Referential Choice and Activation Factors: A Neural Network Approach

In Kibrik’s sample-based study of referential choice presented at DAARC 2000, a quantitative approach was proposed to calculate an activation score of a specific referent at a particular point of discourse from a range of partly interdependent factors, such as distance to the antecedent, referent animacy, etc. The activation score then determines the referential choice. The advantage of that multi-factorial approach is that it is explanatory and testable. That approach, however, was mathematically rather unversed and had some shortcomings. The list of factors and their interaction to yield the activation score were hand-coded. The interaction was purely additive, ignoring possible non-linear interdependencies between the factors. In this paper we propose a more sophisticated neural network analysis of the same data. We trained a feed-forward network on the data. It classified 96% of all data correctly with respect to the actual referential choice. A pruning procedure allowed to produce a minimal network and revealed that out of ten input factors five were sufficient to predict the data almost correctly, and that the logical structure of the remaining factors can be simplified. This is a pilot study necessary for the preparation of a larger neural network-based study. The neural network approach allows to model the set of necessary and sufficient activation factors explaining referential choice in discourse. Its big advantage over classical statistical methods is that the type of regularities it can detect in the data is less constrained.