Translation Memories guarantee consistency: Truth or fiction?

Translation Memory (TM) is a key technology in the translation industry, its success based on several assumptions: that it saves time, provides cost savings, and enhances consistency. Our research tests the assumption of consistency in TMs. We examine English/German and English/Japanese TM data from two commercial companies with a view to measuring the levels of consistency (or inconsistency) between TMs across product releases. To meet our aim we first had to develop a method for interrogating TMs for consistency. Following a pilot study, we chose to categorise translation units based on whether the TM translation process had introduced consistency or inconsistency. Additionally, we examined source text segments that contained minor inconsistencies where the corresponding target text introduced further inconsistency. By identifying categories of inconsistency along with investigating the accompanying metadata, this case study hopes to highlight the types of inconsistency that can occur in TM data and to suggest how TM tools might overcome this problem.