Abstract:
Aiming at providing linguists with automated experimentation tools, we particularly focused our research on unsupervised neural network models because of their ability to extract/discover categorical structures in real (raw) data, the classification of which is unknown. To further extend the experimentation tools automatism, one aspect of our study deals with the parametric constraints of the models. We thus investigate the models and the various related architectures in order to provide them with evolutive abilities. Among those neural network models, we particularly studied the Kohonen's SOM algorithm for its ability to topologically represent the data structure (between data class relations). In the SOM algorithm, the neighborhood dynamics is a complex parameter which exerts a strong influence on data class differentiation and distribution in the topology. After a report of grammatical results obtained with the classical algorithm, we present a dynamic topology which controls border effects of the planar topology. We then describe an implementation of a dynamic Minimal Spanning Tree topology which prevents from dimensional reduction problems.