| Demand for the services of translators is on the increase, and consequently so is the demand for tools to help them improve their productivity. This thesis proposes a novel tool intended to give a translator interactive access to the most powerful translation technology available: a machine translation system. The main new idea is to use the target text being produced as the medium of interaction with the computer. In contrast to previous approaches, this is natural and flexible, placing the translator in full control of the translation process, but giving the tool scope to contribute when it can usefully do so.; A simple version of this idea is a system that tries to predict target text in real time as a translator types. This can aid by speeding typing and suggesting ideas, but it can also hinder by distracting the translator, as previous studies have demonstrated. I present a new method for text prediction that aims explicitly at maximizing a translator's productivity according to a model of user characteristics. Simulations show that this approach has the potential to improve the productivity of an average translator by over 10%.; The core of the text prediction method presented here is the statistical model used to estimate the probability of upcoming text. This must be as accurate as possible, but also efficient enough to support real-time searching. I describe new models based on the technique of maximum entropy that are specifically designed to balance accuracy and efficiency for the prediction application. These outperform equivalent baseline models used in prior work by about 50% according to an empirical measure of predictive accuracy, with no sacrifice in efficiency. |
