LDC provided translations for a large volume of BOLT Chinese and Egyptian Arabic source data, including discussion forum threads in Phase 1 and text message (SMS) and chat conversations in Phase 2. Data was selected and segmented into sentence units (discussion forum) and message units (SMS/chat) for translation That ensured that only the appropriate data was translated and also allowed for the alignment of the resulting parallel text at the sentence or message level.
All translations were created by outsourcing to selected agencies. Incoming training data was manually reviewed by scoring a sample of each batch to ensure translation quality. LDC staff performed manual corrections on devtest and evaluation data. Enhancement of evaluation data included the insertion of translation variants (alternates) into the translations to represent ambiguity in the source.
LDC provided post-editing of machine translation output to compute "edit distance" between machine translations and human gold standard translations. The post editor's task was to compare computer-translated texts against the same texts translated by humans, modifying the computer translation until its meaning is identical to the human translated sentence.
Post editors used a custom interface, which presented kits consisting of machine - and human - translation pairs and provided a workspace for editing the machine translation. First-pass editors completed each kit with the goal of making the meaning of the machine translation match the meaning of the human reference translation with the smallest possible number of edits. A second-pass editor then checked for errors and for places where the same meaning could be achieved with fewer edits. Each kit was edited by two separate teams, and kits with a large discrepancy in scores between the two teams were manually reviewed by task mangers.