitive and Motor
Disabilities, Utrecht University, Utrecht, The Netherlands
6Department of Psychology, College of Social Sciences and
Humanities, Koç University, Istanbul, Turkey
7SoftBank Robotics, Paris, France
be pleasant for children to enjoy, challenging so that chil-
dren remain motivated to learn new skills, and pedagogically
sound to ensure that children receive input that optimises their
learning gain. One domain in which robots for learning are
developed is second language (L2) tutoring (e.g., [1,33,64]).
While much progress has been made in this field, there has
not been an effective one-on-one L2 tutoring programme that
can be structurally applied in educational settings for various
language communities.
The L2TOR project1(pronounced as ‘el tutor’) aims to
bridge this gap by developing a lesson series that helps
preschool children, around the age of 5 years, learn basic
vocabulary in an L2 using an autonomous social robot as
tutor [8]. In particular, we develop one-on-one, personalised
interactions between children and the SoftBank NAO robot
for teaching English to native speakers of Dutch, German,
and Turkish, and for teaching Dutch or German to Turkish-
speaking children living in the Netherlands or Germany. To
ensure a pedagogically sound programme, lessons are being
developed in close collaboration with developmental psy-
chologists and pedagogists.
1http://www.l2tor.eu.
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
326 International Journal of Social Robotics (2018) 10:325–341
Personalising the interactions between child and robot is
crucial for successful tutoring [45]. Personalisation can be
achieved by creating some initial common ground between
child and robot, and by having the robot adapt to the individ-
ual progress of children. Constructing initial common ground
helps to promote long-term interactions between child and
robot [33], and can be achieved by framing the robot as a
peer and by explaining (dis)similarities between robots and
humans. However, to keep children motivated to learn, it
is important to keep the learning targets within the child’s
Zone of Proximal Development [70]. Throughout the lessons
the target should be sufficiently challenging for the child:
not too challenging as this may frustrate the learner and not
too easy as this may bore the learner. Moreover, interactions
should be designed such that the robot provides a scaffold that
allows the child to acquire the desired language skills. For
instance, by providing non-verbal cues (e.g., gestures) that
help to identify a word’s referent or by providing appropriate
feedback, it is possible for children to reinforce successfully
acquired skills or to correct suboptimal (or wrong) skills.
The L2TOR approach relies on the current state-of-the-
art in HRI technology, which offers promising opportunities,
but also poses major challenges. For instance, NAO has the
ability to produce speech in various languages, making it
possible for the robot to address the child in both the native
language (L1) and in the L2. However, at present, automatic
speech recognition (ASR) for child speech is not performing
to a sufficiently reliable standard, and thus using ASR is cur-
rently infeasible [37]. This not only limits the ability to rely
on verbal interactions since the robot is unable to respond to
children’s speech, but it also limits the ability to monitor and
respond to children’s L2 productions. Hence, our design has
to find ways to work around such technological limitations.
The paper aims to present a number of guidelines that
help researchers and developers to design their own social
robot, especially for, though not necessarily limited to, L2
tutoring. After a brief review of L2 learning from a devel-
opmental psychology point of view, Sect. 3reviews some
previous research on language tutoring using social robots.
In Sect. 4, we will present our guidelines relating to pedagog-
ical considerations, child–robot interactions and interaction
management. These issues will be discussed in light of some
of our early experiments. Section 5discusses our approach to
evaluating the L2TOR system, which is designed to demon-
strate the (potential) added value of using social robots for
L2 tutoring.
2 Second Language Learning
Learning an L2 is important in today’s society. In the Euro-
pean Union (EU), for example, 54 percent of the population
can hold a conversation in at least two languages, and 25 per-
cent are able to speak three languages [20]. Consequently,
L2 teaching has become an essential part of primary educa-
tion. In 2002, the EU proposed a multilingualism policy of
teaching an L2 to all young children. The policy suggests
every European citizen learns practical skills in at least two
languages aside from their L1 [4]. According to a recent sur-
vey, the vast majority of European citizens (98 percent of the
respondents in this survey) believe that mastering a foreign
language is useful for the future of their children [20].
Preschool years are vital for L2 learning, because later
academic success depends on early language skills [29].
For children learning English as their school language, their
English vocabulary size predicts their performance in English
reading tests [57]. Although learning an L2 comes naturally
for some children, for many others it is a challenge that they
must overcome. For children from immigrant families or
minority communities, the language used at school is often
different from the language used at home. These children,
thus, not only start learning the school language later than
their peers, but also continue to receive relatively less input
in each of their languages [30]. Hence, novel ways to expose
children to targeted L2 input must be considered.
Patterns of L2 learning largely mirror those of L1 learning,
which requires both the quantity and the quality of language
input to be sufficient [27]. Children do not learn language
just by listening to speech; rather, interactive experience is
essential [39]. L2 learning is no exception, and several fac-
tors such as interactivity must be considered (see [38]for
a review). In addition to quantity, socio-pragmatic forms
of interaction involving joint attention, non-verbal interac-
tion, feedback, and temporal and semantic contingencies are
expected to contribute to L2 learning [3,9,59,66]. However,
there are also some notable differences between L1 and L2
learning. For example, in L2 education it is important to con-
sider from whom children are learning the L2. Place and Hoff
[56] found that hearing English from different speakers and
the amount of English input provided by native speakers is
critical for learning English as L2. Another notable differ-
ence between L1 and L2 learning is that children may rely
on their L1 when learning an L2 (e.g., [75]). Thus, we may
need to be cautious about factors such as negative transfer or
interference, in which some concepts and grammar in the L2
are hard to acquire because children are thinking in their L1
[67].
When children are learning more than one language, the
amount of input a child hears in each language predicts
vocabulary size in each language [30,55]. Bilingual chil-
dren tend to have a smaller vocabulary size in each language
compared to their monolingual peers [54], although the com-
bined or conceptual vocabulary size of both languages is
often equal to that of monolinguals [31,54]. The amount of
language input also affects language processing speed and
trajectories of vocabulary learning, and thus early language
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 327
input may have cascading effects on later language learn-
ing. Hurtado et al. [32] found that the amount of language
input bilingual children receive at 18months of age predicts
their speed of recognizing words and the size of their vocab-
ulary at 24 months. To properly foster development of two
or more languages, adults must carefully consider a good
balance between languages [67].
Although both monolingual and bilingual children moni-
tor and respond to social pragmatic cues, bilingual children
have heightened sensitivity to those non-linguistic cues,
probably due to an early communicative challenge they face
because of less than perfect mastery in one of the languages
[74]. Brojde et al. [10] found that bilingual children rely
more on eye gaze than their monolingual counterparts when
learning novel words. Yow and Markman [76] also demon-
strated that 3- and 4-year-old bilingual children were better at
understanding and using gestures and gaze direction to infer
referential intent. Thus, especially for children with advanced
L2 knowledge, we may be able to boost their learning process
by making use of these pragmatic cues.
As the demand for early L2 education increases, the usage
of additional teaching opportunities in terms of educational
tablet games, or electronic vocabulary trainers becomes more
and more important to increase the quantity of L2 input.
Moreover, especially with regard to young children, the con-
sideration of embodied technologies (e.g., virtual agents or
robots) seems reasonable, because they invite intuitive inter-
actions that would add to the quality of the L2 input. The
question then becomes: how should such a robot be designed?
3 Robots for Language Tutoring
In recent years, various projects have started to investigate
how robot tutors can contribute to (second) language learn-
ing. In this section, we review some of these studies, focusing
on: (a) the evidence that robots can promote learning; (b) the
role of embodiment in robot tutoring; and (c) the role of social
interactions in tutoring.
3.1 Learning from Robots
There has been an increased focus on how social robots
may help engage children in learning activities. Robots have
been shown to help increase interaction levels in larger
classrooms, correlating with an improvement in children’s
language learning ability [22]. How best to apply this knowl-
edge in the teaching of a foreign language has been explored
by different researchers from various perspectives. Alemi et
al. [1] employed a social robot as an assistant to a teacher
over a 5-week period to teach English vocabulary to Ira-
nian students. They found that the class with the robot
assistant learned significantly more than that with just the
human teacher. In addition, the robot-assisted group showed
improved retention of the acquired vocabulary. This builds
on earlier findings by [33] where a 2-week study with a robot
situated in the classroom revealed a positive relation between
interacting with a robot and vocabulary acquisition. Further
results by [64] also confirm that the presence of a robot leads
to a significant increase in acquired vocabulary. Movellan
et al. [50] selected 10 words to be taught by a robot, which
was left in the children’s classroom for 12 days. At the end
of the study, children showed a significant increase in the
number of acquired words when taught by the robot. Lee et
al. [42] further demonstrated that robot tutoring can lead not
just to vocabulary gains, but also improved speaking ability.
In their study, children would start with a lesson delivered
by a computer, then proceed to pronunciation training with
a robot. The robot would detect words with an expanded
lexicon based on commonly confused phonemes and correct
the child’s pronunciation. Additionally, the children’s confi-
dence in learning English was improved.
All of these studies show the capacity of various robots
as tutors for children (with the children’s age ranging from 3
to 12 years old) learning an L1 or L2 ‘in the wild’. How-
ever, what exactly is it that gives robots the capacity for
tutoring? Moreover, how does this compare to other digi-
tal technologies, such as tablets and on-screen agents? Is it
merely the embodiment of the robot, or rather the quality
of social interactions? These questions are explored in the
following sections.
3.2 Embodiment
The impact of embodiment and social behaviour for children
learning English as their L1 has been explored in a laboratory
setting. Neither [24] nor [71] found significant differences
due to the embodiment of the robot in their studies on chil-
dren’s vocabulary acquisition. However, this may be due in
part to methodological limitations. Gordon et al. [24] only
found an average of one word learned per interaction, leaving
very little room for observing differences; similarly [71] only
compared the learning of six words. These studies were con-
ducted with children between the ages of 3 and 8 years. The
relatively small gains are therefore quite surprising, due to
the speed at which children at this age acquire language [40].
Given the non-significant results or the small effect sizes in
these studies, it is difficult to draw conclusions on what could
make robot language tutoring effective.
Rosenthal-von der Pütten et al. [58] found that language
alignment, i.e., the use of similar verbal patterns between
interacting parties, when using an L2 appears to not be
affected when using a virtual robot as opposed to a real one.
Participants completed a pre-test and were then invited for
a second session at a later date. During the second session
the participants were asked to play a guessing game with an
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
328 International Journal of Social Robotics (2018) 10:325–341
agent, either the real NAO robot or a virtual representation
of one. The study reported whether the participants used the
same words as the agent, but no significant difference was
found. This may be due to some issues with the experimental
design: the authors suggest the post-test was given straight
after a relatively long session with the agent, and participants
may have been fatigued.
Moriguchi et al. [49] looked at age differences for young
children and how they learned from a robot compared to
a person. Children between the ages of 4 and 5 years were
taught using an instructional video: one group of children
was shown a video in which a human taught them new words,
while another group of children was shown a video with the
same material, but using a robot tutor. While children aged 5
were able to perform almost as well when taught by a robot,
those aged 4 did not seem to learn from the robot at all. It is
unknown as to whether this result would transfer to the use of
a physically-present robot, rather than one shown on a video
screen.
These studies above do not provide support that the mere
physical presence of the robot has an advantage for language
learning. However, there is evidence for the physical presence
of a robot having a positive impact on various interaction
outcomes, including learning [46]. The lack of a clear effect
of a physical robot on language learning might be due to
a scarcity of experimental data. However, it is also likely
that the effectiveness of robot tutors lies not in their physical
presence, but instead in the social behaviour that a robot
can exhibit and the motivational benefits this carries. This
is explored in the next section.
