17 janvier 2013
info:eu-repo/semantics/OpenAccess
Joss Moorkens et al., « A Virtuous Circle: Laundering Translation Memory Data using Statistical Machine Translation. Session 3 - Machine and Human Translation: Finding the Fit? », HAL-SHS : linguistique, ID : 10670/1.zxftw1
This study compares consistency in target texts produced using Translation Memory (TM) with that of target texts produced using Statistical Machine Translation (SMT), where the SMT engine is trained on the same texts as are reused in the TM workflow. These comparisons focus specifically on noun and verb inconsistencies, as such inconsistencies appear to be highly prevalent in TM data (Moorkens 2012). We then go on to substitute inconsistent TM target text nouns and verbs by consistent nouns and verbs from the SMT output, and to test (1) whether this results in improvements in overall TM consistency and (2) whether an SMT engine trained on the 'laundered' TM data performs better than the baseline engine. Improvements were observed in both TM consistency and SMT performance, a finding that indicates the potential of this approach to TM/MT integration.