Machine Trans EN 4

Machine Translation - A challenge or a chance for human translators?

Overview Page of Machine Translation

30 Chapters（0/30)

Machine_Trans_EN_1 Machine_Trans_EN_2 Machine_Trans_EN_3 Machine_Trans_EN_4 Machine_Trans_EN_5 Machine_Trans_EN_6 Machine_Trans_EN_7 Machine_Trans_EN_8 Machine_Trans_EN_9 Machine_Trans_EN_10 Machine_Trans_EN_11 Machine_Trans_EN_12 Machine_Trans_EN_13 Machine_Trans_EN_14 Machine_Trans_EN_15 Machine_Trans_EN_16 Machine_Trans_EN_17 Machine_Trans_EN_18 Machine_Trans_EN_19 Machine_Trans_EN_20 Machine_Trans_EN_21 Machine_Trans_EN_22 Machine_Trans_EN_23 Machine_Trans_EN_24 Machine_Trans_EN_25 Machine_Trans_EN_26 Machine_Trans_EN_27 Machine_Trans_EN_28 Machine_Trans_EN_29 Machine_Trans_EN_30 ...

Back to translation project overview

To the To Do list

4 王李菲(A Comparison Between Machine Translation of Netease and Traditional Human Translation——A Case Study of The Economists Articles) Machine_Trans_EN_4

Abstract

Machine translation is a subfield of artificial intelligence and natural language processing that investigates transforming the source language into the target language. On this basis, the emergence of neural machine translation, a new method based on sequence-to-sequence model, improves the quality and accuracy of translation to a new level. As one of the earliest companies to invest in machine translation in China, Netease launched neural machine translation in 2017, which adopts the unique structure of neural network to encode sentences, imitates the working mechanism of human brain, and generates a translation which is more professional and more in line with the target language context. This paper takes the excerpts of some articles in The Economist as the corpus for analysis, and aims to explore the types and causes of common errors, as well as the advantages and challenges of each, through the comparative analysis of Netease neural machine translation and human translation, and finally to forecast the future development trend and make a conclusion of this paper.

Key words

Machine Translation; Human Translation; Contrastive Analysis

题目

网易有道机器翻译与人工翻译的译文对比——以经济学人语料为例

摘要

机器翻译研究将源语言所表达的语义自动转换为目标语言的相同语义，是人工智能和自然语言处理的重要研究分区。在此基础上，一种基于序列到序列模型的全新机器翻译方法——神经机器翻译的出现让译文的质量和准确度提升到了新的层次。网易作为国内最早投身机器翻译的公司之一，在2017年上线的神经网络翻译采用了独到的神经网络结构，模仿人脑的工作机制对句子进行编码，生成的译文更具专业性，也更符合目的语语境。本文以经济学人内的文章选段为分析语料，旨在通过对网易神经机器翻译和人工翻译的英汉译文进行对比分析，探究常见错误类型及生成原因，以及各自存在的优势与挑战，最后展望未来发展趋势，并对本文做出总结。

关键词

机器翻译；人工翻译；对比分析

1. Introduction

Nowadays, economic globalization has accelerated overwhelmingly, and considerable resources are poured into the business field. As a branch of global language English, business English is proposed under the theoretical framework of English for Specific Purpose (ESP). It serves international business activities, which is a professional subject requiring specialized English. (Jordan, 1997) As the medium that helps people with different cultural backgrounds to understand each other, business translation is required to be “formal, accurate, standardized and smooth”, which challenges both the machine translation and human translator.

With the urgent requirement for more precise and higher quality translation, recent years have witnessed the rapid development of neural machine translation (NMT), which has replaced traditional statistical machine translation (SMT) to become a new mainstream technique, playing a crucial part in many fields, like business, academic, and industry. Compared with SMT, the NMT model is more like an organism. The model has many parameters that can be adjusted and optimized for the same goal, making the combination and interaction more organic and the overall translation effect better, which fantastically matched the demands of business translation.

The Economist is an international news and Business Weekly offering clear coverage, commentary, and analysis of global politics, business, finance, science, and technology. Highly professional and formal articles put forward a tricky problem to machine translation and human translators. Given this, this paper makes a comparative analysis of Netease neural machine translation and human translation, aiming to explore the types and causes of common errors, as well as the advantages and challenges of each. In the end, this paper will forecast the future trends, hoping to promote the development of translation studies in China.