3.3 Social Behaviour
Social behaviour has previously been studied in the con-
text of children learning languages. Saerbeck et al. [60]
explored the impact of ‘socially supportive’ behaviours on
child learning of the Toki Pona language, using an iCat robot
as a tutor. These behaviours included verbal and non-verbal
manipulations which aimed to influence feedback provision,
attention guiding, empathy, and communicativeness. It was
found that the tutor with these socially supportive behaviours
significantly increased the child’s learning potential when
compared to a neutral tutor. This study used a variety of mea-
sures including vocabulary acquisition, as other studies have,
but also included pronunciation and grammar tests. Another
study which did not only consider vocabulary acquisition
was [26]. French and Latin verb conjugations were taught
by a NAO robot to children aged 10 to 12 years old. In one
condition, the robot would look towards the student whilst
they completed worksheets, but in the other, the robot would
look away. Although gaze towards the child was predicted to
lead to greater social facilitation effects, and therefore higher
performance, this was not observed.
Kennedy et al. [36] investigated the effects of verbal
immediacy on the effect of learning in children. A NAO was
used to teach French to English-speaking children in a task
involving the gender of nouns and the use of articles ‘le’
and ‘la’. A high verbal immediacy condition was designed
in which the robot would exhibit several verbal immediacy
behaviours, for example calling the child by name, providing
positive feedback, and asking children how they felt about
their learning. When contrasted with a robot without this
behaviour, no significant learning differences were observed.
However, children showed significant improvement in both
conditions when comparing pre- and post-test scores, and
were able to retain their acquired knowledge as measured by
means of a retention test. This suggests that the particularities
of robot behaviour do not manifest themselves in the short-
term, but could be potentially be observed over the longer
term.
In [2], a robot acted as a teaching assistant for the purpose
of teaching English to Iranian students. A survey found that
students who were taught by the robot were significantly less
anxious about their lessons than those that were not. This was
thought to be due to a number of factors, including the fact
that the robot was programmed to make intentional mistakes
which the students could correct, which could have made
students less concerned about their own mistakes.
3.4 Summary
In summary, promising results have been found for the use of
robots as constrained language tutors for children and adults,
with the presence of the robot improving learning outcomes
[1,2,33,64]. However, the impact of robot embodiment in this
context has not been explored in depth, leaving an important
question largely unanswered: do robots hold an advantage
over tablets or virtual characters for language tutoring? The
impact of social behaviour is also less clear, with some pos-
itive results [60], but also inconclusive results [26]. Robots
open up new possibilities in teaching that were previously
unavailable, such as the robot taking the role of a peer. By
having an agent that is less like a teacher and more like a peer,
anxiety when learning a new language could be reduced [2].
Despite an increasing interest, there are still relatively few
studies that have considered robot language tutoring, leaving
space to explore novel aspects of language learning.
4 Designing Robot Tutoring Interactions for
Children
Several design issues with respect to robot-guided L2 tutor-
ing have to be considered before an evaluation of robot-child
tutoring success is possible. In particular, multiple design
choices have to be considered to create pleasant, challeng-
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 329
ing, and pedagogically sound interactions between robot and
child [69]. First, we will discuss pedagogical issues that
ensure optimal conditions for language learning. Second,
we will present various design issues specifically relating
to the child–robot interactions. Finally, we will discuss how
to manage personalised interactions during tutoring. The sec-
tion builds on some related work as well as various studies
conducted in the context of the L2TOR project.
4.1 Pedagogical Issues
It is imperative to understand how previous research findings
can be put into practice to support successful L2 acquisition.
Although the process of language learning does not dras-
tically differ between L1 and L2, there are a few notable
differences as we already discussed in Sect. 2. For the L2TOR
project a series of pedagogical guidelines was formulated,
based on existing literature and pilot data collected within
our project. These guidelines concern: (a) age differences;
(b) target word selection; (c) the use of a meaningful con-
text and interactions to actively involve the child; and (d) the
dosage of the intervention. These specific aspects were cho-
sen based on a review of the literature showing that they are
the most crucial factors to consider in designing an interven-
tion for language teaching in general and specifically L2 (see
e.g., [29,51]).
4.1.1 Age Effects
From what age onward can we use social robots to support
L2 learning effectively? From a pedagogical point of view,
it is desirable to start L2 tutoring as early as possible, espe-
cially for children whose school language is an L2, because
this could bridge the gap in language proficiency that they
often have when entering primary school [29]. Various stud-
ies have targeted children as young as 3years focusing on
interactive storytelling in the L1 [22]oronL2tutoring[73].
However, preschool-aged children (3 to 5 years old) undergo
major cognitive, emotional and social developments, such as
the expansion of their social competence [15]. So, whereas
older children may have little difficulty engaging in an inter-
action with a robot, younger children may be more reliant
on their caregivers or show less engagement in the interac-
tion. Therefore, we may expect that child–robot interactions
at those ages will also present some age-related variation.
Clarifying these potential age differences is essential as, in
order to be efficient, interactive scenarios with robots must
be tailored to the diverging needs of children.
In [6], we sought to determine whether there are age-
related differences in first-time interactions with a peer-tutor
robot of children who have just turned 3 and children
who are almost 4 years old. To this end, we analysed the
engagement of 17 younger children (Mage =3.1 years,
SD
age =2 months) and 15 older children (Mage =3.8 years,
SD
age =1 month) with a NAO robot as part of the larger
feedback experiment discussed in Sect. 4.2.6. These children
first took part in a group introduction to familiarise them with
the NAO robot; a week later they had a one-on-one tutoring
session with the robot. We analysed the introductory part of
this one-on-one session, which consisted of greeting, bond-
ing with, and counting blocks with the robot. All speech was
delivered in Dutch, except for the target words (i.e., ‘one’,
‘two’, ‘three’, and ‘four’), which were provided in English.
We analysed the children’s engagement with the robot as
measured through eye-gaze towards the task environment
(robot and blocks) compared to their gazes outside the task
environment (experimenter, self, and elsewhere), as this is
suggested to indicate how well the child is “connected” with
thetask[62].
In short, the analyses revealed that the older children gazed
significantly longer towards the robot than the younger chil-
dren, and that the younger children spent more time looking
elsewhere than the older children. Moreover, the average time
the older children maintained each gaze towards the robot
was longer than that of the younger children.
It is possible that the 3-year-olds have trouble being
engaged with a language learning task, but it may also be
that the NAO robot is somewhat intimidating for 3-year-olds.
As such, for them either group interactions [22]oramore
“huggable” robot (e.g., Tega) [73] could be more appropri-
ate. Moreover, [49] also found children at the age of 5 years
to be more responsive to robot tutoring. Drawing from these
findings about 3-year-olds, combined with experiences from
other pilots with 4- and 5-year-olds, we decided to develop
the L2TOR tutoring system for 5-year-olds, as they generally
appear to feel more comfortable engaging one-on-one with
the robot than 3- and 4-year-olds.
4.1.2 Target Words
Another important aspect to consider is what words are
taught. Previous research recommends that vocabulary items
should be taught in semantic clusters and embedded in a con-
ceptual domain [11,51]. For L2TOR, three domains were
chosen: (a) number domain: language about basic number
and pre-mathematical concepts; (b) space domain: language
about basic spatial relations; and (c) mental states domain:
language about mental representations such as ‘being happy’
and propositional attitudes such as ‘believe’ or ‘like’. These
domains were selected for their feasibility, as well as their rel-
evance and applicability in L2 tutoring sessions in a preschool
setting. Appropriate words to be taught for each domain are
words that children should be familiar with in their L1, as the
goal of the intervention is not to teach children new math-
ematical, spatial, and mental state concepts, but rather L2
labels for familiar concepts in these three domains. This will
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
330 International Journal of Social Robotics (2018) 10:325–341
enable children to use their L1 conceptual knowledge to sup-
port the learning of L2 words. To select appropriate target
words and expressions that children are familiar with in their
L1, a number of frequently used curricula, standard tests,
and language corpora were used. These sources were used
both for identifying potential targets, and for checking them
against age norms to see whether they were suitable for the
current age group (for more details, see [53]). Thus, target
words selection should be based both on semantic coherence
and relevance to the content domain and on children’s L1
vocabulary knowledge.
4.1.3 Meaningful Interaction
An additional aspect of L2 teaching is the way in which new
words are introduced, which may come to affect both learn-
ing gains as well as the level of engagement. Research has
indicated that explicit instruction on target words in mean-
ingful dialogues involving defining and embedding words
in a meaningful context yields higher word learning rates
than implicit instruction through fast mapping (i.e., mapping
of a word label on its referent after only one exposure) or
extracting meaning from multiple uses of a word in context
as the basic word learning mechanisms [48,51]. Therefore,
for the L2TOR project, an overall theme for the lessons was
selected that would be familiar and appealing to most chil-
dren, and, as such, increase childrens engagement during the
tutoring sessions. This overall theme is a virtual town that the
child and the robot explore together, and that contains var-
ious shops, buildings, and areas, which will be discovered
one-by-one as the lesson series progresses. All locations are
familiar to young children, such as a zoo and a bakery. Dur-
ing the lessons, the robot and the child discover the locations,
and learn L2 words by playing games and performing simple
tasks (e.g., counting objects or matching a picture and a spe-
cific target word). The child and the robot are awarded a star
after each completed session, to keep children engaged in the
tasks and in interacting with the robot. Thus, the design cho-
sen for L2TOR is thought to facilitate higher learning gains
as it involves explicit teaching of target words in a dialogue
taking place in a meaningful context. Moreover, this design
should facilitate engagement as it involves settings that are
known and liked by children.
4.1.4 Dosage of Language Input
The final pedagogical aspect that was identified in the lit-
erature concerns the length and intensity, or dosage, of the
intervention. Previous research has shown that vocabulary
interventions covering a period of 10 to 15 weeks with one
to four short 15- to 20-min sessions per week are most effec-
tive. As for the number of novel words presented per session,
the common practice is to offer 5 to 10 words per session, at
least in L1 vocabulary interventions [47]. However, not much
is known about possible differences between L1 and L2 inter-
ventions with regard to this aspect. Therefore, to determine
the number of target words to be presented in the L2TOR
project lesson series, a pilot study was conducted. In this
study, we taught English wordsto one hundred 4- and 5-year-
old Dutch children with no prior knowledge of English. We
started by teaching the children 10 words; when these were
established, more words were added. The results showed that,
for children to learn any of these words at all, the maximum
number of L2 words that could be presented in one session
was six. We also found that a high number of repeated presen-
tations of each word was necessary for word learning: each
word in our study was presented 10 times. Yet, children’s
accuracy rates in the translation and comprehension tasks in
our study were lower than in earlier work on L1 learning. A
possible explanation might be that the items included in the
study were relatively complex L2 words (e.g., adjectives like
‘empty’) rather than concrete nouns such as ‘dog’ or ‘house’.
These items are probably more difficult for children who had
no prior exposure to the target language. However, within the
L2TOR project the choice was made to include these rela-
tively complex items given their relevance for L2 learning
within an academic context [52]. Thus, it was decided that
in all the lessons included within the L2TOR project a max-
imum of six words will be presented in each lesson and each
word will be repeated at least ten times throughout the lesson.
4.2 Child–Robot Interaction Issues
Not only pedagogical issues need to be considered when
designing a social robot tutor, but also other issues relating
to how the interactions between the robot and child should
be designed. As mentioned, we focus on how to design the
interactions to be pleasant, challenging, and pedagogically
sound. In this section, we discuss six aspects that we deem
important: (a) first encounters; (b) the role of the robot; (c)
the context in which the interactions take place; (d) the non-
verbal behaviours and (e) verbal behaviours of the robot; and
(f) the feedback provided by the robot.
Before elaborating on these guidelines, it is important to
remind the reader that in L2TOR, we are designing the robot
to operate fully autonomously. Ideally, this would include
the possibility to address the robot in spoken language and
that the robot can respond appropriately to this. However,
as previously mentioned, current state-of-the-art in speech
recognition for child speech does not work reliably. Kennedy
et al. [37] compared several contemporary ASR technolo-
gies and have found that none of them achieve a recognition
accuracy that would allow for a reliable interaction between
children and robots. We have therefore decided to mediate the
interactions using a tablet that can both display the learning
context (e.g., target objects) and monitor children’s responses
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 331
to questions. This has the consequence that the robot cannot
monitor children’s L2 production autonomously, but it can
monitor children’s L2 comprehension through their perfor-
mance with respect to the lesson content presented on the
tablet.
