Abstract
The advantages of neural machine translation (NMT) have been extensivelyvalidated for offline translation of several language pairs for differentdomains of spoken and written language. However, research on interactivelearning of NMT by adaptation to human post-edits has so far been confined tosimulation experiments. We present the first user study on online adaptation ofNMT to user post-edits. Our study involves 29 human subjects whose post-editingeffort and translation quality were measured on about 4,500 interactions of ahuman post-editor and a machine translation system integrating an onlineadaptive learning algorithm. Our experimental results show a significantreduction of human post-editing effort due to online adaptation in NMTaccording to several evaluation metrics, including hTER, hBLEU, and KSMR.Furthermore, we found significant improvements in BLEU/TER between NMT outputsand human references, and a strong correlation of these improvements withquality improvements of post-edits.