Towards Stream Translation: Adaptive Computation Time for Simultaneous Machine Translation

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 228–236 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 228 Towards Stream Translation: Adaptive Computation Time for Simultaneous Machine Translation Felix Schneider Karlsruhe Institute of Technology felix.schneider@kit.edu Alexander Waibel Karlsruhe Institute of Technology alexander.waibel@kit.edu Abstract Simultaneous machine translation systems rely on a policy to schedule read and write op- erations in order to begin translating a source sentence before it is complete. In this paper, we demonstrate the use of Adaptive Compu- tation Time (ACT) as an adaptive, learned pol- icy for simultaneous machine translation using the transformer model and as a more numer- ically stable alternative to Monotonic Infinite Lookback Attention (MILk). We achieve state- of-the-art results in terms of latency-quality tradeoffs. We also propose a method to use our model on unsegmented input, i. e. without sentence boundaries, simulating the condition of translating output from automatic speech recognition. We present first benchmark re- sults on this task. 1 Introduction Simultaneous machine translation (MT) must ac- complish two tasks: First, it must deliver correct translations on incomplete input as early as possi- ble, i. e. before the source sentence is completely spoken. Second, in a realistic usage scenario, it must deal with unsegmented input, either speech directly or automatic transcriptions without punc- tuation or sentence boundaries. Until now, staged models (Niehues et al., 2016), which have a sep- arate component to insert punctuation (Cho et al., 2012) achieved the best results in this task. In this paper, we will present the first step towards an end- to-end approach. In recent years, a number of approaches for neu- ral simultaneous machine translation have been proposed. They generally build on the com- mon encoder-decoder framework (Sutskever et al., 2014), with the decoder deciding at each step whether to output a target language token based on the currently available information (WRITE) or to wait for one more encoder step in order to have more information available (READ). In order to do this, the decoder relies on a wait policy. The published policies can be broadly di- vided into two categories: • Fixed policies, which rely on pre-programmed rules to schedule the read and write operations, such as wait-k (Ma et al., 2019a) and wait-if (Cho and Esipova, 2016). • Learned policies, which are trained either jointly with the translation model or sepa- rately. Examples include MILk (Arivazha- gan et al., 2019) and the models of Satija and Pineau (2016) and Alinejad et al. (2018) However, all of the above approaches train and evaluate their models on individual sentences. We want to work towards a translation system that can work on a continuous stream of input, such as text without punctuation and sentence segmentation. In a realistic usage scenario, segmentation informa- tion is not available and an end-to-end solution without a separate segmentation component is de- sirable. We therefore propose the use of Adaptive Computation Time (Graves, 2016) for simultane- ous machine translation. This method achieves a better latency-quality trade-off …

Content truncated to 500 words