BLEU and constraint completion on IATE and Wiktionary
This CSV transcribes Table 2 of Translate-and-Revise. It gives BLEU and constraint completion rate for eight systems or prompting settings on two English-to-German lexically constrained translation test sets.
How did revision affect translation quality and constraint completion?
With GPT-3.5-turbo-0613 on IATE, adding revision to the constrained prompt raises constraint completion from 96.2% to 98.9% at the same BLEU of 32.0. On Wiktionary, completion rises from 97.2% to 98.9% while BLEU changes from 32.1 to 32.0. Both metrics are included for all eight settings, so a gain on one can be examined alongside its effect on the other.
Each row is one reported system–test-set result. This author-prepared transcription uses the paper’s absolute scores and includes all five neural machine translation baselines. Table 2 in the CCL paper
Data files
Related paper
Translate-and-Revise: Boosting Large Language Models for Constrained Translation · Official BibTeX