Keystroke-level models for user performance with word prediction

• Published in: Augmentative and alternative communication 1997, Vol.13 (4), p.239-257
• Main Authors: Koester, Heidi Horstmann, Levine, Simon
• Format: Article
• Language: English
• Subjects: augmentative communication; group designs; rate enhancement; user modeling; word prediction
• ISSN: 0743-4618; 1477-3848
• DOI: 10.1080/07434619712331278068

Two modeling studies have been performed to develop quantitative models of user performance with word prediction and test their predictions against empirical data. In the first study, the model structure represented performance as a linear combination of two user parameters, keypress and list search time. Two types of simulations were performed using this structure: Model 1A, in which user parameter values were determined independently, and Model 1B, which used parameter values derived from subjects' data. Model simulations of overall session performance, word entry times, and item selection times were compared to actual performance of able-bodied and spinal cord injured subjects transcribing text with and without word prediction over the course of seven test sessions. The average errors for Models 1A and 1B in modeling subjects' word entry times were 27% and 16%, respectively. The second study used a revised model for list search time in an attempt to improve model accuracy and increase understanding of the list search process. The model revision led to only a small improvement in accuracy but did provide insight into how list search time depends on the context of the search. The results point out the need to understand a user's characteristics before applying a model, but they are an encouraging demonstration of the ability of analytical models to represent user performance with word prediction.