TER

Translation Edit Rate is one of the standard error metrics for automatic evaluation ([Snover_2006]). The whole idea is described in the Research paper. Briefly, the formula is following:

TER = # of edits / average # of reference words

The value of TER is between 0 and 1, where lower is better.

Implementation Details

The current implementation is based on SacreBleu library( [SacreBleu]), which contains TER module.

Our settings are following:

  • we use no case distinction (TER default)

  • we apply normalization (for Chinese + Japanese languages the specific tokenization is applied as well)

  • we remove the brackets around tags (<bold> becomes bold)

  • For Thai we also apply the Apache OpenNLP tokenizer

Reference

[Snover_2006]

Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla and John Makhoul: A Study of Translation Edit Rate with Targeted Human Annotation. Proceedings of Association for Machine Translation in the Americas. 2006.