BiTIIMT: A Bilingual Text-infilling Method for Interactive Machine Translation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1958 - 1969 May 22-27, 2022 c⃝2022 Association for Computational Linguistics BiTIIMT: A Bilingual Text Infilling Method for Interactive Machine Translation Yanling Xiao1∗ Lemao Liu2† Guoping Huang2 Qu Cui2 Shujian Huang1† Shuming Shi2 Jiajun Chen1 1National Key Laboratory for Novel Software Technology, Nanjing University, China 1Collaborative Innovation Center of Novel Software Technology and Industrialization, Nanjing 2Tencent AI Lab, China 1xiaoyl@smail.nju.edu.cn, {huangsj,chenjj}@nju.edu.cn 2{redmondliu, donkeyhuang, qucui, shumingshi}@tencent.com Abstract Interactive neural machine translation (INMT) is able to guarantee high-quality translations by taking human interactions into account. Ex- isting IMT systems relying on lexical con- strained decoding (LCD) enable humans to translate in a flexible translation manner be- yond left-to-right. However, they typically suf- fer from limitations in translation efficiency and quality due to the reliance on LCD. In this work, we propose a novel BiTIIMT system, Bilingual Text-Infilling for Interactive Neural Machine Translation. The key idea to BiTI- IMT is the Bilingual Text-infilling (BiTI) task which aims to fill missing segments in a manu- ally revised translation for a given source sen- tence. We propose a simple yet effective so- lution by casting this task as a sequence-to- sequence task. The benefits of our solution are that it performs efficient decoding with the same complexity as the standard decod- ing in NMT and makes full use of revised words for better translation prediction. Experi- ment results show that BiTiIMT performs sig- nificantly better and faster than state-of-the-art LCD-based IMT on three translation tasks. 1 Introduction Recent years have witnessed significant develop- ment in neural machine translation(NMT) (Bah- danau et al., 2015; Vaswani et al., 2017). Despite their success, their translation in quality still can not meet the requirements in industrial applications. On the other hand, interactive neural machine trans- lation (IMT) (Foster et al., 1997; Langlais et al., 2000; Simard et al., 2007; Barrachina et al., 2009; González-Rubio et al., 2013; Cheng et al., 2016; Weng et al., 2019; Huang et al., 2021) is able to guarantee high-quality translation: it is an iterative collaboration process between human and machine that involves multiple interactive steps to obtain a satisfactory translation. ∗Work was done during internship at Tencent AI Lab. †Corresponding authors. Traditional IMT generates a translation in the left-to-right completing paradigm (Langlais et al., 2000; Sanchis-Trilles et al., 2014; Peris et al., 2017a; Knowles and Koehn, 2016; Zhao et al., 2020) where human translators are required to re- vise words in the translation prefix. This strict left- to-right manner limits its flexibility because some human translators may enjoy their translation man- ners beyond left-to-right. As a result, another part of works (Weng et al., 2019; Huang et al., 2021) propose an alternative IMT paradigm under which human translators can revise words at arbitrary po- sitions of a translation. The essential technique to this paradigm is lexically constrained decoding (LCD) (Hokamp and Liu, 2017; Post and Vilar, 2018), which extends beam search in the decoding stage to include …

Content truncated to 500 words