End-to-End Simultaneous Translation System for IWSLT2020 Using Modality Agnostic Meta-Learning

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 62–68 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 62 End-to-End Simultaneous Translation System for the IWSLT2020 using Modality Agnostic Meta-Learning Houjeung Han, Mohd Abbas Zaidi, Sathish Indurthi, Nikhil Kumar Lakumarapu, Beomseok Lee, Sangha Kim Next AI Solution Lab, Samsung Research, Seoul, South Korea {h.j.han, abbas.zaidi, s.indurthi, n07.kumar, bsgunn.lee, sangha01.kim}@samsung.com Abstract In this paper, we describe end-to-end simul- taneous speech-to-text and text-to-text transla- tion systems submitted to IWSLT2020 online translation challenge. The systems are built by adding wait-k and meta-learning approaches to the Transformer architecture. The systems are evaluated on different latency regimes. The simultaneous text-to-text translation achieved a BLEU score of 26.38 compared to the com- petition baseline score of 14.17 on the low latency regime (Average latency ≤ 3). The simultaneous speech-to-text system improves the BLEU score by 7.7 points over the compe- tition baseline for the low latency regime (Av- erage Latency ≤ 1000). 1 Introduction Simultaneous Neural Machine Translation (SNMT) addresses the problem of live interpretation in ma- chine translation. In a traditional neural machine translation model, the encoder first reads the entire source sequence, and then the decoder generates the translated target sequence. On the other hand, a simultaneous neural machine translation model alternates between reading the source sequence and writing the target sequence using either a fixed or an adaptive policy. This would allow the model to avoid intolerable delay in live or streaming transla- tion scenarios. In this work, we build a simultaneous translation system for text-to-text(t2t) and speech-to-text(s2t) problems based on Transformer wait-k model (Ma et al., 2019a). We adopt the meta-learning approach presented in (Indurthi et al., 2020) to deal with the data scarcity issue in the speech-to-text translation task. The system architecture and data processing techniques are designed for the IWSLT 2020 on- line translation task (Ansari et al., 2020). However, these techniques can be applied to current and fu- ture SNMT models as well. We conduct several experiments on both text-to-text and speech-to-text problems to evaluate the proposed system. Our ex- perimental results reveal that the proposed system achieves significant performance gains over the pro- vided competition baselines on both the translation tasks. 2 Simultaneous Translation 2.1 Base Model The machine translation task involves converting an input sequence x = (x1, x2, . . . , xn), xi ∈ Rdx in the source language to the output sequence y = (y1, y2, . . . , yk), yt ∈ Rdy in the target lan- guage. In the simultaneous translation task, the model produces the output in an online fashion as the input is read. Hence, while producing an output yt, the complete input sequence might not have been processed . Our model derives from the transformer wait- k model proposed in (Ma et al., 2019a). Similar to (Ma et al., 2019b), the encoder consists of uni- directional transformer blocks unlike bi-directional transformer blocks used in the original transformer. The decoder starts producing the translation after having …

Content truncated to 500 words