3.0 Introduction

Up to now, I have mainly been concerned with sketching the general background of CLASPnet (Chapter 1), and with briefly surveying a small number of related neural network models (Chapter 2). Nonetheless, the overall picture of CLASPnet should already be clear: it is a connectionist model which shows that a modular recurrent neural network can be taught to spot high-level properties of clauses and sentences on the basis of low-level semantic and orthographic information about the words of these clauses and sentences.

In this chapter, I will discuss the concrete implementation of CLASPnet. The chapter is divided into four sections: 3.1 provides a description of the architecture of the network; 3.2 deals with the context-free grammar developed to generate testing and training corpora; 3.3 summarizes the main results obtained during training and testing the network; and 3.4, finally, describes a number of experiments carried out to find out the relative importance of various factors. The general discussion of all the results will be presented in Chapter 4.