In this project, my team and I have explored using different levels of linguistic representations to improve Machine Translation Evaluation. We have looked at combining discourse trees, semantics and syntax to improve the state-of-the-art. We have used both structured (trees) and distributed (vector) representations to perform this task. Currently, we're looking at how humans evaluate translations using eyetracking.
The meeting and lecture translation project aims to tackle the challenges of spoken translation in spontaneous situations where there are: 1) multiple speakers (meetings) and 2) single speakers but highly technical content (lectures).