篇章神经机器翻译综述

A Survey of Document-level Neural Machine Translation

篇章机器翻译旨在使用计算机将一个篇章从一种语言自动翻译成另一种语言,是机器翻译中一项富有挑战性的任务。近年来,随着神经机器翻译的快速发展,篇章神经机器翻译成为了机器翻译研究的热门方向。研究者们提出了许多基于神经网络的篇章机器翻译模型,并取得了不错的效果。相比于传统句子神经机器翻译,篇章神经机器翻译通过建模并利用篇章级别的上下文信息来产生质量更高的译文。本文首先简单介绍了篇章翻译任务的定义和特点;其次分三个方面对篇章神经机器翻译现有研究进行了介绍:上下文建模、模型训练、模型分析;最后分析了篇章神经机器翻译研究当前面临的主要难点,并探讨未来可能的研究方向。

Document-level machine translation aims to use a computer to automatically translate a whole document from one language to another,which is a challenging task in machine translation.In recent years,with the rapid development of neural machine translation(NMT),document-level NMT has become one of hot research topics in the community of machine translation.Various document-level machine translation models based on neural networks have been proposed and achieved good results.By modeling document-level context information,document-level NMT models significantly outperform conventional sentence-level NMT models.In this paper,we first briefly introduce the definition and characteristics of document-level NMT.Then,we review the studies in three aspects:context modeling,model training and model analysis.Finally,we analyze the main problems currently faced by document-level NMT and discuss possible future research directions.