4.2.1 Introducing the Robot
The first encounter between robot and child plays a large
role in building the child’s trust and rapport with the robot,
and to create a safe environment [72], which are necessary
to facilitate long-term interactions effectively. For example,
[21] has shown that a group introduction in the kindergarten
prior to one-on-one interactions with the robot influenced
the subsequent interactions positively. Moreover, [72]have
shown that introducing the robot in a one-to-many setting
was more appreciated than in a one-on-one setting, because
the familiarity with their peers can reduce possible anxiety
in children.
We, therefore, developed a short session in which the robot
is introduced to children in small interactive groups. In this
session, the experimenter (or teacher) first tells a short story
about the robot using a picture book, explaining certain sim-
ilarities and dissimilarities between the robot and humans
in order to establish some initial common ground [14,33].
During this story, the robot is brought into the room while
in an animated mode (i.e., turned on and actively looking
around) to familiarise the children with the robot’s physical
behaviour. The children and the robot then jointly engage in a
meet-and-greet session, shaking hands and dancing together.
We observed in various trials that almost all children were
happy to engage with the robot during the group session,
including those who were a bit anxious at first, meaning
these children likely benefited from their peers’ confidence.
Although we did not test this experimentally, our introduction
seems to have a beneficial effect on children’s one-on-one
interaction with the robot.
4.2.2 Framing the Robot
One of the questions that arises when designing a robot tutor
is: How should the robot be framed to children, such that
interactions are perceived to be fun, while at the same time
be effective to achieve language learning? We believe it is
beneficial to frame the robot as a peer [5,7,24], because chil-
dren are attracted to various attributes of a robot [33] and
tend to treat a robot as a peer in long-term interactions [64].
Moreover, framing the robot as a peer could make it more
acceptable when the flow of the interaction is suboptimal
due to technical limitations of the robot (e.g., the robot being
slow to respond or having difficulty interpreting children’s
behaviours). In addition, framing the robot as a peer who
learns the new language together with the child sets the stage
for learning by teaching [64].
While the robot is framed as a peer and behaves like a
friend of the child, the tutoring interactions will be designed
based on adult-like strategies to provide the high quality input
children need to acquire an L2 [39], such as providing timely
and sensible non-verbal cues or feedback. So, in L2TOR we
frame the robot as a peer, it behaves like a peer, but it scaffolds
the learning using adult-like teaching strategies.
4.2.3 Interaction Context
To facilitate language learning, it is important to create a con-
textual setting that provides references to the target words
to be learned. The embodied cognition approach, on which
we base our project, states that language is grounded in real-
life sensorimotor interactions [28], and consequently predicts
that childrens interactions with real-life objects will benefit
vocabulary learning [23]. From this approach, one would
expect children to learn new words better if they manipu-
late physical objects rather than virtual objects on a tablet,
as the former allows children to experience sensorimotor
interactions with the objects. However, for technical reasons
discussed earlier, it would be convenient to use a tablet com-
puter to display the context and allow children to interact
with the objects displayed there. The question is whether
this would negatively affect learning. Here, we summarise
the results from an experiment comparing the effect of real
objects versus virtual objects on a tablet screen on L2 word
learning [68]. The main research question is whether there is
a difference in L2 vocabulary learning gain between children
who manipulate physical objects and children who manipu-
late 3D models of the same objects on a tablet screen.
In this experiment, 46 Dutch preschoolers (Mage =5.1
years, SD
age =6.8 months; 26 girls) were presented with a
story in Dutch containing six L2 (English) target words (i.e.,
‘heavy’, ‘light’, ‘full’, ‘empty’, ‘in front of,’ and ‘behind’).
These targets were chosen as children should benefit from
sensorimotor interactions with objects when learning them.
For example, learning the word ‘heavy’ could be easier when
actually holding a heavy object rather than seeing a 3D model
of this object on a tablet screen. Using a between-subjects
design, children were randomly assigned to either the tablet
or physical objects condition. During training, the target
words were each presented ten times by a human. Various
tests were administered to measure the children’s knowledge
of the target words, both immediately after the training and
one week later to measure children’s retention of the target
words.
Independent-samples t-tests revealed no significant dif-
ferences between using a tablet or physical objects on any
of the tasks, as indicated by childrens mean accuracy scores
on the direct and delayed post-tests (see Fig. 1;all pval-
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
332 International Journal of Social Robotics (2018) 10:325–341
Fig. 1 Mean accuracy scores on
the direct post-test (top) and the
delayed post-test (bottom).
Purple bars refer to the object
condition; orange bars to the
tablet condition. Reprinted from
[68]. (Color figure online)
ues >.243). In the receptive tests (the comprehension task
and sorting task), children scored significantly above chance
level (indicated by the black line), irrespective of condition
(all pvalues <.001). Interestingly, in both conditions, the
mean scores on the Dutch-to-English translation task were
higher for the delayed post-test than for the immediate post-
test (both pvalues <.001), possibly indicating some sort of
“sleep effect”. These findings indicate that it does not mat-
ter much whether the context is presented through physical
objects or a tablet computer.
Displaying the context (i.e., target objects) on a tablet does
not seem to hamper learning, which is convenient, since using
a tablet makes designing contexts more flexible and reduces
the need to rely on complex object recognition and tracking.
Because of this, the lessons in the L2TOR project are dis-
played on a tablet, which is placed between the child and
the robot (see Fig. 2). This tablet not only displays the target
objects (e.g., a set of elephants in a zoo), but also allows chil-
Fig. 2 The L2TOR setup includes the NAO robot standing to the side
of the child with a tablet in between them
dren to perform actions on these objects (e.g., placing a given
number of elephants in their cage). Since at present ASR for
children is not performing reliably [37], the robot cannot
monitor children’s pronunciation or other verbal responses.
We therefore focus on language comprehension rather than
language production and use the tablet to monitor compre-
hension. The use of a tablet in the interaction allows us to
monitor the child’s understanding of language and to control
the interaction between child and robot.
4.2.4 Non-verbal Behaviour
Human language production is typically accompanied by
non-verbal cues, such as gestures or facial expressions. It is
therefore not surprising that research in children’s language
development has shown that the use of gestures facilitates L2
learning in various ways (e.g., [25,59,65]). Gestures could
take the form of deictic gestures, such as pointing to refer to
physical objects near the child, or of iconic gestures used
to emphasize physical features of objects or actions in a
more representational manner. Such iconic gestures help to
build congruent links between target words and perceptual
or motor information, so learners may benefit not only from
observing gestures, but also by way of execution, such as
enactment and imitation [23,25].
Due to its physical presence in the child’s referential
world, a robot tutor has the ability to use its physical embod-
iment to its advantage when interacting with the child, for
example, through the manipulation of objects in the real
world, or simply through the use of gestures for various com-
municative purposes. We believe that the robot’s ability to
use gestures is one of the primary advantages of a robot as
tutor compared to a tablet computer, since it can enrich the
language learning environment of the child considerably by
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 333
exploiting the embodiment and situatedness of the robot to
facilitate the child’s grounding of the second language.
Even though a growing body of evidence suggests that
non-verbal cues, such as gestures aid learning, translating
human’s non-verbal behaviour to a robot like NAO remains
a challenge, mostly due to hardware constraints. For instance,
the NAO robot is limited by its degrees of freedom and con-
straints with respect to its physical reach, making it unable
to perform certain gestures. Motions may sometimes seem
rigid, causing the robot’s movements to appear artificial
rather than human-like. Especially when certain subtleties
are required when performing a gesture, such shortcomings
are not desirable. A noteworthy complication comes with
the NAO’s hand, which has only three fingers that cannot
move independently of one another. This makes an act such
as finger-counting, which is often used for the purpose of
explaining numbers or quantities, practically impossible.
This, thus, requires a careful design and testing of appro-
priate referential gestures, because otherwise they may harm
learning [35].
4.2.5 Verbal Behaviour
One potential advantage of using digital technologies, such
as robots, is that they can be programmed to speak multiple
languages without an accent. However, NAO’s text-to-speech
engines do generate synthetic voices and have few prosodic
capacities. Yet, studies have shown that children rely on
prosodic cues to comprehend spoken language (e.g., [16]).
Moreover, adults typically use prosodic cues to highlight
important parts of their speech when addressing children.
In addition, the lack of facial cues of the NAO robot may
potentially hinder the auditory-visual perception processes
of both hearing-impaired and normal-hearing children [19].
These limitations pose the question to what extent children
can learn the pronunciation of L2 words sufficiently well.
To explore this, a Wizard-of-Oz (WoZ) experimental pilot
was devised using the NAO robot and a tablet for tutor-
ing and evaluating English children counting up to five in
German. The task involved multiple steps to gradually teach
children to count, in L2, animals shown on screen. First, the
robot-tablet concept was introduced, with the robot describ-
ing content displayed on the tablet screen, and the children
were trained on how and when to provide answers by means
of touching images on said screen. The children then pro-
ceeded with the main task, which involved the counting of
animals, first in English and later in German. The interaction
was managed by using multiple utterances from a WoZ con-
trol panel in order to prompt the children to give the answer
only after they were asked to. The WoZ operator triggered
appropriate help and feedback from the robot to the child
when required. Finally, at the end of the task, the robot asked
the children to count up to five again with the robots help
0
1
2
3
4
5
12345
Rang
Parcipant ID
First Pronunciaon Final Pronunciaon
Fig. 3 Pronunciation ratings from seven German native speakers for 5
child participants. Three of the children improve over the course of the
interaction, although one child has initially accurate pronunciation that
drops over time, possibly due to fatigue
and then without any help at all. The purpose of this step was
to evaluate whether the children were able to remember the
pronunciation of the German numbers and if they were able
to recall them with no support.
Voice and video recordings were used to record the inter-
actions with five children aged 4 to 5 years old. The first
and final repetitions of the children pronouncing the German
words were recorded and rated for accuracy on a 5-point
Likert scale by seven German-native coders; intraclass corre-
lation ICC(2,7)=.914, indicating “excellent” agreement
[13]. Based on these ratings, our preliminary findings are that
repetitions generally improve pronunciation. Several chil-
dren initially find it hard to pronounce German numbers but
they perform better by the end (Fig. 3). This may be because
some children had trouble recalling the German numbers
without help. We believe that the task needs updating to
improve the children’s recall (by, for example, including
more repetitions). In addition, it should be noted that children
generally find it difficult to switch from English to German.
To conclude, children can learn the pronunciation of the
L2 from the robot’s synthetic voice, but we should compare
this to performance ratings of children that have learned the
L2 from native speakers. It is worth noting that they seem
to have some reservation speaking a foreign language, but
whether or not this is due to the presence of the robot is
unknown.
4.2.6 Feedback
A typical adult-like strategy known to support language
learning is the use of appropriate feedback [3]. Adult care-
givers tend to provide positive feedback explicitly (e.g., ‘well
done!’) and negative feedback implicitly by recasting the cor-
rect information (e.g., ‘that is the rabbit, now try again to
touch the monkey’). However, evidence suggests that a peer
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
334 International Journal of Social Robotics (2018) 10:325–341
does not generally provide positive feedback and that they
provide negative feedback explicitly without any correction
(e.g., ‘no, that is wrong!’). So, when the robot is framed as a
peer, should it also provide feedback like a peer?
To explore this, we carried out an experiment to investigate
the effect the type of feedback has on children’s engagement
[17,18]. In the experiment, sixty-five 3-year-old children (30
boys, 35 girls; Mage =3.6 years, SDage =3.6 months)
from different preschools in the Netherlands participated.
Six children stopped with the experiment before it was fin-
ished and were excluded from the data. The children were
randomly assigned to one of three conditions, varying the
type of feedback: adult-like feedback, peer-like feedback,
and no feedback. The adult-like feedback of the robot used
reformulations to correct the children in case they made a
mistake (e.g., ‘three means three’, where the text in ital-
ics represents what the robot said in the L2, here English;
the rest was said in the L1, here Dutch) and positive feed-
back (‘well done!’) when children responded correctly. In the
peer-like condition, only explicit negative feedback without
correction was provided whenever children made a mistake
(‘that is wrong!’) and no feedback was provided when they
responded correctly. In the no feedback condition, the robot
simply continued with the next task without providing any
feedback.
