PRHLT’s Submission to WLAC 2022

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1182 - 1186 December 7-8, 2022 ©2022 Association for Computational Linguistics PRHLT’s Submission to WLAC 2022 Ángel Navarro and Miguel Domingo and Francisco Casacuberta PRHLT Research Center Universitat Politècnica de València {annamar8,midobal,fcn}@prhlt.upv.es Abstract This paper describes our submission to the Word-Level AutoCompletion Shared Task of the WMT 2022. We participate in the pair of languages, English–German, in both ways. We propose a segment-based interactive ma- chine translation approach whose central core is a machine translation (MT) model that pre- dicts the complete translation from the context provided by the task and picks the word we were trying to autocomplete from there. We show with this approach that it is possible to use the MT models in the autocompletion task by performing minor changes at the decoding step and obtaining good accuracy. 1 Introduction Machine translation (MT) has significantly im- proved in recent years with the emergence of neural machine translation (NMT), but it still cannot as- sure high-quality translations for all tasks (Toral, 2020). For those scenarios with rigorous translation quality requirements, it is critical for professional translators to manually validate the translations gen- erated by the NMT system. The computer-aided translation (CAT) tools show up to improve the vali- dation and editing process carried out by translators. Researchers approached CAT tools from many di- rections with the aim of reducing the human effort of correcting the automatic translations. Among CAT tools such as translation memory (Zetzche, 2007), augmented translation (Lommel, 2018) and terminology management (Verplaetse and Lam- brechts, 2019); we can find autocompletion tools, which help professional translators by providing new partial translations according to the validated parts they have supplied to the system. Word level autocompletion (WLAC) (Li et al., 2021) is a new shared task introduced in WMT22. Its aim is to complete a target word given a source sentence, a sequence of characters typed by the human translator and a translation context. Four types of context are possible: Zero-contex: no context is given. Suffix: a sequence of translated words located af- ter the word to autocomplete. Prefix: a sequence of translated words located prior to the word to autocomplete. Bi-contex: A combination of the suffix and the prefix type. That is, there is a sequence of translated words located after the word to auto- complete, and a sequence of translated words located prior to the word to autocomplete. Note that, in all cases, the word to autocomplete is not necessarily consecutive to these contexts. We have experimented with a similar CAT tool from the interactive machine translation (IMT) framework. In this field of research, the transla- tion is generated in a collaborative process between the human translator and the MT model. Among the different approaches, the segment-based IMT (Domingo et al., 2017; Peris et al., 2017) proto- col presents certain similarities with WLAC: at each step, the user validates sequences of translated words—the context—and makes a correction—the word to autocomplete. Therefore, in this work we have …

Content truncated to 500 words