NEURAL MACHINE TRANSLATION BY JOINTLY LEARNING TO ALIGN AND TRANSLATE

通过标齐和翻译的联合学习实现神经机器翻译

ABSTRACT

Neural machinetranslation is a recently proposed approach to machine translation. Unlike thetraditional statistical machine translation, the neural machine translationaims at building a single neural network that can be jointly tuned to maximizethe translation performance. The models proposed recently for neural machinetranslation often belong to a family of encoder–decoders and encode a sourcesentence into a fixed-length vector from which a decoder generates atranslation. In this paper, we conjecture that the use of a fixed-length vectoris a bottleneck in improving the performance of this basic encoder–decoderarchitecture, and propose to extend this by allowing a model to automatically(soft-)search for parts of a source sentence that are relevant to predicting atarget word, without having to form these parts as a hard segment explicitly.With this new approach, we achieve a translation performance comparable to theexisting state-of-the-art phrase-based system on the task of English-to-Frenchtranslation. Furthermore, qualitative analysis reveals that the(soft-)alignments found by the model agree well with our intuition.

神经机器翻译是最近提出的机器译方法。和传统的统计机器翻译不同，神经机器翻译目的是，建立一个单一的神经网络可以联合地改良以最大化翻译性能。神经机器翻译的模型刚刚提出来，基本上还是属于编码和解码的一族，也是属于解码一个来源句子到一个固定长度的向量，从中解码产生一个翻译。在本文中。我们推测运用固定长度的向量的是一个改进这种基本的编码和解码结构的瓶颈，据此，提出了一个拓展的模型允许该模型自动的软性的搜索来源句子的部分来预测目标词，而没有形成这些部分作为一个硬性的明显的划分。你用这种新的方法，我们在英语-法语的翻译方面实现了可以和最佳方法媲美的方法。更进一步的。定量分析显示模型创造的软性的对齐和我们的直觉一致。

硬性划分和软性搜索，这是Bengio的最新的大作。