TER
TER, the translation edit rate, measures how many edits, insertions, deletions and substitutions, it would take to turn machine output into a reference translation. Lower is better. Because it maps onto real editing effort, it is used to estimate and price post-editing before the work starts.
How it works
TER counts the minimum edits, insertions, deletions, substitutions and shifts, needed to turn the machine output into the reference, then divides by the reference length. A lower number means fewer changes, so closer to the reference. Zero means identical.
Because an edit in the metric roughly stands for an edit a human would make, TER maps onto real post-editing effort more directly than overlap scores, which is why it is used to estimate and price editing work.
How SourceTarget uses it
SourceTarget uses TER-style edit distance to estimate how much post-editing a job will need before it starts, which feeds both the quote and the routing of work. Segments with high expected edit rates are prioritised for the editor.
TER compared with BLEU score
| TER | BLEU score | |
|---|---|---|
| Direction | Lower is better | Higher is better |
| Measures | Edits to reach the reference | Overlap with the reference |
| Best for | Estimating editing effort and cost | Comparing systems' overlap |