2. The Development Process of Machine Translation

Since the IBM model was put forward by the researcher Peter Brown in the early 1990s, statistical methods have gradually become the mainstream of machine translation research（杜金华,2013:1）. This method has extensively promoted the development of machine translation technology. After a variety of statistical machine translation models have emerged, such as the phrase-based translation model, hierarchical phrase translation model, and syntactic translation model, the translation quality has been dramatically improved.

Since 2002, with the technical maturity and stability of statistical machine translation, especially phrase-based machine translation, statistical machine translation technology has been making solid strides towards practical and commercial application. Therefore, with the rapid development of technology, people have gradually built-up confidence in machine translation, and the social demand for machine translation has been increasing day by day, with higher and higher expectations.

However, from the perspective of academic research, both phrase-based translation models and syntactic translation models have experienced a rapid development stage, and the existing theoretical methods and technical models have begun to show "bottlenecks" in the improvement of translation performance. (罗季美，2012：84） In addition, from the perspective of industrialization and utilitarianism, there is an urgent need for a more practical machine translation system, but the gap between the results of machine translation and the requirements of human beings is still very large. Therefore, for the researcher, while excited to see the score of machine translation system evaluation is getting higher and higher, and the performance of online machine translation systems developed by Google, Baidu, Netease, and other enterprises is developing with each passing day, they are facing more and more challenges.

Aiming to solve these problems, many technological giants are striving to find a new way to improve both the quality and efficiency of machine translation. There was a breakthrough which bought machine translation to a new level. Since 2014, the sequence-to-sequence neural machine translation has developed rapidly, compared with the statistical machine translation, the translation quality received a significant boost（李亚超，2018:2734）.

The previous statistical machine translation was more like a mechanical system. Each module has its own function and goal and then outputs the translation results through mechanical splicing. Its primary disadvantage is that the model contains low syntactic and semantic components, so it will encounter problems when dealing with languages with significant syntactic differences, such as Chinese-English. Sometimes the result is unreadable even though it is “word-for-word”.

On the contrary, neural machine translation is consisted of several components, including phrase conditions, partial conditions, sequential conditions, primitive models, and so on. Its core is deep learning of artificial intelligence which can imitate the working mechanism of human brain and adopt a unique neural network structure to model the whole process of translation. The entire model is composed of a large number of “neurons,” and each “neuron” has to complete some simple tasks, and then through the combination of all of them to coordinate the work, a much better translation text appears. （李亚超，2018：2735）

Since neural machine translation puts more emphasis on context and the whole text, it produces more coherent and comprehensible content to readers, and be widely accepted and used in various fields in a concise time. In 2017, at the GMIC (Global Mobile Internet Congress), Duan Yitao, the chief scientist of Netease, delivered a keynote speech titled “Machine Translation has Its Own Way” and announced a piece of exciting news: the neural machine translation technology independently developed by Netease has been officially launched. This technology was launched by Youdao (a subsidiary brand of Netease) this time has been jointly developed by it and Netease Hangzhou Research Institute for over two years. It will serve Youdao Dictionary, Youdao Translator, Youdao Web version, Youdao E-reader, and other applications, expecting to bring super-convenient product experience to users. In addition, the Youdao Translation officer also launched photo translation, by which users only need to take pictures of the text, then the results of neural network translation will be shown in real-time. As a pioneer of machine translation in China, the development process of Netease YouDao is precisely a paradigm of the history of machine translation in China. Therefore, in this paper, the neural machine translation technology developed by Netease will be compared with human translators. The same excerpts selected from The Economist are translated by both of them, then the different versions will be analyzed by the translation criterion so as to figure out their respective strengths and weaknesses, bringing considerations to the current translation situation and references to future development.

