How ecological regularities can shape linguistic structures

Paul Vogt, 

Tilburg University, The Netherlands, 

p.a.vogt@uvt.nl

A hot topic in language evolution and computation is modelling the
emergence of compositional structures in language, see, e.g., (Batali
1998, Kirby 2001).  However, these models typically take a compositional
structure of the meaning space for granted.  Moreover, these models
assume a predefined meaning space and all the agents in these models
have to do is develop a syntactic language.  I agree that this is
important research from which we learn a lot, but these studies are
bound to overlook crucial aspects of symbol grounding, at least to some
extent. 

One trap that may appear is that one overlooks the possiblity that
agents can exploit the interaction with the environment.  In this paper
I will illustrate, using computational modelling, how agents can exploit
regularities of their ecological niche to shape the compositional
structures they evolve culturally in language.  In this model, agents
develop a compositional structure based on a number of perceptual
features (3 features to represent colour and 1 to represent shape).  The
implicit goal is to develop a compositional language in which sentences
are expressed by two components.  Initially, the agents have no clue
which features belong to colour and which to shape.  Naturally, we hope
to find that the emergent components distinguish between colours and
shapes. 

The model combines the principles behind the Talking Heads experiment
(Steels et al.  2002) with the iterated learning model as was
implemented in (Kirby 2001), and is described in detail elsewhere (Vogt
2003).  In the iterated learning model, language evolves by iterating a
cycle in which learners learn language by observing the linguistic
behaviour of adults, until the adults `die', learners become adults and
new learners enter the population.  When learners enter the population,
they have no categories (meanings), words or grammar; these develop
during their `lifetime'. 

The environment of the agents contains a given number of distinctive
shapes, which can have a fixed number of different colours.  Initially,
perceptual features are categorised holistically, i.e.  by forming
categories as regions in a conceptual space that covers all quality
dimensions (perceptual feature dimensions).  By finding regularities in
the categories that the agents form on different occassions, the agents
are able to group those quality dimensions that have similar values. 
Syntactic structures emerge based on a similar heuristic, which was
adapted from Kirby's (2001) model.  Combining the two mechanisms, the
model exploits a co-development of semantic and syntactic structres. 
The resulting induction mechanisms are similar to those that have
recently been proposed as a model for human language acquisition
(Tomasello 2000). 

Simulations are presented that show how a compositional language can
emerge from scratch.  Moreover, the languages that emerge typically
reflect the regularities found in the perceptual features agents detect
when seeing their environment, and contains linguistic structures
concerning colours and shapes both at the syntactic and semantic level. 
Summarising, the simulations show that a compositional language can
evolve through a combination of cultural evolution (at the syntactic
level), simple induction mechanisms and the interaction of agents with
their environment. 


REFERENCES

J.  Batali.(1998).  Computational simulations of the emergence of
grammar.  In J.  R.  Hurford, M.  Studdert-Kennedy, and C.  Knight,
(editors), Approaches to the Evolution of Language, Cambridge, UK. 
Cambridge University Press. 

S.  Kirby.  (2001).  Spontaneous evolution of linguistic structure: an
iterated learning model of the emergence of regularity and irregularity. 
IEEE Transactions on Evolutionary Computation, 5(2):102--110. 

L.  Steels, F.  Kaplan, A.  McIntyre, and J.  Van Looveren.  (2002). 
Crucial factors in the origins of word-meaning.  In A.  Wray, (editor),
The Transition to Language, Oxford, UK.  Oxford University Press. 

M.  Tomasello.  (2000).  Do young children have adult syntactic
competence? Cognition, 74:209--253. 

P.  Vogt.  (2003).  Iterated learning and grounding: from holistic to
compositional languages.  In S.  Kirby, (editor), Language Evolution and
Computation, Proceedings of the workshop at ESSLLI.