From Simultaneous to Streaming Machine Translation by Leveraging Streaming History

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6972 - 6985 May 22-27, 2022 c⃝2022 Association for Computational Linguistics From Simultaneous to Streaming Machine Translation by Leveraging Streaming History Javier Iranzo-Sánchez and Jorge Civera and Alfons Juan Machine Learning and Language Processing Group Valencian Research Institute for Artificial Intelligence Universitat Politècnica de València Camí de Vera s/n, 46022 València, Spain {jairsan,jorcisai,ajuanci}@vrain.upv.es Abstract Simultaneous Machine Translation is the task of incrementally translating an input sentence before it is fully available. Currently, simul- taneous translation is carried out by translat- ing each sentence independently of the previ- ously translated text. More generally, Stream- ing MT can be understood as an extension of Simultaneous MT to the incremental transla- tion of a continuous input text stream. In this work, a state-of-the-art simultaneous sentence- level MT system is extended to the stream- ing setup by leveraging the streaming history. Extensive empirical results are reported on IWSLT Translation Tasks, showing that lever- aging the streaming history leads to significant quality gains. In particular, the proposed sys- tem proves to compare favorably to the best performing systems. 1 Introduction Simultaneous Machine Translation (MT) is the task of incrementally translating an input sentence be- fore it is fully available. Indeed, simultaneous MT can be naturally understood in the scenario of trans- lating a text stream as a result of an upstream Au- tomatic Speech Recognition (ASR) process. This setup defines a simultaneous Speech Translation (ST) scenario that is gaining momentum due to the vast number of industry applications that could be exploited based on this technology, from person-to- person communication to subtitling of audiovisual content, just to mention two main applications. These real-world streaming applications moti- vate us to move from simultaneous to streaming MT, understanding streaming MT as the task of simultaneously translating a potentially unbounded and unsegmented text stream. Streaming MT poses two main additional challenges over simultaneous MT. First, the MT system must be able to lever- age the streaming history beyond the sentence level both at training and inference time. Second, the system must work under latency constraints over the entire stream. With regard to exploiting streaming history, or more generally sentence context, it is worth men- tioning the significant amount of previous work in offline MT at sentence level (Tiedemann and Scherrer, 2017; Agrawal et al., 2018), document level (Scherrer et al., 2019; Ma et al., 2020a; Zheng et al., 2020b; Li et al., 2020; Maruf et al., 2021; Zhang et al., 2021), and in related areas such as lan- guage modelling (Dai et al., 2019) that has proved to lead to quality gains. Also, as reported in (Li et al., 2020), more robust ST systems can be trained by taking advantage of the context across sen- tence boundaries using a data augmentation strat- egy similar to the prefix training methods proposed in (Niehues et al., 2018; Ma et al., 2019). This data augmentation strategy was suspected to boost re-translation performance when compared …

Content truncated to 500 words