2.2 Robots creating linguistic systems
Others have applied the principle of grounding word
semantics in perceptual inputs of robots to acqui-
sition and evolution of artificial languages, e.g. in
(Steels, 1999) and (Steels and Kaplan, 2001). Those
languages are used by populations of software agents
which teleport themselves into active vision heads so
they may perceive the scenes they describe to each
other. These experiments are demonstrations of a
computational model of language evolution. The
resulting languages are not natural; they are con-
structed specifically and exclusively for the purpose
of playing the naming game and not intended for
interaction with humans. Therefore, a human in-
teracting with an agent from the experiment must
learn the new language in order to engage in com-
munication, which will be restricted by the rules of
the game.
Also interesting is the work of Holly Yanco
(Yanco, 1994), where a team of robots must collec-
tively perform an atomic task by following a desig-
nated leader (human or robot). Based on certain as-
sumptions about the task and the environment, e.g.,
perfect communication, no inference of goals, and the
simplicity of tasks, the robots create an artificial lan-
guage to assist their collaborative effort. Again, the
development of this language reflects its very spe-
cific purpose, which is not necessarily human-robot
interaction.
2.3 Adaptable behavior-based architectures
We combine the task of word and concept acquisi-
tion with the behavior-based approach to robotics
(Brooks, 1986) by conceiving of a concept as a
kind of behavior, a process in a competitive hier-
archy. However, traditionally behavior-based sys-
tems have fixed architectures of behaviors designed
in advance for a specific purpose. Our purpose, on
the contrary, is to build and modify an architecture
of protoverbal behaviors at runtime, as the robot
learns to extract them from experience. We there-
fore take an approach similar to that of certain re-
cent expansions in behavior-based robotics, which
have included adaptive architectures. For example,
in (Nicolescu and Mataric, 2001) scenarios are de-
scribed, in which humans act as teachers and col-
laborators for robots. Robots learn to combine ex-
isting behaviors into control networks from human
demonstrations of tasks.
The work reported here builds on the Lateral be-
havior architecture (Fitzpatrick, 1997), which was
extended and modified to enable adaptable behav-
ioral architectures in (Varchavskaia, 2002).
2.4 Novel approach of reported work
The problem we are attempting to address in the cur-
rent work is that of demonstrating the development
of a communicative system by an artificial creature
in a social environment populated by benevolent hu-
mans. This communicative system finds expression
in strings of natural language. However, the em-
phasis of the current approach is not on learning an
extensive vocabulary or a verisimilar grammar, but
on acquiring words with functional meanings. In-
deed, the questions of grammar development are not
addressed here at all, hence we refer to the task as
“protolanguage”, meaning the presyntactic stage in
language development. Reported here are very pre-
liminary results obtained by taking this pragmatic
approach to language acquisition. The main novelty
of this research is the framework proposed here, in
which concepts have procedural as well as declara-
tive meaning. Concepts (i.e., words with meanings)
are processes in a concurrent “mental” architecture
which compete to be expressed by the creature. To
the best of our knowledge, this approach has not yet
been explored in machine learning of natural lan-
guage.
This work was most influenced and inspired by
certain results in developmental linguistics, to which
we now turn.
3. Development of Meaning and Lan-
guage in Humans
Human infants are surprisingly adept at learning
about the structure of the environment, how to be-
have in it, and how to express themselves and un-
derstand others all at once, in the space of a few
years. Bloom (Bloom, 2000) has shown that chil-
dren are good at innately facilitated learning of so-
cially transmitted information. This means that one
of the most important attributes of the spoken lan-
guage in the infant’s environment is that it fulfills a
social function, to which the infant is sensitive and
which enables her to acquire a new word or concept
after a very limited number of examples, perhaps
even one (see also (Pinker, 1999) for overview of ex-
periments suggesting extremely fast word learning by
young children and infants).
In a seminal work (Halliday, 1975), most inspira-
tional for the design of Kismet’s Protolanguage Mod-
ule, Halliday makes the very important distinction
between what he calls the mathetic and the prag-
matic functions of human natural language. The
mathetic function is to provide an encoding of in-
formation channelled through speech or text. When
designing robots who develop their abilities in a man-
ner similar to humans, we cannot simply focus on the
mathetic function of language and neglect its prag-
matic aspect, which is to provide the speaker with a