Data Augmentation for Inline Tag-Aware Neural Machine Translation

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 886 - 894 December 7-8, 2022 ©2022 Association for Computational Linguistics Data Augmentation for Inline Tag-Aware Neural Machine Translation Yonghyun Ryu, Yoonjung Choi, Sangha Kim Samsung Research Seoul, Republic of Korea {yonghyun.ryu, yj0807.choi, sangha01.kim}@samsung.com Abstract Despite the wide use of inline formatting, not much has been studied on translating sentences with inline formatted tags. The detag-and- project approach using word alignments is one solution to translating a tagged sentence. How- ever, the method has a limitation: tag reinser- tion is not considered in the translation process. Another solution is to use an end-to-end model which takes text with inline tags as inputs and translates them into a tagged sentence. This ap- proach can alleviate the problems of the afore- mentioned method, but there is no sufficient parallel corpus dedicated to such a task. To solve this problem, an automatic data augmen- tation method by tag injection is suggested, but it is computationally expensive and augmen- tation is limited since the model is based on isolated translation for all fragments. In this paper, we propose an efficient and effective tag augmentation method based on word alignment. Our experiments show that our approach out- performs the detag-and-project methods. We also introduce a metric to evaluate the place- ment of tags and show that the suggested metric is reasonable for our task. We further analyze the effectiveness of each implementation detail. 1 Introduction While most machine translation studies are focused on plain text, the textual information that we en- counter every day on the internet contains words with different styles and links within the sentence. Various styling of any part of the text is called in- line formatting and is represented by markup and markdown tags. The inline formatting not only improves the readability of documents but also pro- vides additional information with tags; so it is im- portant to correctly translate sentences including tag information. In addition, the widespread use of formatting tags in the computer-based document system makes it inevitable to increase the demand for translating web text or structured documents containing inline tags. There are two main approaches to translating segments with inline tags. One solution is the detag- and project method (Hanneman and Dinu, 2020). It first strips tags from the source sentence and trans- lates only the plain text. Then, the removed tags are reinserted into the translation results using word alignments, which can be induced from attention weight in the model or an external aligner such as SimAlign (Jalili Sabet et al., 2020). This method does not take into account the re-insertion of tags in the translation process, making it difficult to restore tags at the proper positions. Another way is to use an end-to-end model which takes sentences including tags as inputs and generates translation results with tags. Since tag information is considered, the translation can be performed with more context, thus this method po- tentially improves the quality of translation and the placement of tags. …

Content truncated to 500 words