This work seeks to model co-occurrence based semantics, in normal speech output of human beings. It is inspired by functional approaches to speech in normal and language-impaired individuals. We show that it is possible to learn a Maximum Likelihood hypothesis of co-occurrence based semantics across a physiological Short Term Memory by minimizing cross entropy error. Earlier attempts at validation that have sought to evaluate the model, have shown that the model when applied as a text classifier, has an accuracy that is comparable to or better than that possible using Support Vector Machines, as in published literature. This paper describes validation using speech output from normal human beings in the Picture Naming Task.
展开▼
