Zero-pronoun Data Augmentation for Japanese-to-English Translation

Abstract

For Japanese-to-English translation, zero pronouns in Japanese pose a challenge, since the model needs to infer and produce the corresponding pronoun in the target side of the English sentence. However, although fully resolving zero pronouns often needs discourse context, in some cases, the local context within a sentence gives clues to the inference of the zero pronoun. In this study, we propose a data augmentation method that provides additional training signals for the translation model to learn correlations between local context and zero pronouns. We show that the proposed method significantly improves the accuracy of zero pronoun translation with machine translation experiments in the conversational domain.

A Survey on Zero Pronoun Translation

“…Language Bias. Most works used Chinese and Japanese datasets as testbed for training ZP models (Song et al, 2020;Ri et al, 2021). However, there were limited data available for other prodrop languages (e.g.…”

“…They trained an external model on the ZP data to recover the ZP information in the input sequence of the MT model (Tan et al, 2019;Ohtani et al, 2019;Tan et al, 2021) or correct the errors in the translation outputs (Voita et al, 2019). Others aimed to up-sample the training data for the ZPT task (Sugiyama and Yoshinaga, 2019;Kimura et al, 2019;Ri et al, 2021). They preferred to improve the ZPT performance via a data augmentation without modifying the MT architecture (Wang et al, 2016a;Sugiyama and Yoshinaga, 2019).…”

“…They preferred to improve the ZPT performance via a data augmentation without modifying the MT architecture (Wang et al, 2016a;Sugiyama and Yoshinaga, 2019). Kimura et al (2019); Ri et al (2021) verified that the performance can be further improved by denoising the pseudo data. 4.…”

“…Modeling ZPs for advanced NMT models, however, has received more attention, resulting in better performance in this field (Wang et al, 2018a;Tan et al, 2021;Hwang et al, 2021). Generally prior works fall into three categories: (1) Pipeline, where input sentences are labeled with ZPs using an external ZP recovery system and then fed into a standard MT model (Chung and Gildea, 2010;Wang et al, 2016a); (2) Implicit, where ZP phenomenon is implicitly resolved by modelling document-level contexts Ri et al, 2021); (3) Endto-End, where ZP prediction and translation are jointly learned in an end-to-end manner Tan et al, 2021).…”

“…Furthermore, COMET (Rei et al, 2020) is a neural framework for training multilingual MT evaluation models which obtains new SOTA levels of correlation with human judgements. • Pronoun-Aware Translation Quality: Previous works usually evaluate ZPT using the BLEU metric (Wang et al, 2016a(Wang et al, , 2018aRi et al, 2021), however, general-purpose metrics cannot characterize the performance of ZP translation. As shown in Table 3, the missed or incorrect pronouns may not affect BLEU scores but severely harm true performances.…”