Raising the TM Threshold in Neural MT Post-Editing: a Case Study onTwo Datasets

  title={Raising the TM Threshold in Neural MT Post-Editing: a Case Study onTwo Datasets},
  author={Anna Zaretskaya},
  • Anna Zaretskaya
  • Published in MTSummit 2019
This study intends to determine whether replacing fuzzy TM matches by suggestions from neural machine translation (NMT) can decrease the post-editing effort. We compare the post-editing distance of TM fuzzy matches and of NMT suggestions based on two datasets. We found that in one of the datasets MT was consistently more useful than TM matches, but in the other dataset it was not. We argue that it is necessary to collect extensive data on PED in TM matches in order to be able to easily optimize… CONTINUE READING

Figures, Tables, Results, and Topics from this paper.

Key Quantitative Results

  • On average, switching from our previous statistical MT framework to the current neural one decreased the post-editing distance by 9.2%, which means an improvement in quality of approximately 29%.
  • We have seen that a customized system can improve the PED by up to 20% compared to a baseline generic system.
  • An experiment that had been conducted at TransPerfect showed that a customization with additional 100 000 new translation units yields about 4% increase of the PE distance over the baseline system, and the quality grows exponentially when adding more data.


Publications referenced by this paper.

Perception vs

Sánchez-Gijón, Pilar, +3 authors Andy.
  • Acceptability of TM and SMT Output: What do translators prefer? EAMT 2018, 21st Annual Conference of the European Association for Machine Translation, Alicante, Spain, 331.
  • 2018

The Challenge of Machine Translation Post-editing: An Academic Perspective

Rico, Celia, +3 authors Olga
  • Trends in ETools and Resources for Translators and Interpreters,
  • 2018