Studying The Impact Of Document-level Context On Simultaneous Neural Machine Translation

Studying The Impact Of Document-level Context On Simultaneous Neural Machine Translation Raj Dabre raj.dabre@nict.go.jp National Institute of Information and Communications Technology Aizhan Imankulova aizhan.imankulova@cogsmart-global.com CogSmart Masahiro Kaneko masahiro.kaneko@nlp.c.titech.ac.jp Tokyo Institute of Technology Abstract In a real-time simultaneous translation setting, neural machine translation (NMT) models start generating target language tokens from incomplete source language sentences, making them harder to translate, leading to poor translation quality. Previous research has shown that document-level NMT, comprising of sentence and context encoders and a decoder, leverages context from neighbouring sentences and helps improve translation quality. In simultaneous translation settings, the context from previous sentences should be even more critical. To this end, in this paper, we propose wait-k simultaneous document-level NMT where we keep the context encoder as it is and replace the source sentence encoder and target language decoder with their wait-k equivalents. We experiment with low and high resource settings using the Asian Language Treebank (ALT) and OpenSubtitles2018 corpora, where we observe minor improvements in translation quality. We then perform an analysis of the translations obtained using our models by focusing on sentences that should benefit from the context where we found out that the model does, in fact, benefit from context but is unable to effectively leverage it, es- pecially in a low-resource setting. This shows that there is a need for further innovation in the way useful context is identified and leveraged. 1 Introduction Neural machine translation (NMT) (Bahdanau et al., 2015; Luong et al., 2016) is an end-to-end approach known to give the state of the art results for a variety of language pairs. In stan- dard NMT, the entire source language sentence is fed to the model, and once the entire target language sentence is generated, it is presented to the user. However, in a real-time translation setting, translation models are expected to present translated words or phrases as they are gen- erated. Furthermore, waiting for the entire source language sentence adds to the latency, and therefore an optimal solution is to have a model that can start generating target language words right after the first few source language words are available for translation. This is known as simultaneous NMT (SNMT) and is known for its poor translation quality, especially in low- resource settings. The concept of waiting for k words or tokens before generating target lan- guage words or tokens is known as wait-k SNMT (Ma et al., 2019). In this paper, we work with the Transformer architecture as the standard NMT model, consisting of a bidirectional en- coder and unidirectional decoder. The decoder is able to attend to all source language tokens when generating target language tokens. However, in the case of the wait-k SNMT model, Proceedings of the 18th Biennial Machine Translation Summit Virtual USA, August 16 - 20, 2021, Volume 1: MT Research Track Page 202 the standard encoder and decoder are replaced with their SNMT equivalents, which are a unidi- rectional encoder and a modified decoder, respectively. The decoder can only look at i …

Content truncated to 500 words