User performance with augmentative communication systems: Measurements and models.

Abstract

Augmentative and alternative communication (AAC) systems can improve the quality of life for individuals with severe disabilities. To design and prescribe AAC systems that provide users with the best performance, the principles of user-system interaction must be better understood. This dissertation focuses on user performance with a particular type of AAC interface known as word prediction. The goal of this work is to improve understanding of the conditions under which word prediction does and does not enhance text generation rate. This goal has been pursued through empirical measurement of actual user performance and by creating and validating quantitative models of user performance with word prediction systems. Two empirical studies were conducted in which a total of 20 subjects (14 able-bodied and 6 with spinal cord injuries) transcribed text with and without word prediction across multiple sessions. Use of word prediction decreased text generation rate for the spinal cord injured subjects and only modestly enhanced it for the able-bodied subjects. This suggests that the cognitive cost of using this word prediction system largely offset the benefit of the keystroke savings provided by the system. Three model implementations were developed and validated against the empirical data. Models 1A and 1B shared a structure which represented performance as a linear combination of two user parameters, keypress and list search time, while Model 2 used a revised model for list search time. For Model 1A, user parameter values were determined independently, while Models 1B and 2 used parameter values derived from subjects' data. All subjects had slower keypress times using word prediction as compared to letters-only typing, and spinal cord injured subjects had much slower list search times than able-bodied subjects. The average errors for Models 1A, 1B, and 2 in simulating subjects' word entry times were 27%, 16%, and 14%, respectively. The two-parameter model was used to simulate expected performance with word prediction across a range of usage conditions. These empirical and modeling studies have demonstrated that: (1) word prediction is not universally effective in enhancing text generation rate; (2) a simple linear model can accurately represent user performance; and (3) simulations using this model can help answer questions regarding the efficacy of word prediction.