Semantic Extraction using Neural Network Modelling and Sensitivity Analysis.pdf

Semantic Extraction Using Neural Network Modelling and Sensitivity Analysis TH Goh, Francis Wong Institute of Systems Science National University of Singapore Heng Mui Keng Terrace, Kent Ridge Singapore 0511 Email: ISSGTH@NUSVM.BITNET thgoh@iss.nus.sg Abstract Neural networks have often been faulted as black box systems that do not explain clearly the way it arrives at its conclusions. In this paper, we present three methods of using neural network modelling and sensitivity analysis to extract semantics from the historical data of a given system. First, neural network modelling and sensitivity analysis is used to determine the decision boundaries of the system. Several test cases including a simple credit rating system are described to illustrate the use and effectiveness of this method. Secondly, it is used for causal inferencing of the system. That is, determining which inputs has the largest effect on the output of the system. Typical problems like the decoder and parity are discussed. Lastly it is used to analyze the historical data for detection of exceptions or special input cases. In all three methods, the system being studied is first modelled using a neural network. Sensitivity analysis is then applied to the trained network to extract semantics learned by the network. Introduction Neural networks have often been faulted as black box systems that do not reveal the way its conclusions are reached. In this paper, we present three methods of using sensitivity analysis to extract semantics from systems that hav