Integrating N-best SMT Outputs into a TM System

摘要

In this paper, we propose a novel frame-work to enrich Translation Memory (TM) systems with Statistical Machine Trans-lation (SMT) outputs using ranking. In order to offer the human translators mul-tiple choices, instead of only using the top SMT output and top TM hit, we merge the N-best output from the SMT system and the k-best hits with highest fuzzy match scores from the TM sys-tem. The merged list is then ranked ac-cording to the prospective post-editing ef-fort and provided to the translators to aid their work. Experiments show that our ranked output achieve 0.8747 precision at top 1 and 0.8134 precision at top 5. Our framework facilitates a tight integration between SMT and TM, where full advan-tage is taken of TM while high quality SMT output is availed of to improve the productivity of human translators.