During the experiment, the robot taught the native Dutch-
speaking children counting words one to four in English. The
interaction consisted of an introductory phase followed by
the tutoring phase. During the introductory phase, the target
words (i.e., ‘one’, ‘two’, ‘three’, and ‘four’) were described
and associated with their concept in sentences such as ‘I have
one head’, ‘I have two hands’, ‘I have three fingers’, and
‘there are four blocks’. We analysed the introductory phase
as part of the age-effects study reported in Sect. 4.1.1.Inthe
tutoring phase, the robot asked the child to pick up a cer-
tain number of blocks that had been placed in front of them.
All instructions were provided in Dutch and only the target
words were provided in English. After the child collected the
blocks, the robot provided either adult-like feedback, peer-
like feedback, or no feedback depending on the experimental
condition assigned to the child.
As a result of the relatively low number of repetitions of
the target words over the course of the interaction, we did not
expect to find any effects with respect to learning gain. How-
ever, the objective was not to investigate the effect feedback
has on learning, but rather on the child’s engagement with
the robot as an indicator of learning potential [12]. As for the
age-effect study, we analysed engagement by annotating the
children’s eye-gaze towards the robot, human experimenter,
to the blocks, and elsewhere, and measured the average time
children maintained their gaze each time they looked at one
of these targets.
Fig. 4 Mean duration per gaze to the robot, blocks, experimenter, and
elsewhere for the three feedback conditions
Results from a repeated measures ANOVA indicated that,
on average, the children maintained their gaze significantly
longer at the blocks and the robot than at the experimenter,
regardless of their assigned condition (see Fig. 4).
However, we did not see any significant differences in
the gaze duration across the three conditions. As such, the
way the robot provides feedback does not seem to affect the
engagement of the child with the robot. This would suggest
that, as far as the child’s engagement with the robot and task
is concerned, it does not matter how the robot provides feed-
back or whether the robot provides feedback at all. Hence,
the choice for the type of feedback that the robot should give
can, thus, solely be based on the effect feedback has on learn-
ing gain. Future work will investigate which type of feedback
is most effective for learning.
4.3 Interaction Management
4.3.1 Objective
To realise robot-child tutoring interactions that provide a
pleasant and challenging environment for the child, while
at the same time being effective for L2 learning, interaction
management plays a crucial role.
As children typically lose interest when a lesson is either
too easy or too difficult, personalisation of the lessons to each
child’s performance is very important. The tutor has to struc-
ture the interaction, needs to choose the skills to be trained,
must adjust the difficulty of the learning tasks appropriately,
and has to adapt its verbal and non-verbal behaviour to the
situation. The importance of personalised adjustments in the
robot’s behaviour has been evidenced in research showing
that participants who received personalised lessons from a
robot outperformed others who received non-personalised
training [5,45]. Suboptimal robot behaviour (e.g., too much,
too distracting, mismatching, or in other ways inappropriate)
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 335
can even hamper learning [35]. Therefore lessons should be
adapted to the knowledge state (i.e., level) of the child [70].
Along these lines, the L2TOR approach is to personalise
language tutoring in HRI by integrating knowledge-tracing
into interaction management [61]. This adaptive tutoring
approach is realised in a model of how tutors form men-
tal states of the learners by keeping track of their knowledge
state and selecting the next tutoring actions based on their
likely effects on the learner. For that purpose, an extended
model based on Bayesian Knowledge Tracing was built that
combines knowledge tracing (what the learner learned) and
tutoring actions in one probabilistic model. This allows for
the selection of skills and actions based on notions of optimal-
ity: the desired learner’s knowledge state as well as optimal
task difficulty.
4.3.2 Proposed Model
A heuristic is employed that maximises the beliefs of all skills
while balancing the single skill-beliefs with one another. This
strategy is comparable to the vocabulary learning technique
of spaced repetition as implemented, for instance, in the Leit-
ner system [43]. For the choice of actions, the model enables
simulation of the impact each action has on a particular skill.
To keep the model simple, the action space only consists of
three different task difficulties (i.e., easy, medium, hard).
4.3.3 Results
As an evaluation, the model was implemented and tested
with a robot language tutor during a game-like vocabulary
tutoring interaction with adults (N=40) [61].
We adopted the game ‘I spy with my little eye’. In this
game, the NAO robot describes an object which is displayed
on a tablet along with some distractors, by referring to its
descriptive features in an artificial L2 (i.e., “Vimmi”). The
student then has to guess which object the robot refers to.
The overall interaction structure, consisting of five phases
(i.e., opening, game setup, test-run, game, closing), as well
as the robot’s feedback strategies were based on our observa-
tions of language learning in kindergartens. After the tutoring
interaction, a post-test of the learned words was conducted.
The results revealed that learners’ performance improved
significantly during training with the personalised robot tutor
(Fig. 5). A mixed-design ANOVA with training phase as a
within-subjects factor and training type as between-subject
factor demonstrated a significant main effect of training
phase (F(1,38)=66.85,p<.001
2=.64), such that
learners’ performance was significantly better in the final
phase as compared to the initial phase. Crucially, partici-
pants who learned in the adaptive condition had a higher
number of correct answers as compared to the control con-
dition (F(1,38)=6.52,p=.02
2=.15). Finally, the
3.5
Correct answer count
srewsna7tsaLsrewsna7tsriF
4.0
4.5
5.0
5.5
6.0
6.5
7.0
Adapve
Random
Fig. 5 Mean numbers of correct answers at the beginning (first 7) and
end (last 7) of the interaction in the different conditions. Adapted from
[61]
Table 1 Results of both post-tests (L1-to-L2 and L2-to-L1): Means
(M) and standard deviation (SD) of correct answers grouped by the
experimental conditions
Adaptive (A) Control (C)
MSD MSD
L1-to-L2 3.95 2.56 3.35 1.98
L2-to-L1 7.05 2.56 6.85 2.48
Adapted from [61]
interaction between training phase and type was also signif-
icant (F(1,38)=14.46,p=.001
2=.28), indicating
that the benefit of the adaptive training developed over time.
The results of the post-test did not show significant dif-
ferences between the two conditions, which may be due
to the way in which responses were prompted during the
training sessions and post-test (Table 1). In the training ses-
sions participants saw pictures relating to the meaning of the
to-be-learned words, whereas in the post-test they received
a linguistic cue in form of a word they had to translate.
Although no main effect of training type emerged in the
post-test, some details are nevertheless worth mentioning. In
the L1-to-L2 post-test, a maximum of ten correct responses
was achieved by participants of the adaptive-model condi-
tion, whereas the maximum in the control condition were
six correct answers. Moreover, there were two participants
in the control condition who did not manage to perform
any L1-to-L2 translation correctly, while in the adaptive-
model condition, all participants achieved at least one correct
response (see Fig. 6).
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
336 International Journal of Social Robotics (2018) 10:325–341
Fig. 6 Participant-wise amount of correct answers grouped by the dif-
ferent conditions for the L1-to-L2 post-test. Adapted from [61]
4.3.4 Outlook
This basic adaptive model will be extended by further inte-
grating skill interdependencies as well as affective user states.
Both have already been shown to improve learning [34,63].
In addition, the model can, and is meant to, provide a basis for
exploiting the full potential of an embodied tutoring agent,
and will therefore be advanced to the extent that the robot’s
verbal and non-verbal behaviour will adapt to the learner’s
state of knowledge and progress. Specifically, it aims to
enable dynamic adaptation of (a) embodied behaviour such
as iconic gesture use, which is known to support vocab-
ulary acquisition as a function of individual differences
across children (cf. [59]); (b) the robot’s synthetic voice to
enhance comprehensibility and prosodic focusing of content
when needed; and (c) the robot’s socio-emotional behaviour
depending on the learners’ current level of motivation and
engagement.
5 Evaluation Framework for Robot L2
Tutoring
In this section, we discuss our plans for evaluating our
robot-assisted L2 vocabulary intervention. While this sec-
tion describes future plans rather than already completed
work, it also provides guidelines for evaluating tutoring
systems similar to the L2TOR system. The first step in
an evaluation is the development of pre- and post-tests
designed to assess children’s learning of the targeted vocab-
ulary through comprehension and translation tasks, as well
as tasks assessing deep vocabulary knowledge (i.e., concep-
tual knowledge associated with a word). Not directly targeted
but semantically-related vocabulary will also be assessed, as
well as general vocabulary and other skills related to word
learning (e.g., phonological memory). This is important as
children learn not only the words directly used, but can also
use these words to bootstrap their further vocabulary learning
in the same as well as related domains [51].
In addition to assessing children’s L2 word learning, we
will evaluate the word learning process during the interac-
tive sessions between children and the robot by observing,
transcribing, and coding video-taped interactions. Measures
will include children’s and the robot’s participation and turn-
taking, the type of questions, recasts and expansions, the
semantic contingency of responses and expansions, and the
coherence and length of episodes within the sessions. All
these aspects are known to promote language learning [9,44].
Therefore, it is important to evaluate how these processes are
taking place within the context of language learning with a
social robot.
Finally, given the importance of motivation, we will
observe how children comply with the robot’s initiatives and
instructions, how involved they are in the intervention, and
to what extent they express positive emotions and well-being
during the lessons [41]. The intervention will consist of mul-
tiple sessions, such that children’s learning, motivation, and
interaction with a social robot can be judged over time.
The design of the evaluation study will be based on a com-
parison between an experimental and a control group. The
experimental group will be taught using the social robot while
the control group will receive a placebo training (e.g., non-
language activity with the robot). This design is very common
in educational research as it enables testing whether children
who participate in an educational programme (L2TOR in
this case) learn more or just as much as children who follow
the normal curriculum. Additionally, learning gains with the
robot will be compared to learning gains using an intelligent
tutoring system on a tablet, to test the additional value of
a social robot above existing technology used in education.
In evaluating the robot-supported program developed within
L2TOR, our aim is not only to assess the effectiveness of
the specific tutoring by the L2TOR robot, but also to provide
recommendations for further technological development and
guidelines for future use of social robots in (L2) language
tutoring situations.
6 Conclusion
In this paper, we have presented guidelines for designing
social robots as L2 tutors for preschool children. The guide-
lines cover a range of issues concerning the pedagogy of L2
learning, child–robot interaction strategies, and the adaptive
personalisation of tutoring. Additional guidelines for eval-
uating the effectiveness of L2 tutoring using robots were
presented.
While the benefits of social robots in tutoring are clear,
there are still a range of open issues on how robot tutors
can be effectively deployed in educational settings. The spe-
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 337
cific focus of this research programme –tutoring L2 skills to
young children– requires an understanding of how L2 learn-
ing happens in young children and how children can benefit
from tutoring. Transferring the tutoring to social robots has
highlighted many questions: should the robot simulate what
human tutors do? Should the robot be a peer or a teacher?
How should the robot blend L1 and L2? How should feed-
back be given?
Our aim is to develop an autonomous robot: this incurs
several complex technical challenges, which cannot currently
be met by state-of-the-art AI and social signal processing.
ASR of child speech, for example, is currently insufficiently
robust to allow spoken dialogue between the robot and the
young learner. We propose a number of solutions, including
the use of a tablet as an interaction-mediating device.
Our and our colleagues’ studies show that social robots
hold significant promise as tutoring aids, but a complex pic-
ture emerges as children do not just learn by being exposed
to a tutoring robot. Instead, introducing robots in language
learning will require judicious design decisions on what the
role of the robot is, how the child’s learning is scaffolded,
and how the robot’s interaction can support this.
Acknowledgements We are grateful to all research assistants, partic-
ipating schools, parents, and their children for their assistance in this
project.
Compliance with Ethical Standards
Conflict of interest The authors declare that they have no conflict of
interest.
Funding The L2TOR project is funded by the H2020 Framework Pro-
gramme of the EC, Grant Number: 688014.
