This paper shows how ASR word lattices can be translated even when exhaustive reordering is required for good translation quality. We propose a method for labeling lattice word hypotheses with position information derived from a confusion network (CN). This information is effectively used in the statistical phrase-based machine translation (MT) search to reduce its complexity, which makes even long-range reordering possible. The proposed method has the benefits of the CN-based MT without having its theoretical drawbacks.
We compare our novel search with the search based on single-best recognition output and on confusion networks. We obtain significant improvements on two translation tasks over the single-best search and gain over the CN search on a task requiring heavy reordering.
Bibliographic reference. Matusov, Evgeny / Hoffmeister, Björn / Ney, Hermann (2008): "Spoken language translation systems ************ ASR word lattice translation with exhaustive reordering is possible", In INTERSPEECH-2008, 2342-2345.