3.Comparative Analysis of Errors in English-Chinese Translation

Based on the introduction above, neural machine translation has been widely used in the business field. However, because of the professional characteristics, there are still numerous human translators devoted themselves to business translation, especially in some important international meetings. In order to be faithful to the style of the original text, when translating, it is necessary to use some conventional terms, professional jargon, and fixed sentence patterns, striving to be accurate and standardized, and achieving the authenticity of the style of writing. These kinds of characteristics make the neural machine translation and human translator highly comparable. Guided by the translation criterion— “faithfulness, expressiveness, and closeness” put forward by the prominent translator Liu Zhongde, （方梦之，2004）the detailed comparative analysis of the same texts translated by Netease neural machine translation and human translator respectively will be presented to show their own strengthens and weaknesses.

3.1 Precision—the selection of vocabulary

Business English is usually related to formal economic activities, such as insurance, contracting, and documentation, which require the words to be precise and professional. A huge number of terminologies plus the polysemy contained in the texts, so the premise of translation is to figure which word fits the context most. In The Economist, there are many articles contain words which created by the authors, or cannot be explained smoothly with its common meaning. Because the author are all masters of various fields, so they may cite many allusions, new hot words, and memes to complement their articles or create a humorous effect. This requires the translator to accurately convey the information by choosing the vocabularies related to the field according to the context and relevant knowledge.

Excerpt 1: Investors are salivating: during the past month the German giant’s share price has surged by 60% while Tesla’s has slipped. That is mainly because of a change of heart about which of the two will win the electric-vehicle (ev) contest. Investors have, it seems, caught "The Loev Bug".

Netease neural machine translation: 投资者对此垂涎三尺:在过去的一个月里，这家德国巨头的股价飙升了60%，而特斯拉的股价却在下滑。这主要是因为两家公司在电动车竞赛中谁将胜出的问题上改变了想法。投资者似乎染上了“勒夫虫”(Loev Bug)。

Human translator: 投资者开始对大众倾注期望，在过去一个月，这家德国巨头的股价飙升了60%，而特斯拉的股价在下滑。这主要是因为观众对于谁能赢得这场电动汽车（EV）竞赛改变了看法。投资者似乎已情迷“万能大众电动汽车”。

In this sentence, “caught ‘The Loev Bug’” is translated into “染上了‘勒夫虫’”by Netease neural machine translation. However, the human translator translates it into “已情迷’万能大众电动汽车’”.

Analysis: There is a difference in the translation of the word “The Loev Bug”. As mentioned above, in The Economist, the author may use some memes in their articles which require the readers command relevant knowledge so as to understand them smoothly. In this sentence, “The Loev Bug” is a translation of the English name of the Disney film “Herbie: ‘The Love Bug’”, “ve” in “love” is changed to “ev” (the English abbreviation of electric vehicle), implying the connection between Volkswagen Group's electric vehicles and The Love Bug.

In these two translation versions, guided by the translation criterion “faithfulness, expressiveness, and closeness” put forward by the translator Liu Zhongde, the version translated by human translator is more suitable. On the one hand, the human translator reveals the true meaning of the phrase “The Loev Bug”, which is more consistent to the original text as well as more comprehensive to the readers. On the other hand, “caught ‘The Loev Bug’” are translated to “情迷‘万能大众电动汽车’”, which is more appropriate to the context, achieving the author’s ideal effects.

Therefore, the version of human translator is better in selecting the proper words.

Excerpt 2: H&M, a retailer, is asking for rent holidays, hurting commercial-property firms.

Netease neural machine translation: 零售商H&M要求出租假期，这损害了商业地产公司的利益。

Human translator: 零售商H&M正在申请免租期，这无疑会损害商业地产开发商的利益。

In this sentence, “rent holidays’” are translated into “出租假期”by Netease neural machine translation. However, the human translator translates it into “免租期”.

Analysis: There is a difference in the translation of the word “rent holidays”. This is a common word which has been wildly used in business field, which means that there is no rent for a specific period. However, the machine translation just translates this phrase word by word, and ignores the context, causing a reading disorder.

One of the translation ideals proposed by Liu Zhongde ── “faithfulness” required that the meaning of the target text should be close to that of original one. The version of machine translation selects the inappropriate word, so when the readers see such information, the acquisition is not accurate enough, so their responses will be different. It is inevitable that there will be certain differences between the communication parties when receiving information, resulting in inaccurate communication and less than expected feedback.

