The University of Edinburgh’s Submission to the IWSLT21 Simultaneous Translation Task

Proceedings of the 18th International Conference on Spoken Language Translation, pages 46–51 Bangkok, Thailand (Online), August 5–6, 2021. ©2021 Association for Computational Linguistics 46 The University of Edinburgh’s Submission to the IWSLT21 Simultaneous Translation Task Sukanta Sen, Ulrich Germann and Barry Haddow University of Edinburgh {ssen, ugermann, bhaddow}@inf.ed.ac.uk Abstract We describe our submission to the IWSLT 2021 shared task1 on simultaneous text-to-text English-German translation. Our system is based on the re-translation approach where the agent re-translates the whole source pre- fix each time it receives a new source token. This approach has the advantage of being able to use a standard neural machine translation (NMT) inference engine with beam search, however, there is a risk that incompatibility be- tween successive re-translations will degrade the output. To improve the quality of the translations, we experiment with various ap- proaches: we use a fixed size wait at the begin- ning of the sentence, we use a language model score to detect translatable units, and we apply dynamic masking to determine when the trans- lation is unstable. We find that a combination of dynamic masking and language model score obtains the best latency-quality trade-off. 1 Introduction In spoken language translation (SLT), there is often a need to produce translations simultaneously, with- out waiting for the speaker to finish. For example, we may be targeting live events such as conferences or meetings where excessive latency will disrupt the user experience. In order to achieve low la- tency SLT, however, translation systems must be able to cope well with incomplete utterances, and we find that we need to trade off latency for trans- lation quality. In research on simultaneous SLT, we would like to understand how to produce the best possible trade-off between these two measures. In the IWSLT 2021 shared task on simultaneous translation, the aim was to build and evaluate si- multaneous SLT systems at three different latency regimes (low, medium and high), as measured us- ing the Average Lagging (AL; Ma et al. (2019)). 1https://iwslt.org/2021/ There are two main approaches to simultaneous translation: streaming (Cho and Esipova, 2016; Ma et al., 2019) where the system appends the output to a growing hypothesis as new inputs are avail- able, and re-translation (Niehues et al., 2016, 2018; Arivazhagan et al., 2020a,b), where, as the name suggests, the system re-translates the whole prefix on every update to a completely new output. Re- translation approach has the advantage that we can use an unmodified, general purpose, optimised MT engine with beam-search, but we have to address the problem of flicker. That is to say, the translation of a prefix may be changed by the translation of an extended prefix. Recent work by Arivazhagan et al. (2020a) has shown that, if measures are taken to mitigate flicker, then re-translation produces re- sults comparable to streaming approach. Since the shared task does not permit any revision of a com- mitted hypothesis (i.e. flicker is not allowed) we focus on adapting the re-translation approach for our submission without introducing …

Content truncated to 500 words