Inria-ALMAnaCH at WMT 2022: Does Transcription Help Cross-Script Machine Translation?

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 233 - 243 December 7-8, 2022 ©2022 Association for Computational Linguistics Inria-ALMAnaCH at the WMT 2022 shared task: Does Transcription Help Cross-Script Machine Translation? Jesujoba O. Alabi1∗ Lydia Nishimwe2 Benjamin Muller2,3 Camille Rey2 Benoît Sagot2 Rachel Bawden2 1Spoken Language Systems (LSV), Saarland University, Saarland Informatics Campus, Germany 2Inria, Paris, France jalabi@lsv.uni-saarland.de3Sorbonne Université, France firstname.lastname@inria.fr Abstract This paper describes the Inria ALMAnaCH team submission to the WMT 2022 general translation shared task. Participating in the lan- guage directions {cs,ru,uk}→en and cs↔uk, we experiment with the use of a dedicated Latin-script transcription convention aimed at representing all Slavic languages involved in a way that maximises character- and word-level correspondences between them as well as with the English language. Our hypothesis was that bringing the source and target language closer could have a positive impact on machine translation results. We provide multiple com- parisons, including bilingual and multilingual baselines, with and without transcription. Ini- tial results indicate that the transcription strat- egy was not successful, resulting in lower re- sults than baselines. We nevertheless submitted our multilingual, transcribed models as our pri- mary systems, and in this paper provide some indications as to why we got these negative results. 1 Introduction This paper describes the Inria ALMAnaCH team submission to the WMT 2022 general translation shared task. We chose to explore the language directions {cs,ru,uk}↔en and cs↔uk in order to concentrate on the Slavic language family. Due to some experimental problems that impacted the into- Slavic directions most heavily, we only submitted {cs,ru,uk}→en and cs↔uk language directions, but we present all results we obtained here. A major area of interest in machine translation (MT) research is transfer between languages, par- ticularly related ones and for lesser resourced lan- guages (Zoph et al., 2016; Kocmi and Bojar, 2018). One way of encouraging transfer is to train multilin- gual models, whereby several language directions are trained simultaneously, often sharing some (Fi- rat et al., 2016) or all model parameters (Ha et al., ∗Contributions made whilst at Inria. 2016; Johnson et al., 2017; Aharoni et al., 2019), with the hope that similarities between the lan- guages can boost performance, particularly for the lower-resourced languages. To encourage lexical sharing and therefore the transfer capacity of such models, joint subword seg- mentation models (Sennrich et al., 2016b) and MT vocabularies are often used (Sennrich et al., 2016a), and techniques such as phonetisation and translit- eration/transcription can be applied to texts in a bid to overcome differences in writing systems and spelling (Nguyen and Chiang, 2017; Chakravarthi et al., 2019; Goyal et al., 2020; Muller et al., 2021). In our submission to the WMT 2022 general translation shared task, we experimented with mul- tilingual models and the use of customised tran- scription into a common writing system designed to maximise lexical sharing, similar to the one used in (Muller et al., 2021). We choose to work with the language directions involving Slavic languages, that is {cs,ru,uk}↔en and cs↔uk. We …

Content truncated to 500 words