ADAPTIVE PREDICTIVE TEXT GENERATION AND THE REACTIVE KEYBOARD

Abstract

This thesis discusses the application of predictive text generation to enhance the communication abilities of physically disabled persons. Predictive techniques exploit the statistical redundancies of language to accelerate and amplify user inputs, thereby increasing communication efficiency. Acceleration is achieved by making more likely language elements faster to select, while amplification is accomplished by selection of concatenated elements. Novel adaptive language models are used to enhance versatility and flexibility. Predictive text generation (PTG) is defined, existing PTG systems are surveyed and a framework is developed for classifying and evaluating them. Simulation studies and a user pilot experiment are persented for a particularly significant PTG system called \fIPredict\fR. Experience with Predict led to the Reactive Keyboard concept of PTG. The research results presented highlight design issues common to the two systems. The Reactive Keyboard concept is introduced along with its prediction technique, data structure and two user interface implementations: \fIRK-button\fR and RK-pointer. A clear distinction is made between the system's user interface and the underlying model it employs. A variable-length n-gram language model is presented which adaptively gathers statistics from the user's text input. A number of alternative model structures are discussed and details of a novel, highly compact, data storage technique are given. Context conditioned candidate strings, which are predicted by the model, are ordered according to popularity and displayed for selection on a VDU.