Abstract

This paper presents the NICT’s participation in the WMT19 unsupervised news translation task. We participated in the unsupervised translation direction: German-Czech. Our primary submission to the task is the result of a simple combination of our unsupervised neural and statistical machine translation systems. Our system is ranked first for the German-to-Czech translation task, using only the data provided by the organizers (“constraint’”), according to both BLEU-cased and human evaluation. We also performed contrastive experiments with other language pairs, namely, English-Gujarati and English-Kazakh, to better assess the effectiveness of unsupervised machine translation in for distant language pairs and in truly low-resource conditions.

Highlights

  • This paper describes the unsupervised neural (NMT) and statistical machine translation (SMT) systems built for the participation of the National Institute of Information and Communications Technology (NICT) to the WMT19 shared News Translation Task

  • Since the pseudo-parallel corpora generated by unsupervised SMT (USMT) and unsupervised NMT (UNMT) are of very different nature, and that USMT and UNMT perform in translation quality, we can expect that the complementarity of both data will be useful to train a better NMT system in contrast to using only data generated either by USMT or UNMT

  • Compared with pseudosupervised MT model trained only on pseudoparallel corpora generated by either UNMT (#3) or USMT (#4), merging pseudo-parallel corpora generated by UNMT and USMT (#5) can improve translation performance

Read more

Summary

Introduction

This paper describes the unsupervised neural (NMT) and statistical machine translation (SMT) systems built for the participation of the National Institute of Information and Communications Technology (NICT) to the WMT19 shared News Translation Task. One translation direction was proposed in the unsupervised track of task: German-to-Czech (de-cs). Our submitted systems are constrained, in other words, we used only the provided monolingual data for training our models and the provided parallel data for development, i.e., validation and tuning. We trained unsupervised NMT (UNMT) and unsupervised SMT (USMT) systems, and combined them through training a pseudo-supervised NMT model with merged pseudo-parallel corpora and n-best list

Data and Preprocessing
Unsupervised NMT
Unsupervised SMT
Pseudo-supervised MT
Combination of PNMT and USMT
Generation of n-best Lists
Reranking Framework and Features
Fine-tuning and Post-processing
Results on the German-to-Czech Task
Contrastive Experiments on English-Gujarati and English-Kazakh
Conclusion

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call

Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.