Learning Interpretation Techniques from Human

Real-time simultaneous translation is a demanding task. The translator must translate partial sentences before knowing how the sentence will end. Computational approaches to do this difficult, important task are limited because we lack a quantitative understanding of the tactics simultaneous translators use. We study characteristics of interpreted text and translated text through an automatic, computational approach. Our analysis reveals that human interpretors apply effective tactics to reduce translation latency, such as segmentation and passivization. These tactics are predictable given only partial text input, a necessary first step before we can integrate these techniques with current batch translation systems.