3.2 Standardization—the perfection of logicality

Except for vocabulary, some distinctions also exist in the syntax of Chinese and English. The expression used by English speakers and Chinese speakers is quite different although they are delivering the same attitude towards the same thing.

For example, when they try to write down a passage, English speakers tend to choose more long sentences which are joined together with various tense, conjunctions, and clauses so that they are more logical and coherent. On the contrary, Chinese speakers prefer to use a series of short sentences, the structure of which is usually loose and conjunctions are rarely used. To sum up, English is a hypotactic language, and all the sentences constitute the grape-like structure; Chinese is a paratactic language and mostly a bamboo-like structure.

Such kind of contrast exists universally in business translation. If the translator fails to notice the syntactic characteristic and logicality of the sentence, the translation will become broken and divert from the original meaning.

Excerpt: Volkswagen has recently started to do the same with its new id.3 ev, and has set out to become the biggest computing company in Germany after sap, a maker of business software. But whereas Tesla has tech in its dna, Volkswagen’s most notable software achievement was to facilitate the ignominious emissions-cheating. Autonomous driving will add a whole new realm of complexity. Mr. Diess himself acknowledges that as software increasingly turns the car into a computer on wheels, it will transform the industry even more than electrification does.

Netease neural machine translation: 大众汽车最近也开始用它的新身份做同样的事情。它已经开始着手成为德国仅次于商业软件制造商sap的最大计算机公司。但特斯拉的DNA中有技术，而大众最显著的软件成就是促进了可耻的排放作弊。自动驾驶将增加一个全新的复杂性领域。迪斯自己也承认，随着软件越来越多地把汽车变成一台带轮子的电脑，它将比电气化更能改变整个行业。

Human translator: 大众最近开始在它的新款ID.3电动汽车上尝试这件事，并立志成为德国仅次于商业软件制造商SAP的计算公司。然而相比DNA里刻着科技的特斯拉，大众在软件上最突出的成就却是为可耻的排放作弊提供便利。自动驾驶又会让复杂程度上升到一个全新的维度。迪斯本人也承认，软件把汽车变得越来越像是带轮子的计算机，这一点对汽车工业的改造将比电气化更加深

Analysis: Both of the machine and human translator choose to divide the first sentence into two or more short sentences, which is more in line with the writing habit of Chinese. The machine translation version splits this long sentence into two sentences, and use the character "它" to introduce the other part, which is more understandable.

However, machine translation can only achieve peer-to-peer translation of individual phrases, and it fails to express the logicality between sentences then confuse the readers. For example, in the version of machine translation, the phrase “But whereas Tesla has tech in its dna” is translated into “但特斯拉的DNA中有技术”. Then there exists two subjects in this sentence, namely, “特斯拉” and “大众”, which cause the incoherence of sentence meaning, as well as reading difficulties. On the contrary, the human translator use the translation technique addition, adding “相比” to the text, so the logicality is more clear, and the expression is also very standard. What’s more, the only one subject emphasize the point of this sentence, giving the readers a quick overview of this paragraph.

In addition, the last sentence is translated into “随着软件越来越多地把汽车变成一台带轮子的电脑，它将比电气化更能改变整个行业。”by translation machine, then cause an ambiguity. In this sentence, the character “它” can be seen as the pronoun of “软件”, “电气化”as well as the whole sentence, affecting the expressive effect of this sentence, then the readers may wonder who “will transform the industry even more than electrification does” indeed. For those working in the business field, simple words and clear expressions are preferred. For expressions which have of the same meaning, using simple and clear expressions saves both writing and reading time. In the version of human translator, instead of using “它”, the translator chooses the phrase “这一点”, which directly refers to the whole thing, that is “software increasingly turns the car into a computer on wheels”.

One of the translation ideals proposed by Liu Zhongde ── “expressiveness” required that the target text should be coherent and clear, there being no need to follow the exact order of words and sentences structure of the original language but reorganize and elaborate to respect the rules of target language. In this excerpt, the human translator pays more attention to the logicality and coherence, using different translation techniques to polish the text, so the original meaning is guaranteed to the maximum extent. At same time, this version is more consistent to the features of business English as well as the Chinese writing habit.

