Text Prediction in Web-based Text-Processing
Summary
This thesis describes the research undertaken to investigate the optimal implementation of text prediction. Using a 16.7Tb large dataset, a double-blind experiment is executed with 156 participants which predicts up to 4 words ahead what a user is typing.
Following the main research question: Which User Interface Configuration Supports an Optimal Implementation of Text Prediction in Physical Keyboard Supported Software? several hypotheses are defined. The optimal implementation is defined as the text prediction system that leads to the least errors over the shortest time it takes a participant to retype a text.
The results show that that optimal implementation is no text prediction. The text prediction system used for the experiment made users type more errors and it took users more time to retype a text with the system enabled than without. A significant result was found in the influence of the user interface used and the amount of words predicted. There are in total four user interfaces tested, and the autofill interface, among many familiar from Google’s search bar, results in the least amount of errors typed.
However, even with this interface enabled,typing without text prediction results in significantly less errors.
The amount of words predicted varies from 0 (the words being typed is completed by the system) to 4 (the word being typed is completed and four additional words are predicted). With three words predicted, the amount of errors typed is the lowest. But likewise to the user interface, the control experiment without prediction performs significantly better.
It is from these results that the author can only conclude that the optimal implementation of text prediction is no text prediction. In further research, the author presents an inverted way of text prediction which might increase the accuracy of the predictions, and a longitudinal study to further understand the significant results.