Open Access This article is distributed under the terms of the Creative
Commons Attribution 4.0 International License (http://creativecomm
ons.org/licenses/by/4.0/), which permits unrestricted use, distribution,
and reproduction in any medium, provided you give appropriate credit
to the original author(s) and the source, provide a link to the Creative
Commons license, and indicate if changes were made.
References
1. Alemi M, Meghdari A, Ghazisaedy M (2014) Employing
humanoid robots for teaching english language in Iranian Junior
High-Schools. Int J Hum Robot 11(3). https://doi.org/10.1142/
S0219843614500224
2. Alemi M, Meghdari A, Ghazisaedy M (2015) The impact of social
robotics on L2 learners’ anxiety and attitude in English vocabulary
acquisition. Int J Social Robot. pp 1–13. https://doi.org/10.1007/
s12369-015- 0286-y
3. Ate¸s-¸Sen AB, Küntay AC (2015) Childrens sensitivity to caregiver
cues and the role of adult feedback in the development of referential
communication. Acquis Ref 15:241–262
4. Barcelona European Council (2002) Presidency Conclusions, part
I, 43.1
5. Baxter P, Ashurst E, Read R, Kennedy J, Belpaeme T (2017a) Robot
education peers in a situated primary school study: personalisation
promotes child learning. PLoS One 12(5):e0178126
6. Baxter P, de Jong C, Aarts A, de Haas M, Vogt P (2017b) The effect
of age on engagement in preschoolers child–robot interactions. In:
Companion proceedings of the 2017 ACM/IEEE international con-
ference on HRI
7. Belpaeme T, Baxter PE, Read R, Wood R, Cuayáhuitl H, Kiefer
B, Racioppa S, Kruijff-Korbayová I, Athanasopoulos G, Enescu
V et al (2012) Multimodal child–robot interaction: building social
bonds. J Hum–Robot Interact 1(2):33–53
8. Belpaeme T, Kennedy J, Baxter P, Vogt P, Krahmer EJ, Kopp S,
Bergmann K, Leseman P, Küntay AC,Göksun T, Pandey AK, Gelin
R, Koudelkova P, Deblieck T (2015) L2TOR—second language
tutoring using social robots. In: Proceedings of first international
workshop on educational robots
9. Bornstein MH, Tamis-LeMonda CS, Hahn CS, Haynes OM (2008)
Maternal responsiveness to young children at three ages: longitudi-
nal analysis of a multidimensional, modular, and specific parenting
construct. Dev Psychol 44(3):867
10. Brojde CL, Ahmed S, Colunga E (2012) Bilingual and monolingual
children attend to different cues when learning new words. Front
Psychol 3:155. https://doi.org/10.3389/fpsyg.2012.00155
11. Bus AG, Leseman PP, Neuman SB (2012) Methods for preventing
early academic difficulties. In: Harris KR, Graham S, Urdan T, Bus
AG, Major S, Swanson HL (eds) APA Educational Psychology
Handbook, vol 3, APA, pp 227–250
12. Chien NC, Howes C, Burchinal M, Pianta RC, Ritchie S, Bryant
DM, Clifford RM, Early DM, Barbarin OA (2010) Childrens class-
room engagement and school readiness gains in prekindergarten.
Child Dev 81(5):1534–1549
13. Cicchetti DV (1994) Guidelines, criteria, and rules of thumb for
evaluating normed and standardized assessment instruments in psy-
chology. Psychol Assess 6(4):284
14. Clark HH (1996) Using language. Cambridge University Press,
Cambridge
15. Denham SA, Blair KA, DeMulder E, Levitas J, Sawyer K,
Auerbach-Major S, Queenan P (2003) Preschool emotional com-
petence: pathway to social competence? Child Dev 74(1):238–256
16. Dominey PF, Dodane C (2004) Indeterminacy in language acqui-
sition: the role of child directed speech and joint attention. J
Neurolinguist 17(2–3):121–145. https://doi.org/10.1016/S0911-
6044(03)00056-3
17. de Haas M, Vogt P, Krahmer E (2016) Enhancing child–robot tutor-
ing interactions with appropriate feedback. In: Proceedings of the
long-term child–robot interaction workshop at RO-MAN 2016
18. de Haas M, Baxter P, de Jong C, Vogt P, Krahmer E (2017)
Exploring different types of feedback in preschooler- and robot
interaction. In: Companion proceedings of the 2017 ACM/IEEE
international conference on HRI
19. Erber NP (1975) Auditory-visual perception of speech. J Speech
Hear Disord 40(4):481–492. https://doi.org/10.1044/jshd.4004.
481
20. European Commission (2012) Special Eurobarometer 386: Euro-
peans and their languages
21. Fridin M (2014a) Kindergarten social assistive robot: first meeting
and ethical issues. Comput Hum Behav 30:262–272
22. Fridin M (2014b) Storytelling by a kindergarten social assistive
robot: a tool for constructive learning in preschool education. Com-
put Educ 70:53–64
23. Glenberg AM (2008) Embodiment for education. Handbook of
cognitive science: an embodied approach, pp 355–372
24. Gordon G, Breazeal C, Engel S (2015) Can children catch curiosity
from a social robot? In: Proceedings of the 10th ACM/IEEE inter-
national conference on human–robot interaction, ACM, pp 91–98
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
338 International Journal of Social Robotics (2018) 10:325–341
25. Hald LA, de Nooijer J, van Gog T, Bekkering H (2016) Optimizing
word learning via links to perceptual and motoric experience. Educ
Psychol Rev 28(3):495–522. https://doi.org/10.1007/s10648- 015-
9334-2
26. Herberg J, Feller S, Yengin I, Saerbeck M (2015) Robot Watchful-
ness Hinders Learning Performance. In: Proceedings of the 24th
IEEE international symposium on robot and human interactive
communication, pp 153–160
27. Hirsh-Pasek K, Adamson LB, Bakeman R, Owen MT, Golinkoff
RM, Pace A, Yust PKS, Suma K (2015) The contribution of
early communication quality to low-income childrens language
success. Psychol Sci 26(7):1071–1083. https://doi.org/10.1177/
0956797615581493
28. Hockema SA, Smith LB (2009) Learning your language, outside-in
and inside-out. Linguistics 47(2):453–479
29. Hoff E (2013) Interpreting the early language trajectories of chil-
dren from low-SES and language minority homes: implications
for closing achievement gaps. Dev Psychol 49(1):4–14. https://
doi.org/10.1037/a0027238
30. Hoff E, Core C, Place S, Rumiche R, Seor M, Parra M (2012) Dual
language exposure and early bilingual development. J Child Lang
39(1):1–27. https://doi.org/10.1017/S0305000910000759
31. Hoff E, Rumiche R, Burridge A, Ribot KM, Welsh SN (2014)
Expressive vocabulary development in children from bilingual
and monolingual homes: a longitudinal study from two to four
years. Early Child Res Q 29(4):433–444. https://doi.org/10.1016/
j.ecresq.2014.04.012
32. Hurtado N, Marchman VA, Fernald A (2008) Does input influence
uptake? Links between maternal talk, processing speed and vocab-
ulary size in spanish-learning children. Dev Sci 11(6):F31–F39.
https://doi.org/10.1111/j.1467-7687.2008.00768.x
33. Kanda T, Hirano T, Eaton D, Ishiguro H (2004) Interactive robots
as social partners and peer tutors for children: a field trial. J Hum
Comput Interact 19(1):61–84
34. Käser T, Klingler S, Schwing AG, Gross M (2014) Beyond knowl-
edge tracing: modeling skill topologies with bayesian networks.
In: Proceedings of the 12th international conference on intelligent
tutoring systems. Springer, Berlin, pp 188–198
35. Kennedy J, Baxter P, Belpaeme T (2015) The robot who tried too
hard: Social behaviour of a robot tutor can negatively affect child
learning. In: Proceedings of the tenth annual ACM/IEEE interna-
tional conference on human–robot interaction, ACM, pp 67–74
36. Kennedy J, Baxter P, Senft E, Belpaeme T (2016) Social robot tutor-
ing for child second language learning. In: Proceedings of the 11th
ACM/IEEE international conference on human–robot interaction,
ACM, pp 67–74
37. Kennedy J, Lemaignan S, Montassier C, Lavalade P, Irfan B,
Papadopoulos F, Senft E, Belpaeme T (2017) Child speech
recognition in human–robot interaction: evaluations and recom-
mendations. In: Proceedings of the 12th ACM/IEEE international
conference on human–robot interaction, ACM, pp 82–90
38. Konishi H, Kanero J, Freeman MR, Golinkoff RM, Hirsh-Pasek
K (2014) Six principles of language development: implications
for second language learners. Dev Neuropsychol 39(5):404–420.
https://doi.org/10.1080/87565641.2014.931961
39. Kuhl PK (2007) Is speech learning ’gated’ by the social
brain? Dev Sci 10(1):110–120. https://doi.org/10.1111/j.1467-
7687.2007.00572.x
40. Kuhl PK (2010) Brain mechanisms in early language acquisition.
Neuron 67(5):713–727
41. Laevers F et al (2012) Pointing the compass to wellbeing-why and
with what kind of result? Every Child 18(3):26
42. Lee S, Noh H, Lee J, Lee K, Lee GG, Sagong S, Kim M (2011)
On the effectiveness of robot-assisted language learning. ReCALL
23(01):25–58
43. Leitner S (1972) So lernt man lernen. Der Weg zum Erfolg, Herder
44. Leseman PP, Rollenberg L, Rispens J (2001) Playing and work-
ing in kindergarten: cognitive co-construction in two educational
situations. Early Child Res Q 16(3):363–384
45. Leyzberg D, Spaulding S, Toneva M, Scassellati B (2012) The
physical presence of a robot tutor increases cognitive learning
gains. In: Proceedings of the 34th annual conference of the cogni-
tive science society
46. Li J (2015) The benefit of being physically present: a survey of
experimental works comparing copresent robots, telepresent robots
and virtual agents. Int J Hum Comput Stud 77:23–37
47. Marulis LM, Neuman SB (2010) The Effects of vocabu-
lary intervention on young children’s word learning: a meta-
analysis. Rev Educ Res 80(3):300–335. https://doi.org/10.3102/
0034654310377087
48. Mol SE, Bus AG, de Jong MT (2009) Interactive book reading in
early education: a tool to stimulate print knowledge as well as oral
language. Rev Educ Res 79(2):979–1007
49. Moriguchi Y, Kanda T, Ishiguro H, Shimada Y, Itakura S (2011)
Can young children learn words from a robot? Interact Stud
12(1):107–118
50. Movellan JR, Eckhardt M, Virnes M, Rodriguez A (2009) Sociable
robot improves toddler vocabulary skills. In: Proceedings of the 4th
ACM/IEEE international conference on human–robot interaction,
pp 307–308
51. Neuman SB, Newman EH, Dwyer J (2011) Educational effects of
a vocabulary intervention on preschoolers’ word knowledge and
conceptual development: a cluster-randomized trial. Read Res Q
46(3):249–272
52. Newcombe NS, Uttal DH, Sauter M (2013) Spatial development.
Oxford handbook of developmental psychology 1:564–590
53. Oudgenoeg-Paz O, Verhagen J, Vlaar R (2017) Lesson series for
three domains. Tech.rep., Universiteit Utrecht, L2TOR Deliverable
54. Pearson BZ, Fernndez SC, Oller DK (1993) Lexical develop-
ment in bilingual infants and toddlers: comparison to monolingual
norms. Lang Learn 43(1):93–120. https://doi.org/10.1111/j.1467-
1770.1993.tb00174.x
55. Pearson BZ, Fernandez SC, Lewedeg V, Oller DK (1997)
The relation of input factors to lexical learning by bilingual
infants. Appl Psycholinguist 18(1):41–58. https://doi.org/10.1017/
S0142716400009863
56. Place S, Hoff E (2011) Properties of dual language expo-
sure that influence two-year-olds’ bilingual proficiency. Child
Dev 82(6):1834–1849. https://doi.org/10.1111/j.1467- 8624.2011.