3.3 Appropriateness—the consistency of style

Besides the accuracy of words and fitness of sentence patterns analyzed above, whether the translator can catch the style of the entire passage mainly determine the quality of translation. Whether the translation manages to be professional, whether the illocutionary meaning is presented to readers, and whether the four maxims of cooperation principles are observed strictly, all these tricky problems should be considered in the whole discourse.

For example, western people tend to tell the results at the beginning, then list the reasons why they made such a decision. But from an eastern perspective, they are used to implicating their idea and saying things the long way round. As a middleman, the translator should search for a suitable style for the discourse, giving full consideration to pragmatic principles in the target language.

Excerpt: He was wandering in a ricefield of dreams. The plants were tall as sorghum, taller than a man. Their panicles hung full as brooms, and each grain was as big as a peanut. After walking a while, he lay down in the leaf-shade with a friend, quite hidden. A rest was a good idea, because the wonder-plants went on and on. In fact, they covered the world.

Netease neural machine translation: 他徘徊在梦的稻田里。这些植物像高粱一样高，比人还高。它们的圆锥花序像扫帚一样挂满，每一颗谷粒都有花生那么大。走了一会儿，他和一个很隐蔽的朋友一起躺在树荫下。休息一下是个好主意，因为神奇的植物不停地生长。事实上，它们覆盖了整个世界。

Human translator: 他漫步在梦中的稻田里。水稻长得有高粱那么高，一人多高。穗子沉甸甸的，有扫帚那么长，籽粒有花生米那么大。走了一会儿，他和朋友躺在稻叶的阴凉下，被遮得严严实实。休息一下也无妨，因为神奇水稻会继续生长。事实上，它们已覆盖全球。

Analysis: This excerpt is selected form an article which commemorates the death of Yuan Longping, the father of hybrid rice. Though it is published in The Economist, this article not only contains Yuan Longping’s contribution to agriculture, but also some graceful description of his daily life. Therefore, when translating, the translator should pay more attention to the style of the entire passage.

In this excerpt, the author creates a wonderful dreamland where the vigorous plants researched by Yuan Longping are grow healthily. Since the original text is poetic and polished, the target text should restore this kind of beauty. However, in the version of machine translation, “它们的圆锥花序像扫帚一样挂满”，“他和一个很隐蔽的朋友一起躺在树荫下”，and “休息一下是个好主意”, these sentences are neither accord with style of the original text, nor with the writing and reading habits of Chinese. In contrast, the human translator use some short sentences to describe the idyllic view, like “ 穗子沉甸甸的，有扫帚那么长。” and“他和朋友躺在稻叶的阴凉下，被遮得严严实实。”， which is more picturesque and smooth. The use of some concrete adjectives like “沉甸甸”and “严严实实”allowing readers to feel more about the density of rice, feeling like they’re in the wonderland.

One of the translation ideals proposed by Liu Zhongde ── “closeness” required that the style of target text should be consistent with the original text. In this excerpt, the human translator gets rid of the construction of English, then restructure the content in a Chinese tone. At same time, this version is closer to the original style, and more readable.

4.1 Machine Translation

As a new paradigm, neural network method represents the mainstream of current machine translation methods. In recent years, especially since 2014, relevant achievements have been constantly emerging, and great success has been achieved. At present, the following three trends are in the ascendant. It is believed that with the progress of technology and in-depth research, machine translation methods will definitely develop vigorously in the future.

4.1.1 Network Integration will Deepen

In the modern society, data types and modes are diverse. It is difficult to deal with massive complicated data in the future only by the text-oriented circular neural network which based on attention currently. Different networks have different advantages. For example, cyclic neural network is good at processing time series data, convolutional neural network is suitable for processing high latitude data, and recursive neural network is effective in processing tree and graph structure data（侯强 2019:33）. Therefore, how to use parallel or serial methods to optimize the combination of different networks, and develop a multi-mode machine translation model suitable for processing heterogeneous data, so that it can be used in real scenes, and even achieve the degree of intelligence, is a subject to be deepened.

