Various recent studies have approached phonology as a complex adaptive system (Lindblom, 1998, De Boer, 1997). In this work we study the emergence of phonology in a population of artificial agents communicating by sounds to define which are the basic rules in interaction that lead to the emergence of the categorisation of phonology.

The tool used for modelling in this experience is a multi-agent system because it allows the study of emerging language phenomena in a population of artificial communicating agents.

The phonetic naming game is a module of the LEMMingS (Language Evolution Monitoring and Modelling System) platform developed at the Laboratory Dynamique Du Langage. It defines a new type of interactions inspired from the imitation game (De Boer, 1999) and the naming game (Steels, 1997) and allows to study phonological aspects of the origins of language in a population of agents exchanging monosyllabic words.

At the beginning of the game, the agents have empty memories. In the interactions, they will have to name the objects of their virtual environment and invent a phonology for their lexicon.

The structure of the agents is composed of two modules and two memories :

• The production module is an acoustic tubes dynamic model based on the Distinctive Region Model of Carré (1995) : a sensibility function allows to define dynamically which are the acoustic modification produced by the gestures of the vocal tract.
• The perception module : the agents segment the acoustic signal into three formants coordinates (values of the 3 first formants of the vowel. The agents are able to retrieve the articulation of a sound from acoustic values with an inverse mapping based on the sensibility function of the DRM model. A threshold is given to the perception module to define the maximal acoustic distance tolerated between two sounds.
• The repertoire of sounds is composed of all the productions stored by the agent. The "phonemes" are represented in the phonetic memory by associations between articulatory gestures and the formants values of the sounds.
• The associative memory stores the lexicon. Words are formed by a sequence of three "phonemes" and are represented as an association by a meaning and a topic.

Two agents of the population randomly chosen play the game. The speaker chooses a topic and sends the associated name to the hearer. The imitator looks in his phonetic memory for the closest representative sounds corresponding to the received sequence. If he has the 3 sounds, he looks if an association corresponds to the topic pointed by the speaker and compares the association to the one sent by the initiator. If the agent has no similar association, he creates a new word for the topic. The sounds of the sequence lacking from the phonetic memory are reproduced with an inverse mapping procedure. The association is then sent to the initiator.

The speaker compares the sounds sequence to the one he has emitted and sends a signal of failure or success to the receiver. At the end of the game the success and use of the sounds and of the associations are updated. If during the game the hearer has stored a new sound, he merges it with the closest sound of his repertoire that is in the merging area.

These first results show a good convergence : the agents have similar vocalic systems although they present inter-individual differences.

Nevertheless, the vocalic systems resulting of the communications in a population of 5 agents for 10 concepts don't show a similarity to any of the phonetic systems encountered in the world's languages.

The agents vocalic systems developed should show a greater realism with the implementation of the coarticulation model producing consonants and a more fine-grained perception module that can take into account syllables transitions.

The game offers a good way to study the lexicon pressure on the phoneme creation and the influence of the population size and the categorisation procedure on the emergence of common vocalic systems in a population of communicating individuals (Lindblom, Mac Neilage and Studdert-Kennedy, 1984).

Although this experience is an oversimplification of the communication in human populations, it offers a good tool to study the minimal conditions for the emergence phonology.

 Conference site: http://www.infres.enst.fr/confs/evolang/