User Modeling for Individuals with Disabilities: A Pilot Study of Word Prediction
Abhishek Agarwal, Richard Simpson · 2005 · Proceedings of the 7th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '05) · doi:10.1145/1090785.1090838
Summary
This short paper describes the development of user models that predict how word prediction (WP) systems affect text entry performance for individuals with disabilities. The research addresses a practical clinical problem: clinicians who assess assistive technology for clients have limited time, limited access to the full range of available devices, and difficulty predicting how a client's initial performance with a device will translate into expert-level performance over time. The authors propose that accurate user models could allow clinicians to simulate client performance across various device configurations, using quantitative predictions to identify the best candidates for detailed evaluation. The study focuses on word prediction specifically because, while WP clearly reduces the number of keystrokes needed to generate text, empirical evidence shows it does not necessarily increase text generation rate (TGR) and may actually reduce it due to the cognitive overhead of monitoring and selecting from the prediction list.
Key findings
The pilot study used a sophisticated multi-computer instrumentation setup: a subject computer with a word prediction interface, an eye tracking system (ISCAN 726 PCI) with desk-mounted camera to monitor visual attention, and a Biopac MP150 data acquisition system with an accelerometer to measure hand movement. Four able-bodied subjects typed using a single finger (simulating physical disability) under varied experimental conditions — WP list size (2 or 8 words), minimum word length for predictions (3 or 5 letters), WP display timing (immediate, after 2 keystrokes), and WP hiding behaviour (never hidden, hidden after 2 keystrokes). The pilot revealed several methodological issues: some subjects typed too quickly to benefit from WP; subjects sometimes entered text without looking at the screen (unobserved by eye tracking); simultaneous key presses occurred; and the accelerometer struggled to distinguish between keys that were close together. Based on these findings, the authors plan to switch to an on-screen keyboard with mouse input to better control text entry speed and capture hand movement data.
Relevance
This work addresses a persistent challenge in assistive technology provision: matching the right technology configuration to individual users without exhaustive trial-and-error. The user modelling approach — simulating performance across configurations to narrow the field before hands-on evaluation — anticipates modern personalisation and recommendation systems in assistive technology. The finding that word prediction can paradoxically slow text generation despite saving keystrokes remains a critical consideration for AT practitioners configuring text entry systems. The cognitive cost of monitoring a prediction list, deciding whether to accept a prediction, and executing the selection can outweigh the keystroke savings, particularly for users who type at moderate speeds. This tradeoff between motor efficiency (fewer keystrokes) and cognitive load (decision delay) is a design tension that persists in modern predictive text systems across all platforms.
Tags: word prediction · user modeling · assistive technology · text entry · eye tracking · keystroke savings · AAC