Period disambiguation using a neural network

Abstract

Summary form only given. A problem that has never been addressed in the literature is the problem of machine recognition of sentences in real-world documents (i.e. identifying the beginning and end of a sentence). A description is given of the problem and experiments that show that a feedforward neural network can be trained, using the backpropagation learning algorithm, to disambiguate periods. The authors also present results that indicate that a low training tolerance improves a neural network's ability to generalize at the cost of a dramatic increase in the number of training iterations.<>