01660.x
57. Proctor CP, Carlo M, August D, Snow C (2005) Native Spanish-
speaking children reading in english: toward a model of compre-
hension. J Educ Psychol 97(2):246–256. https://doi.org/10.1037/
0022-0663.97.2.246
58. Rosenthal-von der Pütten AM, Straßmann C, Krämer NC (2016)
Robots or agents–neither helps you more or less during second lan-
guage acquisition. In: Proceedings of the international conference
on intelligent virtual agents. Springer, Berlin, pp 256–268
59. Rowe ML, Silverman RD, Mullan BE (2013) The role of pictures
and gestures as nonverbal aids in preschoolers’ word learning in a
novel language. Contemp Educ Psychol 38(2):109–117
60. Saerbeck M, Schut T, Bartneck C, Janse MD (2010) Expressive
robots in education: varying the degree of social supportive behav-
ior of a robotic tutor. In: Proceedings of the SIGCHI conference on
human factors in computing systems, ACM, pp 1613–1622. https://
doi.org/10.1145/1753326.1753567
61. Schodde T, Bergmann K, Kopp S (2017) Adaptive robot lan-
guage tutoring based on Bayesian knowledge tracing and predictive
decision-making. In: Proceedings of the 12th ACM/IEEE interna-
tional conference on human–robot interaction, ACM
62. Sidner CL, Lee C, Kidd CD, Lesh N, Rich C (2005) Explorations in
engagement for humans and robots. Artif Intell 166(1–2):140–164
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 339
63. Spaulding S, Gordon G, Breazeal C (2016) Affect-aware student
models for robot tutors. In: Proceedings of the 2016 international
conference on autonomous agents and multiagent systems, pp 864–
872
64. Tanaka F, Matsuzoe S (2012) Children teach a care-receiving robot
to promote their learning: field experiments in a classroom for
vocabulary learning. J Hum Robot Interact 1(1):78–95
65. Tellier M (2008) The effect of gestures on second language mem-
orisation by young children. Gestures Lang Dev 8(2):219–235.
https://doi.org/10.1075/gest.8.2.06tel
66. Tomasselo M, Todd J (1983) Joint attention and lexical acquisition
style. First Lang 4:197–212
67. Unsworth S (2014) Comparing the role of input in bilingual acqui-
sition across domains. In: Grter T, Paradis J (eds) Trends language
acquisition research, vol 13, John Benjamins Publishing Company,
pp 181–201. https://doi.org/10.1075/tilar.13.10uns
68. Vlaar R, Verhagen J, Oudgenoeg-Paz O, Leseman P (2017) Com-
paring L2 word learning through a tablet or real objects: what
benefits learning most? In: Proceedings of the R4L workshop at
HRI’17
69. Vogt P, de Haas M, de Jong C, Baxter P, Krahmer E (2017)
Child-robot interactions for second language tutoring to preschool
children. Frontiers in Human Neuroscience 11: https://doi.org/10.
3389/fnhum.2017.00073
70. Vygotsky L (1978) Mind in society. Harvard University Press,
Cambridge
71. Westlund JK, Dickens L, Jeong S, Harris P, DeSteno D, Breazeal C
(2015) A comparison of children learning new words from robots,
tablets, and people. In: Proceedings of the 1st international confer-
ence on social robots in therapy and education
72. Westlund KJM, Martinez M, Archie M, Das M, Breazeal C (2016a)
Effects of framing a robot as a social agent or as a machine
on children’s social behavior. In: Proceedings of the 25th IEEE
international symposium on robot and human interactive commu-
nication, IEEE, pp 688–693
73. WestlundJMK, Martinez M, Archie M, Das M, Breazeal C (2016b)
A study to measure the effect of framing a robot as a social agent
or as a machine on children’s social behavior. In: The eleventh
ACM/IEEE international conference on human robot interaction,
IEEE Press, pp 459–460
74. Wermelinger S, Gampe A, Daum MM (2017) Bilingual toddlers
have advanced abilities to repair communication failure. J Exp
Child Psychol 155:84–94. https://doi.org/10.1016/j.jecp.2016.11.
005
75. White J (1996) An input enhancement study with ESL children:
effects on the acquisition of possessive determiners. Ph.D. disser-
tation, McGill University, Montreal, Canada
76. Yow WQ, Markman EM (2011) Young bilingual children’s height-
ened sensitivity to referential cues. J Cognit Dev 12(1):12–31.
https://doi.org/10.1080/15248372.2011.539524
Tony Belpaeme is a Professor at the Centre for Robotics and Neural
Systems at the University of Plymouth (UK) and at Ghent University
(Belgium). He received his Ph.D. in Computer Science from the Vrije
Universiteit Brussel (Belgium). He leads a team studying cognitive
robotics and human–robot interaction. Starting from the premise that
intelligence is rooted in social interaction, Belpaeme and his research
team try to further the science and technology behind artificial intel-
ligence and social robots. This results in a spectrum of results, from
theoretical insights to practical applications. He is the coordinator of
the H2020 L2TOR project, a large-scale European project bringing 7
partners together to study how robots can be used to support the learn-
ing of a second language by children.
Paul Vogt is an Associate Professor at the Department of Cogni-
tive Science and Artificial Intelligence at Tilburg University in the
Netherlands. He received an M.Sc. in Cognitive Science and Engineer-
ing from the University of Groningen (Netherlands), and obtained a
Ph.D. at the Artificial Intelligence Laboratory of the Vrije Universiteit
Brussel (Belgium). His research focuses on understanding the cultural,
social and cognitive mechanisms that underlie the evolution and acqui-
sition of language and communication. Vogt is particularly interested
in investigating how humans and machines can ground the meaning
of linguistic utterances in the real world, and how they learn language
from each other through social interactions. To study this, he has used
a variety of techniques, ranging from agent-based modelling, child–
robot interaction and psycholinguistic experiments to ethnographic
research of children’s language acquisition in different cultures.
Rianne van den Berghe is a Ph.D. candidate at the Department
of Special Education at Utrecht University in the Netherlands. She
received an MA in Linguistics from Utrecht University, in which she
focused on (second) language acquisition and discourse processing. In
her Ph.D. research, she investigates the way robot-assisted language
lessons should be designed in order to optimize children’s learning
gains and engagement.
Kirsten Bergmann is a postdoctoral researcher at the Cluster of Excel-
lence for Cognitive Interaction Technology at Bielefeld University in
Germany. She received her Ph.D. in Computer Science from Biele-
feld University. For the past ten years she has worked on empirically
grounded and cognitively plausible models of multimodal commu-
nicative behaviour, with a particular focus on coordination of speech
and gestures in artificial agents. In current projects she is developing
embodied tutoring systems, and has been investigating the role of mul-
timodal communication in educational settings.
Tilbe Göksun is an Assistant Professor of Psychology at the Depart-
ment of Psychology at Koç University in Istanbul, Turkey. She received
her Ph.D. in Developmental Psychology from Temple University in
Philadelphia, PA, USA and worked as a postdoctoral researcher in
the Center for Cognitive Neuroscience at the University of Pennsylva-
nia in Philadelphia, PA, USA. She leads the Language and Cognition
Lab at Koç University. Her research examines the interaction between
language and thought processes, focusing on first and second lan-
guage acquisition, event perception, relational and spatial language,
neuropsychology of language, and the role of gestures in these pro-
cesses.
Mirjam de Haas is a Ph.D. candidate at the Tilburg center for Cogni-
tion and Communication at Tilburg University in the Netherlands. She
received her M.Sc. in Artificial Intelligence from the Radboud Uni-
versity Nijmegen, the Netherlands. Her Ph.D. focuses on the design
of child robot tutor interactions and she is interested in how to keep
the children motivated and engaged throughout the different lessons.
Junko Kanero is a Postdoctoral Researcher in the Department of Psy-
chology at Koç University in Istanbul, Turkey. She received her Ph.D.
in Developmental Psychology and Neuroscience from Temple Univer-
sity in Philadelphia, PA, USA. Her research examines language as a
window into human cognition, using a broad range of methodologies
including behavioural measures, eye tracking, EEG, and fMRI. She is
interested in language development in infancy and childhood, neural
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
340 International Journal of Social Robotics (2018) 10:325–341
processing of language, and how language and non-linguistic cogni-
tion interact.
James Kennedy is a Postdoctoral Associate at Disney Research, Los
Angeles, USA. James received his Ph.D. from Plymouth University,
UK in 2017 for his work using social robots to tutor children. Dur-
ing his Ph.D., he worked as a Research Assistant on the EU-funded
DREAM project and collaborated with the ALIZ-E, and L2TOR
projects, focusing on the use of social robots in applications involving
children. His research interests lie in Human–Robot Interaction and
Socially Intelligent Agents.
Aylin C. Küntay is a professor of psychology and the Dean of College
of Social Sciences and Humanities in Koç University. She received her
Ph.D. in Developmental Psychology in 1997 from the University of
California at Berkeley. Her work is on the relation of early commu-
nicative and language development to social interaction and cognitive
development in bilingual and monolingual children.
Ora Oudgenoeg-Paz is a postdoctoral researcher at the Department of
Special Education: Cognitive and Motor Disabilities at Utrecht Uni-
versity (the Netherlands). She received her Ph.D. in Pedagogics from
Utrecht University. Her research focuses on early language and motor
development and the link between the two. She studies how motor
development, sensorimotor interactions and early language exposure
facilitate the development of (spatial) cognition and (spatial) language.
Her work concerns both first and second language acquisition.
Fotios Papadopoulos is a Reseach Fellow at Centre for Robotics and
Neural Systems at the University of Plymouth (UK). He received his
Ph.D. in 2012 from the University of Hertfordshire. His research inter-
ests lies within Human–Robot interaction, robot engagement, haptic
communication, and robot tele-operation.
Thorsten Schodde is a Ph.D. candidate at the Cluster of Excellence
for Cognitive Interaction Technology at Bielefeld University in Ger-
many. He also received a Master of Science in Intelligent Systems
from Bielefeld University. His Ph.D. research focuses on planning of
an adaptive teaching course for second language learning lessons for
preschool children. The major aim is the maximization of the learning
gain while engagement and motivation are kept high.
Josje Verhagen is a post-doctoral researcher at the Department of
Special Education: Cognitive and Motor Disabilities at Utrecht Uni-
versity, the Netherlands. She received her Ph.D. in 2009 on adult sec-
ond language acquisition from the Max Planck in Psycholinguistics
(Nijmegen) and Free University (Amsterdam). Her research interests
are first and second language acquisition, and bilingualism. In her cur-
rent research, she studies how specific properties of the language input
affect acquisition, and how language development in young children
relates to more general cognitive development.
Christopher D. Wallbridge is a Ph.D. candidate at the Centre for
Robotics and Neural Systems at the University of Plymouth (UK). He
received a Master of Science in Robotics from the University of Ply-
mouth. His research is on the natural use of spatial concepts by robots,
including their use in multiple languages.
Bram Willemsen is a Researcher at the Tilburg Center for Cogni-
tion and Communication at Tilburg University, the Netherlands. He
received his M.Sc. in Communication and Information Sciences (cum
laude) from Tilburg University, the Netherlands. His research inter-
ests concern problems related to Natural Language Understanding and
data-driven approaches to dialogue modelling. His work within the
L2TOR project focuses on the realization of context-aware generation
of verbal and non-verbal robot behaviours.
Jan de Wit is a Ph.D. candidate at the Tilburg center for Cognition and
Communication at Tilburg University in the Netherlands. He received
his M.Sc. in Game and Media Technology from Utrecht University
and his PDEng in User System Interaction from Eindhoven University
of Technology, the Netherlands. He is on a quest to design technology
that contributes to society in a fun and light-hearted way. In his Ph.D.
research, he is exploring the role of robot-performed gestures in chil-
dren’s second language learning.
Vasfiye Geçkin is a lecturer at School of Foreign Languages, Bogazici
University, Istanbul, Turkey. She received her Ph.D. (a cotutelle degree)
in Linguistics from Macquarie University (Australia) and the Univer-
sity of Potsdam (Germany) in 2015, on acquisition of logical operators
by monolingual and bilingual children. Her research focuses on lan-
guage comprehension and production of bilingual and monolingual
children.
