Octanove Labs’ Japanese-Chinese Open Domain Translation System

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 166–171 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 166 Octanove Labs’ Japanese-Chinese Open Domain Translation System Masato Hagiwara Octanove Labs LLC Seattle, WA masato@octanove.com Abstract This paper describes Octanove Labs’ submis- sion to the IWSLT 2020 open domain transla- tion challenge. In order to build a high-quality Japanese-Chinese neural machine translation (NMT) system, we use a combination of 1) parallel corpus filtering and 2) back transla- tion. We have shown that, by using heuristic rules and learned classifiers, the size of the parallel data can be reduced by 70% to 90% without much impact on the final MT perfor- mance. We have also shown that including the artificially generated parallel data through back-translation further boosts the metric by 17% to 27%, while self-training contributes lit- tle. Aside from a small number of parallel sen- tences annotated for filtering, no external re- sources have been used to build our system. 1 Introduction Building a robust, open domain machine transla- tion (MT) system for non-English, Asian languages remains a challenge since many MT research ef- forts have focused mainly on European languages (such as English and German) and/or on particular domains (such as news). This is especially the case when there is lack of high-quality, human-curated parallel corpora and one needs to bootstrap an MT system from noisy parallel data crawled from the Web. This is the exact setting of the IWSLT 2020 open domain translation challenge (Ansari et al., 2020), where the organizers provide large, noisy parallel datasets crawled from the Web and the participants build open-domain machine translation systems be- tween Japanese (JA) and Chinese (ZH). The partic- ipants are also encouraged to only use the provided datasets to train their models. Therefore, the key to building high-quality MT systems seems to be in how to filter and make the most of the provided, noisy datasets. Existing parallel (EP) Web crawled (WC) (a) Annotate (a) Annotate (b) Filter (b) Filter Official Baseline BASELINE FILTERED (c) Existing parallel filtered (EPf) (c) Web crawled filtered (WCf) (e) Web crawled remainder (WCr) (f) (g) (g) (f) Official dev. dataset (DEV) (d) Our dev. dataset (DEV+) COMBINED (ZH→JA) COMBINED (JA→ZH) FINAL (ZH→JA) FINAL (JA→ZH) Back-Translation Parallel Corpus Filtering JA (MT) ZH (MT) ZH JA Figure 1: Overview of the data/training pipeline Based on these insights, we used a combina- tion of 1) parallel corpus filtering (Koehn et al., 2018, 2019) and 2) back-translation (Sennrich et al., 2016; Edunov et al., 2018) techniques as our main strategy. For 1), we showed that we can reduce the size of the parallel corpora by 70% to even 90% without much impact on the final MT perfor- mance. As demonstrated in (Chen et al., 2019), we also verified that artificially generated parallel data through back-translation can further help im- prove the performance by 17% to 27% depending on the direction. We used the vanilla Transformer (Vaswani et al., 2017) as our NMT model. In …

Content truncated to 500 words