Hierarchical Multi-task learning framework for Isometric-Speech Language Translation

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 22 - 31 May 26-27, 2022 c⃝2022 Association for Computational Linguistics Simultaneous Neural Machine Translation with Prefix Alignment Yasumasa Kano and Katsuhito Sudoh and Satoshi Nakamura Nara Institute of Science and Technology, Japan kano.yasumasa.kw4@is.naist.jp Abstract Simultaneous translation is a task that requires starting translation before the speaker has fin- ished speaking, so we face a trade-off between latency and accuracy. In this work, we focus on prefix-to-prefix translation and propose a method to extract alignment between bilingual prefix pairs. We use the alignment to segment a streaming input and fine-tune a translation model. The proposed method demonstrated higher BLEU than those of baselines in low la- tency ranges in our experiments on the IWSLT simultaneous translation benchmark. 1 Introduction Simultaneous machine translation (SimulMT) is a task to start outputting translation before observ- ing the whole input sentence. SimulMT is more difficult than the translation with the whole input sentence because it cannot use the latter part of the sentence as context. SimulMT has to decide whether to wait for more input or to output partial translation using the input so far, in real-time. The translation quality should become better if we can use longer inputs and vice versa. We have to han- dle such a trade-off between the quality and latency of the translation by decision policies to choose the next action between read (waiting for the next input segment) and write (outputting a translation segment) for a given input-output history (Gu et al., 2017). Neural Machine Translation (NMT) models used for SimulMT can be roughly categorized into policy-dependent and policy-independent. A policy-dependent model is trained with the constraints given by the policy, in order to trans- late an input prefix into an output prefix. Ma et al. (2019) proposed a simple method with a fixed pol- icy called wait-k, where the NMT first takes k read actions followed by alternating write and read ac- tions until the end of the translation output. Ari- vazhagan et al. (2019) proposed a joint training framework for flexible policies and the correspond- ing NMT model using a latency-augmented loss function and Monotonic Infinite Lookback (MILk) attention. In contrast, a policy-independent model is a standard NMT model to translate the whole in- put into the whole output and used for SimulMT along with a given policy in the inference. We can share one NMT model for different policies, so the quality-latency trade-off can be controlled easily. Dalvi et al. (2018) achieved some latency reduction with a small loss in BLEU by the use of a fixed policy called STATIC-RW. Ma et al. (2019) also applied their wait-k policy using a sentence- based NMT model, called test-time wait-k. Zhang et al. (2020) proposed a flexible policy to predict segment boundaries in an input. Once a bound- ary is found, the segment is translated using a sentence-based NMT model. The model based on their segmentation demonstrated better results in quality-latency trade-off than those …

Content truncated to 500 words