Laura Hoffmann is Postdoctoral Researcher at the Cluster of Excel-
lence Cognitive Interaction Technology (CITEC) associated with
Bielefeld University, Germany. Her research is about the psycholog-
ical impact of interactive technologies, in particular social robots. She
uses quantitative and qualitative methods to understand how humans
make sense of artificial others; how they perceive and evaluate them
according to specific characteristics like embodiment, morphology or
behavior.
Stefan Kopp is a Professor of Computer Science and head of the
Social Cognitive Systems Group at Bielefeld University. He received
his Ph.D. in 2004 for work on the synthesis of multimodal commu-
nicative behaviour of embodied agents. After a Postdoc at Northwest-
ern University, he was research fellow at Bielefeld’s Center for Inter-
disciplinary Research (ZiF) and is now principal investigator of the
Center of Excellence “Cognitive Interaction Technology” (CITEC).
Kopp and his team develop computational accounts of behavioural
and cognitive abilities needed to act as a socially intelligent interac-
tion partner. These are embedded in artificial systems like virtual 3D
avatars or social robots, which are applied and evaluated in assisted
living, industrial, or educational settings.
Emiel Krahmer is a Professor of Language, Cognition and Computa-
tion at the Tilburg School of Humanities and Digital Sciences, where
he co-leads the Language, Communication and Cognition research
group. He received his Ph.D. in Computational Linguistics in 1995,
after which he worked as a postdoc in the Institute for Perception
Research at the Eindhoven University of Technology before mov-
ing to Tilburg University. In his current work he studies how peo-
ple communicate with each other, both in text and in speech, with
the aim of subsequently improving the way computers communicate
with human users. To achieve this, he combines computational mod-
elling and experimental studies with human participants. Much of his
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
International Journal of Social Robotics (2018) 10:325–341 341
research is funded through external grants, including an NWO VICI
grant.
Ezgi Mamus is a Ph.D. candidate at Radboud University in the Nether-
lands. She received her MA in Cognitive Psychology from Bogazici
University in Istanbul, Turkey. Her Ph.D. research focuses on the influ-
ence of perceptual modality (e.g., sound vs. vision) on gestural repre-
sentations of spatial events.
Jean-Marc Montanier is an engineer researcher at SoftBank Robotics
Europe. He obtained his Ph.D. from Université Paris-Sud XI in 2013,
on the study of autonomous auto-adaptation in swarm robotics. Since
then he worked on topics relative to autonomous learning in multi-
agent Systems. At SoftBank Robotics Europe, he is looking at the
latests trends in Artificial Intelligence in order to transfer them to
industrialized robots.
Cansu Oranç is a Ph.D. candidate at the Department of Psychology at
Koç University in Istanbul, Turkey. She received her M.Sc. degree in
Behavioural Science from Radboud University Nijmegen, the Nether-
lands. Her research focuses on the factors affecting children’s learning
of new information from different technological sources such as elec-
tronic books and robots.
Amit Kumar Pandey is Head Principal Scientist (Chief Scientist) at
SoftBank Robotics Europe, Paris, France, also serving as the scientific
coordinator and principal investigator of the European Union (EU)
collaborative projects of the company. Earlier for 6 years he worked
as researcher in Robotics and AI at LAAS-CNRS (French National
Center for Scientific Research), France. Among other responsibilities,
he is also the founding coordinator of Socially Intelligent Robots and
Societal Applications (SIRo-SA) Topic Group (TG) of euRobotics. He
is also the recipient of the second best Ph.D. thesis (tie) in Robotics in
Europe for the prestigious euRobotics Georges Giralt Award.
123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com
... Research has shown that social robots can have many benefits in the education field (Alves-Oliveira et al., 2016;Belpaeme, Vogt, et al., 2018;Donnermann et al., 2022;Gleason & Greenhow, 2017;Ramachandran et al., 2016;Rosenberg-Kima et al., 2020;Smakman et al., 2020;Vincent et al., 2015). For example, social robot tutors can be beneficial, as suggested by research in which social robots were used to assist children in learning a second language (e.g., Vogt et al., 2019) or solving fraction problems (Ramachandran et al., 2016). ...
... They found that students interacting with a robot with a personalized and more human-like behavior scored higher on the exam and had an increase in intrinsic motivation related to the course content in general compared to students who interacted with a robot that did not adapt itself to the participants. Other research came to a similar conclusion that designing the educational robots as more anthropomorphic, or at least fitting their demographic (i.e., children), results in better learning rates and positive social interactions (Belpaeme, Vogt, et al., 2018;Vincent et al., 2015). Furthermore, some papers have shown results that an anthropomorphic social robot encourages responses that are beneficial for learning because it invites social interaction with the robot . ...
Conference Paper
Full-text available
Social robots are becoming increasingly relevant in education, for example, by using them as tutors. To create a more empathetic and engaging learning environment, it is important to consider the anthropomorphism of these social robots. However, an ethnic perspective on the use of anthropomorphization is still lacking when it comes to improving learning gains. Therefore, this research focuses on whether personalized, ethnicity-based anthropomorphization of a robot can enhance learning gains. To this end, history lessons were tutored with a Furhat robot, with groups of participants interacting with a Furhat whose face matched the ethnicity of the participants, in an experimental setting. Our results showed that participants who interacted with the robot displaying the personalized, ethnicity-based anthropomorphization learned more than participants interacting with a robot displaying a robotic appearance. These findings highlight the importance of incorporating cultural diversity into educational technologies to foster more effective and inclusive learning environments.
... Specifically, the robot acts as a tutor who role-plays a predetermined scenario, and the scenario is proceeded by the learner's input to a tablet. Such a system integrating a robot and a tablet is recommended in the guidelines for designing social robots as second-language tutors [7]. Despite such limited lessons, it is meaningful to compare the learning outcomes when using RALL systems and human tutors. ...
... Further, it is difficult to accurately recognize speech in children [24]; moreover, it is difficult to accurately recognize speech in a second language, even in adults [11]. The guidelines for designing social robots as second-language tutors recommend that using a tablet makes the design context more flexible and reduces the need to rely on complex object recognition and tracking [7]. Therefore, in this study, we combined a tablet and a robot. ...
Article
Full-text available
This study explores how much current mainstream Robot-Assisted Language Learning (RALL) systems produce outcomes compared to human tutors instructing a typical English conversation lesson. To this end, an experiment was conducted with 26 participants divided in RALL (14 participants) and human tutor (12 participants) groups. All participants took a pre-test on the first day, followed by 30 min of study per day for 7 days, and 3 post-tests on the last day. The test results indicated that the RALL group considerably improved lexical/grammatical error rates and fluency of speech compared to that for the human tutor group. The other characteristics, such as rhythm, pronunciation, complexity, and task achievement of speech did not indicate any differences between the groups. The results suggested that exercises with the RALL system enabled participants to commit the learned expressions to memory, whereas those with human tutors emphasized on communication with the participants. This study demonstrated the benefits of using RALL systems that can work well in lessons that human tutors find hard to teach.
... Kennedy et al [8] in their study with primary school children in a language learning setting cast the robot in the role of a tutor. Belpaeme et al [17] designing social robots as second language learning tutors. Davison [18] tested how primary school children's selfregulated learning processes developed over a four-monthlong study during which a social robot acted as a tutor and instructor to various learning tasks. ...
Conference Paper
Full-text available
Research has shown the potential of social robots to support learning in science, technology, and language. We contribute to this field by exploring how robots can support music learning. We report on a within-subjects experiment where 50 young learners practiced the piano in the presence of a robot assuming a non-evaluative and a self-assessment enhancing role implemented in a Wizard-of-Oz fashion. We examined whether the robot can make piano practice more fun, and whether initiating self-assessment to support self-regulated learning is a useful strategy for the robot. We collected quantitative self-report data to assess fun, learning, interest, engagement, and effort. We found a direct positive effect of fun on learning in the context of musical instrument practice. Path modeling showed a positive influence of having fun on learners' attitudes, interests, and learning outcomes in music education, particularly with the self-assessment robot role exhibiting superiority.
... For human-human explanation, it has been shown that it helps the listener to understand the intended meaning and structure if the speaker uses gestures in an interaction [10,42]. It has also been shown that iconic gestures can support the long-term learning of second language vocabulary in children [4,7,23,81,82]. ...
Preprint
Full-text available
In human interaction, gestures serve various functions such as marking speech rhythm, highlighting key elements, and supplementing information. These gestures are also observed in explanatory contexts. However, the impact of gestures on explanations provided by virtual agents remains underexplored. A user study was carried out to investigate how different types of gestures influence perceived interaction quality and listener understanding. This study addresses the effect of gestures in explanation by developing an embodied virtual explainer integrating both beat gestures and iconic gestures to enhance its automatically generated verbal explanations. Our model combines beat gestures generated by a learned speech-driven synthesis module with manually captured iconic gestures, supporting the agent's verbal expressions about the board game Quarto! as an explanation scenario. Findings indicate that neither the use of iconic gestures alone nor their combination with beat gestures outperforms the baseline or beat-only conditions in terms of understanding. Nonetheless, compared to prior research, the embodied agent significantly enhances understanding.
... Согласно психолого-педагогическому представлению о цифровизации, цифровые устройства создаются и используются для человека, а не вместо него. Нейроцифровые устройства и роботы (смарт-устройства) в области образования чаще всего создаются, применяются, исследуются и совершенствуются в качестве сверстников или компаньонов в обучении со студентами [25], в роли обучающих студентов репетиторов [7], в качестве педагогов, использующих фронтальный режим лекции [31], во взаимодействии «один на один» и в образовательных диалогах [12]. Они используются даже для активизации формирования и использования метакогнитивных навыков, включая любопытство и стремление понять себя и мир [12], а также так называемое «мышление роста» (growth mindset) и критическое мышление [27]. ...
Article
The aim of the study is to analyse the psychological and pedagogical problems of using neurodigital and smart technologies in modern vocational education as leading trends in its digitalization. Neurodigital technologies and robots should be developed, implemented, and improved into modern and future vocational education, as truly high-tech, complex products that help a person in his development, set and solve “super tasks”, and improve the subjects of education, educational relations.
Chapter
In the realm of Education 5.0, individuals investing in English language learner (ELL) educator training recognize the imperative to equip students with an understanding of the challenges and possibilities emerging from the integration of technology into the English language classroom. Offering a comprehensive overview, this chapter delineates the utilization of an educational robotics approach in EFL education and contemplates potential outcomes for students studying English education.
Article
Full-text available
The benefit of social robots to support child learning in an educational context over an extended period of time is evaluated. Specifically, the effect of personalisation and adaptation of robot social behaviour is assessed. Two autonomous robots were embedded within two matched classrooms of a primary school for a continuous two week period without experimenter supervision to act as learning companions for the children for familiar and novel subjects. Results suggest that while children in both personalised and non-personalised conditions learned, there was increased child learning of a novel subject exhibited when interacting with a robot that personalised its behaviours, with indications that this benefit extended to other class-based performance. Additional evidence was obtained suggesting that there is increased acceptance of the personalised robot peer over a non-personalised version. These results provide the first evidence in support of peer-robot behavioural personalisation having a positive influence on learning when embedded in a learning environment for an extended period of time.
Article
Full-text available
In this digital age social robots will increasingly be used for educational purposes, such as second language tutoring. In this perspective article, we propose a number of design features to develop a child-friendly social robot that can effectively support children in second language learning, and we discuss some technical challenges for developing these. The features we propose include choices to develop the robot such that it can act as a peer to motivate the child during second language learning and build trust at the same time, while still being more knowledgeable than the child and scaffolding that knowledge in adult-like manner. We also believe that the first impressions children have about robots are crucial for them to build trust and common ground, which would support child-robot interactions in the long term. We therefore propose a strategy to introduce the robot in a safe way to toddlers. Other features relate to the ability to adapt to individual children’s language proficiency, respond contingently, both temporally and semantically, establish joint attention, use meaningful gestures, provide effective feedback and monitor children’s learning progress. Technical challenges we observe include automatic speech recognition (ASR) for children, reliable object recognition to facilitate semantic contingency and establishing joint attention, and developing human-like gestures with a robot that does not have the same morphology humans have. We briefly discuss an experiment in which we investigate how children respond to different forms of feedback the robot can give.
