Recurrent neural network classifier for three layer conceptual network and performance evaluation

Abstract

Natural language has traditionally been handled using symbolic computation and recursive processes. Classification of natural language by using neural network is a hard problem. Past few years several recurrent neural network (RNN) architectures have emerged which have been used for several smaller natural language problems. In this paper, we adopt Elman RNN classifier for disease classification for a doctor patient-dialog system. We find that the Elman RNN is able to find a representation for natural language. Contextual analysis in dialog is also a major problem. A three layers memory structure was adopted to address the challenge which we referred to as ”Three Layer Conceptual Network” (TLCN). This highly efficient network simulates the human brain by discourse information. An extended case structure framework is used to represent the knowledge. We used the same case frame structure to train and examine the RNN classifier. This system prototype is based on doctor-patients dialogs. The over all system performance achieved 84% accuracy. Disease identification accuracy depends on number of disease and number of utterances. The performance evaluation is also discussed in this paper.