Improving Word-Based Predictive Text Entry with Transformation-Based Learning
David J. Brooks, Mark Lee · 2008
Abstract. Predictive text interfaces allow for text entry on mobile phones, using only a 12-key numeric keypad. While current predictive text systems require only around 1 keystroke per character entered, there is ambiguity in the proposed words that must be resolved by the user. In word-based approaches to predictive text, a user enters a sequence of keystrokes and is presented with an ordered list of word proposals. The aim is to minimise the number of times a user has to cycle through incorrect proposals to reach their intended word. This paper considers how contextual information can be incorporated into this prediction process, while remaining viable for current mobile phone technology. Our hypothesis is that Transformation-Based Learning is a natural choice for inducing predictive text systems. We show that such a system can: a) outperform current baselines; and b) correct common prediction errors such as “of ” vs. “me ” and “go ” vs. “in”. 1