Dr Blaise Thomson

Blaise was elected to a Junior Research Fellowship in October 2010. He researches the development of dialogue systems, which are computers that can interact using spoken language. His research on dialogue systems differs from traditional methods in two key ways:

• Traditional systems base all their decisions on the most likely sequence of words that a person said. However, a computer can compute a list of possible utterances, along with estimates of their probability. By using probability theory and efficient computational techniques, this allows the system to know when it is unsure and to remember alternative hypotheses for what was said. The result is that the computer can quickly correct any misunderstandings.
• When developing a dialogue system, it is difficult to know what questions should be asked, and in what order. A relatively new approach is for the system to learn this automatically. If a particular strategy proves to be effective at helping people then the system will repeat that approach when it encounters a similar situation. Over time, the system can learn optimal decisions for every time it gets a chance to speak.