Provided by Cognitive Sciences ePrint Archive
Behavior-Based Early Language Development on a
Humanoid Robot
Paulina Varshavskaya (Varchavskaia)
Artificial Intelligence Lab, MIT
NE43-937, 200 Technology Square
Cambridge, MA 02139 US
Abstract
We are exploring the idea that early lan-
guage acquisition could be better modelled on
an artificial creature by considering the prag-
matic aspect of natural language and of its de-
velopment in human infants. We have imple-
mented a system of vocal behaviors on Kismet
in which “words” or concepts are behaviors in
a competitive hierarchy. This paper reports
on the framework, the vocal system’s archi-
tecture and algorithms, and some preliminary
results from vocal label learning and concept
formation.
1. Introduction
We are exploring the idea that early language acqui-
sition could be better modelled on an artificial crea-
ture by considering the pragmatic aspect of natural
language and its development in human infants. We
believe that this will contribute to a solution to the
“Grounding Problem” (Harnad, 1990) by providing
a new level of grounding in intentions and function.
The pragmatic approach to language acquisition is
to consider first of all the intentions of a speech act.
Language is not viewed as a denotational symbolic
system for reference to objects and relationships be-
tween them, as much as a tool for communicating
intentions. The utterance is a way to manipulate
the environment through the beliefs and actions of
others.
We develop a system of vocal behaviors for the
robot Kismet (Breazeal, 2000) which exemplifies the
approach we believe should be taken to natural lan-
guage acquisition by machines. The robot’s youthful
appearance dictates the sort of interaction that hu-
mans will have with it: scenarios of social scaffolding
similar to the kinds of interactions that teachers have
with human infants.
One step forward is to enable the robot to use
the scaffolded environment to its advantage in order
to learn to perform tasks and behave appropriately.
Learning to communicate with the teachers using a
shared semantic basis is one aspect of learning to be-
have in the world and manipulate it. We augment
the existent motivational and behavioral systems of
the robot with a set of vocal behaviors, regulatory
drives, and learning algorithms, which together con-
stitute Kismet’s Protolanguage Module.
In what follows, we first take a look at previous
work in robotics in Section 2., and in human lan-
guage development in Section 3. Then we present
the architecture of Kismet’s protolanguage module
in Section 4. and the algorithms in Section 5. Some
preliminary results from experiments with the new
system can be found in Section 6., and a discussion
in Section 7.
2. Previous Work
The current work was inspired by and built upon
results and ideas in robotic language acquisition and
adaptable behavior-based robotic architectures.
2.1 Robots acquiring natural language
(Roy, 1999) and (Oates et al., 2000), approached the
problem of acquisition of natural categories and la-
bels by robots from the point of view of perceptual
grounding. The robot analyzes the visual scene and
the speech stream into segments, the best correlation
between which will form a perceptual concept-label
pair which is acquired by the robot, as for example in
the development of CELL (Roy, 1999). CELL is em-
bodied in an active vision camera and acquires lexical
units from the following scenario: a human teacher
places an ob ject in front of the robot and describes it.
The visual system extracts color and shape proper-
ties of the ob ject, and CELL learns on-line a lexicon
of color and shape terms grounded in the represen-
tations of ob jects. The terms learned need not be
pertaining to color or shape exclusively - CELL has
the potential to learn any words. Associations be-
tween linguistic and contextual channels are chosen
on the basis of maximum mutual information.