Participants: Philipp Koehn, Logan Athan, Paola Valli, Rudolf Rosa, Sravana Reddy
Agenda for Monday: Attend lab, look at Caitra, read:
We will meet Tuesday after the lectures
With increasing use of machine translation as a tool for human translators, there are a number of ideas to provide more useful assistance than just plain post-editing machine translation output.
One interesting paradigm is interactive machine translation, where the translator types in the translation, while the system suggests the next words or phrases. We have a prototype implementation of this: Caitra.
The partial translation of the user (sometimes called a prefix) is matched against the search graph. If not exact match is found, the closest match according to string edit distance is returned.
A commonly used tool by translators are translation memories (TM). When translating material that is similar to previously translated data (for instance, a revision of a product manual), then finding sentences that have been previously translated, or previously translated sentences that are very close (called "fuzzy matches), can be a useful starting point for human translator.
Such fuzzy matches can be also reformulated as rules for a machine translation system. We have demonstrated in prior work that this is a helpful, achieving better results than plain SMT or TM. This approach could be easily integrated into Moses as a MT Marathon project.
Other possible projects: confidence measures on sentence or word level.