Easy Guided Decoding in Providing Suggestions for Interactive Machine Translation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7840–7852 July 9-14, 2023 ©2023 Association for Computational Linguistics Easy Guided Decoding in Providing Suggestions for Interactive Machine Translation Ke Wang∗, Xin Ge∗, Jiayi Wang, Yu Zhao, Yuqi Zhang† Alibaba Group Inc. {moyu.wk,shiyi.gx,joanne.wjy,kongyu, chenwei.zyq}@alibaba-inc.com Abstract Machine translation technology has made great progress in recent years, but it cannot guarantee error-free results. Human translators perform post-editing on machine translations to correct errors in the scene of computer-aided transla- tion. In favor of expediting the post-editing pro- cess, many works have investigated machine translation in interactive modes, in which ma- chines can automatically refine the rest of trans- lations constrained by human’s edits. Transla- tion Suggestion (TS), as an interactive mode to assist human translators, requires machines to generate alternatives for specific incorrect words or phrases selected by human translators. In this paper, we utilize the parameterized ob- jective function of neural machine translation (NMT) and propose a novel constrained de- coding algorithm, namely Prefix-Suffix Guided Decoding (PSGD), to deal with the TS prob- lem without additional training. Compared to the state-of-the-art lexically constrained decod- ing method, PSGD improves translation qual- ity by an average of 10.87 BLEU and 8.62 BLEU on the WeTS1 and the WMT 2022 Trans- lation Suggestion datasets2, respectively, and reduces decoding time overhead by an aver- age of 63.4% tested on the WMT translation datasets. Furthermore, on both of the TS bench- mark datasets, it is superior to other supervised learning systems trained with TS annotated data. 1 Introduction and Related Work The emergence of machine translation technology (Lopez, 2008; Koehn, 2009) assists human transla- tion to improve translation efficiency (Green et al., 2014, 2015; Herbig et al., 2020). Even though there is a quality gap between the outputs of machine ∗indicates equal contribution. †indicates the corresponding author. 1https://github.com/ZhenYangIACAS/WeTS 2https://www.statmt.org/wmt22/ translation-suggestion-task.html translation (MT) systems and manual translations by professional translators, MT can still practically reduce time in comparison with translating from scratch (Läubli et al., 2013). Later, the advances at the sequence-to-sequence model (Sutskever et al., 2014; Bahdanau et al., 2014; Vaswani et al., 2017) further made a breakthrough in translation tech- nology, inspiring the industry to transform human translation into computer-aided translation (CAT) (Knowles and Koehn, 2016; Santy et al., 2019) to a great extent. CAT usually relies on an MT en- gine and a platform with a user-friendly interface (Bowker, 2002; Lengyel et al., 2004; Bowker and Fisher, 2010; Bowker, 2014; Pal et al., 2016; Chat- terjee, 2019), with which humans perform post- editing (PE) on machine translations to achieve final results with quality standards. In the past, the post-editing process was typi- cally static and machines would no longer respond to humans’ modifications once humans started post- editing. Recent works (Domingo et al., 2016; González-Rubio et al., 2016; Peris et al., 2017) investigate interactive protocols and algorithms so that humans and machines can collaborate and ma- chines automatically refine the translations accord- ing to the human’s edits. …

Content truncated to 500 words