摘 要

目前，随着科学技术和信息技术的飞速发展，以及全球经济竞争和跨文化交流愈加频繁，翻译工具的发展需求日益迫切。此外，由于高素质翻译人才缺乏， 而且精良的翻译成果需要大量的人力和时间，因此，十分有必要开发更有效的机器翻译工具，以更好地交流互通各国之间的信息。近年来，随着机器翻译技术的发展，机器翻译译文变得更加合理，但仍让人感受到僵化的机器翻译风格。而在

2016 年，谷歌发布全新机器翻译系统，即谷歌神经机器翻译（GNMT）系统。与流行的信念“机器翻译永远无法赶上人类翻译”相反，如今 GNMT 正经历着前所未有的改变，特别对科技类文本来说，GNMT 译文可能比人工翻译更快更准确。

但是，作为机器翻译的共同弱点，神经网络翻译技术目前并不是绝对正确， 还会犯些人工译员轻易能看出的毫无逻辑的错误，这些问题都有望在未来靠更先进的技术纠正。简单具体的来说，本文将着重分析 GNMT 系统产生的错误类型， 旨在能够提高译后编辑效率，并希望为机器翻译系统的革新提供灵感。

关键词：神经网络翻译；错误分析；译后编辑；深度学习

Contents

1. Introduction 1
2. Demonstration 2
   1. Description of research task and process 2
   2. Characteristics of scientific English 3
   3. Types of prominent GNMT errors 4
   4. Application of GNMT system in translation tasks 6
      1. Words 6
      2. Phrases 8
         1. Prepositional phrase 8
         2. Verb phrase 9
      3. Sentences 10
         1. Passive sentence 10
         2. Subordinate clause 10

2.5 Significance of NMT error analysis 11

1. Conclusion 12

Notes 13

Works Cited 14

Bibliography 15

Error Analysis of Machine Translation based on GNMT system

Introduction

MT technology, by definition, is a sub-field of computational linguistics that investigates the use of software to translate text or speech from one language to another. The origin of MT tools can be traced back in the 17th century when a few of scholars came up with compiling mechanical dictionaries. However, not until 1940s that MT technology came into public eyes. At that time, W.Weaver and A.D.Booth put forward the suggestion to apply computer science into translation. Furthermore, it was in 1949 that W.Weaver officially advanced the idea of machine translation in the memorandum titled Translation.(Fu Sihan 2017) However, he regarded MT as a matter of mechanical decoding, regardless of other complex rules of linguistics, syntax and semantic meaning. After that, MT technology had gone through its ups and downs before the prosperous development in the 20th century. During that time, many MT tools had been launched, using intermediary resources to conduct indirect translation. One great breakthrough took place in 1990s when the research personnel of IBM Peter Brown and others proposed IBM model, introducing statistical MT models to the industry in which the phrase-based and syntax-based method became popular and spread to all parts of the world. (Du Jinhua 2013)

Nevertheless, the glorious period of statistical MT models has become the history. Google Neural Machine Translation (GNMT) system, launched in 2016, revolutionized the entire MT industry. It is a promising advanced approach with great potential of handling many disadvantages of traditional MT systems in that it is based on end-to-end learning method and possesses numerous data and machine memory. It combined deep learning method with MT technology, trying and making progress in changing the stiffen translation output by traditional MT tools. With the help of GNMT system, present-day translators conveniently preprocess the material that needs to be translated. This is the common way for today’s translators to process abstruse original texts, using machine

translation tools to preprocess original texts and then transfer the output to human post- editing.