Conference Paper
Full-text available
An increasing number of human-robot interaction (HRI) studies are now taking place in applied settings with children. These interactions often hinge on verbal interaction to effectively achieve their goals. Great advances have been made in adult speech recognition and it is often assumed that these advances will carry over to the HRI domain and to interactions with children. In this paper, we evaluate a number of automatic speech recognition (ASR) engines under a variety of conditions, inspired by real-world social HRI conditions. Using the data collected we demonstrate that there is still much work to be done in ASR for child speech, with interactions relying solely on this modality still out of reach. However, we also make recommendations for child-robot interaction design in order to maximise the capability that does currently exist.
Article
The purpose of this study was to examine the hypothesis that helping preschoolers learn words through categorization may enhance their ability to retain words and their conceptual properties, acting as a bootstrap for self‐learning. We examined this hypothesis by investigating the effects of the World of Words instructional program, a supplemental intervention for children in preschool designed to teach word knowledge and conceptual development through taxonomic categorization and embedded multimedia. Participants in the study included 3‐ and 4‐year‐old children from 28 Head Start classrooms in 12 schools, randomly assigned to treatment and control groups. Children were assessed on word knowledge, expressive language, conceptual knowledge, and categories and properties of concepts in a yearlong intervention. Results indicated that children receiving the WOW treatment consistently outperformed their control counterparts; further, treatment children were able to use categories to identify the meaning of novel words. Gains in word and categorical knowledge were sustained six months later for those children who remained in Head Start. These results suggest that a program targeted to learning words within taxonomic categories may act as a bootstrap for self‐learning and inference generation. كان الغرض من هذه الدراسة هو فحص فرضية أن مساعدة الأطفال في الروضة تعلم كلمات عن طريق التصنيف بإمكانه أن يعزز من قدراتهم على تذكر الكلمات وخصائصها المفاهيمية، بوصفها محفزا للتعلم. قمنا بفحص هذه الفرضية وذلك بالتحقيق في أثر البرنامج التعليمي "عالم المفردات" (WOW) الذي يعتبر تدخلا تكميليا لأطفال الروضة ومصمما من أجل تعليم معرفة الكلمات والتطور المفاهيمي، من خلال النظام التصنيفي ووسائل الإعلام. تضمنت الدراسة مشاركين تتراوح أعمارهم بين 3 و4 سنوات ينتمون إلى 28 روضة أطفال، التي تنتمي إلى 12 مدرسة, خضعت عشوائيا للبحث ومراقبة المجموعات. جرى تقييم الأطفال على معرفة الكلمات واللغة التعبيرية والمعرفة المفاهيمية وأصناف وخصائص المفاهيم في تدخل لمدة سنة كاملة. أظهرت النتائج أن الأطفال الذين خضعوا للبرنامج التعليمي "عالم المفردات" (WOW) قد تفوقوا بصورة مستمرة على نظرائهم الذين خضعوا للمراقبة. علاوة على ذلك, فإن أطفال البحث قد تمكنوا من استخدام الأصناف والتعرف على معنى الكلمات الجديدة. استمر اكتساب الكلمات والمعرفة التصنيفية بعد 6 أشهر بالنسبة للأطفال الذين بقوا في روضة الأطفال. توحي هذه النتائج إلى أنه يمكن لبرنامج يهدف إلى تعلم الكلمات أن يعمل كمحفز للتعلم الذاتي، وتوليد الاستنتاجات. 本研究旨在考查一个假设:帮助学龄前儿童通过分类来学习单词,可提高他们记忆单词及其概念属性的能力,从而发展他们的自主学习能力。作者通过调查「单词世界」(WOW)教学计划的影响来考查这个假设。该教学计划是一个学龄前儿童补充干预计划,旨在透过使用分类学的分类方法及嵌入式多媒体,教授单词知识和发展单词概念。研究参与者是来自12所学校中的28个「启蒙计划」学前儿童班里的3‐4岁儿童,他们被随机分配到干预组和对照组。在一年的干预中,儿童接受单词知识、表达语言、概念知识、类别和概念属性的评估。结果显示,「单词世界」(WOW)干预组的表现一致地优于对照组;此外,干预组儿童可以使用类别来确定新词的词义。仍然留在「启蒙计划」班里的儿童,其单词和类别知识的改进保持稳定至干预后6个月。这些研究结果显示,针对透过分类学的分类方法来学习单词的教学计划,可引导儿童凭自己的力量学习自学和产生推断。 Cette étude avait pour but d'examiner l'hypothíse qu'aider des enfants d'âge préscolaire à apprendre des mots en les catégorisant pourrait favoriser leur capacité à retenir les mots et leurs propriétés, agissant alors comme une amorce d'auto‐apprentissage. Nous avons examiné cette hypothíse en analysant les effets du matériel d'enseignement le Monde des Mots (MDM), un supplément pédagogique destiné aux enfants d'âge préscolaire conçu pour la connaissance des mots et le développement des concepts à l'aide d'une catégorisation taxinomique avec multimédia impliqué. Ont participé à l'étude des enfants de 3 et 4 ans provenant de 28 classes Head Start issus de 12 écoles assignées au hasard au groupe de traitement et au groupe contrôle. Les enfants ont été évalués sur leur connaissance des mots, l'expression orale, les connaissances conceptuelles, les catégories et les propriétés des concepts tout au long de l'année de l'intervention. Les résultats ont montré que les enfants du groupe de traitement MDM ont de maniíre systématique dépassé leur contrepartie du groupe contrôle; en outre, les enfants soumis au traitement ont été en mesure de se servir des catégories pour trouver le sens des mots nouveaux. Les bénéfices dans la connaissance des mots et les connaissances catégorielles sont demeurés six mois plus tard chez les enfants restés dans Head Start. Ces résultats suggírent qu'un programme visant l'apprentissage de mots au sein de catégories taxonomiques peut agir comme une amorce pour l'auto‐apprentissage et la production d'inférences. Проверялась гипотеза о том, что категоризация лексики при обучении дошкольников новым словам может существенно увеличить способность к запоминанию слов и их концептуальных свойств и стимулировать малышей к самообучению. Для расширения их словарного запаса и умения работать с концептами авторы исследовали учебный модуль “Мир слов” (WOW), разработанный в качестве дополнительного вмешательства для дошкольников, которые обучаются по программе Head Start. WOW знакомит детей с таксономической классификацией посредством мультимедийных средств. Трех‐ и четырехлетние дети из 28‐и дошкольных групп в 12‐и школах были случайным образом включены либо в экспериментальную, либо в контрольную группу. На протяжении годичного обучения оценивалось знание слов, выразительность речи, знание концептов, их свойств и категорий, к которым они могут быть причислены. Дети, обучавшиеся по программе WOW, стабильно показывали более высокие результаты, чем их ровесники из контрольных групп. Помимо прочего, эти дети способны использовать категоризацию для определения значений новых слов. Через полгода после окончания обучения эти дети продолжали опережать сверстников по знанию слов и умению категоризировать. Это свидетельствует о том, что программа, предлагающая изучение слов в рамках таксономических категорий, может помочь вырастить поколение, которое будет способно к самообучению и к самостоятельным выводам. La meta de este estudio fue el de investigar la hipótesis que ayudar a los preescolares a aprender palabras por medio de la categorización mejoraría su capacidad de retener palabras y sus propiedades conceptuales, sirviendo de arranque para el auto aprendizaje. Investigamos esta hipótesis estudiando los efectos del programa de enseñanza World of Words (Mundo de palabras; WOW por sus siglas en inglés), una intervención adicional para niños preescolares diseñada para el aprendizaje de palabras y el desarrollo conceptual por medio de la categorización taxonómica y el uso de diversos medios. En este estudio participaron niños de 3 y 4 años de 20 aulas de Head Start en 12 escuelas escogidas al azar en cuanto a grupos de tratamiento y de control. En un año completo de intervención, se evaluaron los estudiantes en cuanto a su conocimiento de palabras, su lenguaje expresivo, su conocimiento conceptual, y las categorías y propiedades de los conceptos. Los resultados mostraron que los niños del grupo de WOW sistemáticamente superaban a los niños del grupo de control; además, los niños del grupo de tratamiento podían usar categorías para encontrar el significado de palabras nuevas. Los adelantos en el conocimiento de palabras y categorías todavía existían 6 meses más tarde para los niños que seguían con Head Start. Estos resultados sugieren que un programa dedicado al aprendizaje de palabras dentro de categorías taxonómicas puede ayudar al autoaprendizaje y la producción de inferencias.
Article
Copyright © 2016, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. Computational tutoring systems, such as educational software or interactive robots, have the potential for great societal benefit. Such systems track and assess students' knowledge via inferential methods, such as the popular Bayesian Knowledge Tracing (BKT) algorithm. However, these methods do not typically draw on the affective signals that human teachers use to assess knowledge, such as indications of discomfort, engagement, or frustration. In this paper we present a novel extension to the BKT model that uses affective data, derived autonomously from video records of children playing an interactive story-telling game with a robot, to infer student knowledge of reading skills. We find that, compared to a control group of children who played the game with only a tablet, children who interacted with an embodied social robot generated stronger affective data signals of engagement and enjoyment during the interaction. We then show that incorporating this affective data into model training improves the quality of the learned knowledge inference models. These results suggest that physically embodied, affect-aware robot tutors can provide more effective and empathic educational experiences for children, and advance both algorithmic and human-centered motivations for further development of systems that tightly integrate affect understanding and complex models of inference with interactive, educational robots.
Conference Paper
This study considered the feedback of a robot during second language tutoring. Traditionally, robots are programmed to provide feedback as teacher; we propose a robot that acts as a peer to motivate preschoolers during the tutoring. We conducted an experiment with 65 preschoolers (M = 3.6 years) in which the robot varied feedback in three conditions: peer-like (explicit negative), adult-like (explicit positive and implicit negative) and no feedback. The results suggest that feedback did not influence children's engagement (measured via eye-gaze), although children who received peer-like feedback seemed to perform more independently during the learning task (requiring less interventions from the experimenter).
Conference Paper
In this study, we investigate the effect of age on preschoolers' engagement - as measured by gaze direction - during a first-time interaction with a social robot. The results revealed significant differences in gaze patterns. Specifically, younger children were more easily distracted, and looked at the robot for a shorter duration and briefer periods of gaze. Moreover, they showed a higher level of reliance on the experimenters. The results have implications for the design of young preschoolers child-robot interactions and specifically for the ways in which the first introductory interactions should occur.
Conference Paper
In this paper, we present an approach to adaptive language tutoring in child-robot interaction. The approach is based on a dynamic probabilistic model that represents the inter-relations between the learner's skills, her observed behaviour in tutoring interaction, and the tutoring action taken by the system. Being implemented in a robot language tutor, the model enables the robot tutor to trace the learner's knowledge and to decide which skill to teach next and how to address it in a game-like tutoring interaction. Results of an evaluation study are discussed demonstrating how participants in the adaptive tutoring condition successfully learned foreign language words.
Article
Recent research has demonstrated enhanced communicative abilities in bilingual children compared with monolingual children throughout childhood and in a variety of domains. The processes underlying these advantages are, however, not well understood. It has been suggested that one aspect that particularly stimulates bilinguals’ communication skills is their daily experience with challenging communication. In the current study, we investigated whether children’s assumed experience with communication failures would increase their skills when it came to repairing communication failure. Non-German bilingual, German bilingual, and monolingual 2.5-year-old toddlers participated in a communication task in which a misunderstanding occurred. We hypothesized that monolingual and German bilingual children would have fewer daily communication failures—and, therefore, less well-trained repair skills—compared with non-German bilinguals. The results showed that non-German bilinguals were more likely to repair the misunderstanding compared with both monolingual children and German bilingual children. The current findings support the view that the communicative advantages of bilingual individuals develop based on their unique experience with interpersonal communication and its difficulties.