Implementing Soft Preferences for Structural Disambiguation

Abstract

A simulation is presented here that demonstrates how a standard BP net, taking whole sentences as input, may be trained to perform a structural disambiguation task and to generalise to novel examples. It is argued that, during learning, the net implements raw attachment preferences in the relationship between the upper and lower weights. An analysis is provided of how these preferences are implemented, and it is shown how each individual word may be assigned a raw preference value (RPV) which may be used as an indicator of its structural bias. Moreover, it is shown how the activation function makes use of the raw preferences by modulating their strength of their biases in a manner that is contextually sensitive to the structural biases of all the other words in a sentence.

This project was funded by an award from the British Telecom Research Laboratories at Martlesham Heath (under the CONNEX iniative). I would like to thank Richard Sutcliffe (RA on the project) for running the simulation and Paul Day for conducting the normative study. I would also like to acknowledge Don Mitchell, Ajit Narayanan. and Peter Wyard for their support and suggestions.