| 研究実績の概要 |
To improve the machine translation (MT) quality in this low resource scenarios, we studied the following in FY2018:
1. We continued our research by studying the topics of data adaptation using large-scale monolingual web corpora and multiple resource adapted system integration as scheduled. We published a journal paper in the journal of information processing, in which we conducted a comprehensive comparison of previous studies in these two topics.
2. We conducted a survey of domain adaptation for MT and published a survey paper at COLING 2018. Our survey paper covered the techniques for improving low resource domain translation in both historical and practical perspectives, which can be a good start point for both researchers and engineers working on this area. We also gave a talk on this topic to translators at JAITS 2019 to promote the practical use of these techniques.
3. We also studied more general topics for MT, which are not limited to low resource scenarios. We proposed a recursive neural network based pre-ordering model to improve the translation quality of distant language pairs such as Japanese-English, and published our work at ACL-SRW 2018 and the journal of natural language processing. We also studied a word rewarding model to improve the translation adequacy using bilingual dictionaries, and published our work at IWSLT 2018.
4. Using the techniques we developed in this project, we attended the MT shared task at WAT 2018. We have showed that our techniques can significantly improve low resource MT such as the Myanmar-English language pair.
|