4.1.2 Processing Moves toward Parallelism

Traditional neural machine translation is a serial-to-sequence model based on recurrent neural networks, which is difficult to process in parallel and slow in training and decoding. In contrast, serial-to-sequence model, self-attentional Transformer model and non-autoregressive model based on convolutional neural network all have better parallel processing ability. During training, their training data and network parameters can be updated synchronously, and encoders and decoders can also work in parallel, thus effectively reducing the time complexity. Parallel processing can accelerate the speed of training process, as well as that of encoding and decoding, and greatly improve the translation efficiency, so it is an important research and development direction in the future.

4.1.3 Training Methods will be more Varied

Classical neural machine translation model training is highly dependent on bilingual parallel corpora. However, except for some vertical fields in high-resource languages such as Chinese and English, most languages in the world lack large-scale, high-quality and extensive parallel corpora, especially low-resource languages. Therefore, how to use reinforcement learning, transfer training, dual learning, joint training, confrontation learning and other training methods to alleviate the lack of corpus resources is also one of the key exploration objects in the future. Training methods have a great influence on the performance of the model, and the training methods vary over time, if all of them can take advantages from each other and avoid the disadvantages, it is expected to further improve the overall performance of the model.

4.2 Human Translator

On the basis of the above, it can be concluded that there are many difficulties in business English translation that machine translation couldn’t overcome yet. Human translators still have their competitiveness in achieving equivalent translation in the business field. Therefore, what attitudes and steps should be taken in future has become much more important currently.

4.2.1 Holding an Appropriate Attitude

Machine translation has been developing for 54 years since its birth, and it has been able to develop independently as a complete artificial intelligence technology. It has now become a discipline and technology comparable to human translation. Therefore, whatever concerns language learning may have about the development of machine translation, it is inexorable the normal development of machine translation and should face it calmly with a proper attitude. As language workers, when machine translation has a certain impact on the employment of the industry, they should constantly "recharge" themselves, improve their ability, and actively treat their translation career.

4.2.2 Cooperating with Machine Translation

As Darwin described in The Origin of Species, “The fittest survive by natural selection and the unfit are eliminated.” Since language workers want to avoid being replaced by machine translation, they must improve their professional skills to survive in this highly mechanized environment. Therefore, in the future, due to the continuous development of artificial intelligence technology, machine translation is bound to move to a new peak. In order to make translation more efficient and applicable, human translators need to make good use of this opportunity to combine their skills with the rapid development of science and technology, and try to form a new situation of "human-machine coupling".

5. Conclusion

Through the above dialectical analysis of some cases in the translation process, it is easy to find that the readability of the version of manual translation is still higher than that of neural machine translation referred to the translation criterion proposed by Liu Zhongde. Though machine translation has an unsurmountable superiority in speed because it can translate a sea of texts in a moment, distinguishing the nuance between different words which may all seem suitable to the original text, and selecting the most appropriate sentence patterns to perfect the logic are the core competitiveness of human translators, so they will not be replaced for a long time, especially in some specific fields which require professional knowledge and life experience.

However, the advancements of machine translation are obvious to all, the quality of translation texts has been improving. With the development and application of new technology, it has always been an important research topic in the field of artificial intelligence and more and more resources has been poured to this field. After the ups and downs of the past 70 years, it has entered a period of rapid development. How to improve the existing neural machine translation model and build a new model with more accurate translation, so as to make a qualitative leap in the overall translation quality is an urgent task at present. It is believed that with the further development of artificial intelligence, machine translation will grow stronger and play a more important role in overcoming natural language barriers and promoting transnational communication in the future.

References

[1] Papineni, K., Roukos, S., Ward, T. & Zhu, Weijing. (2002) BLEU: a Method for Automatic Evaluation of Machine Translation, Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), Philadelphia,311-18.

[2]杜金华,张萌,宗成庆,孙乐.中国机器翻译研究的机遇与挑战——第八届全国机器翻译研讨会总结与展望[J].中文信息学报,2013,27(04):1-8.

[3]方菲. 商务英语机器翻译汉译错误分析[D].广东外语外贸大学,2017.

[4]胡开宝,李翼.机器翻译特征及其与人工翻译关系的研究[J].中国翻译,2016,v.37;No.239(05):10-14.

[5]侯强,侯瑞丽.机器翻译方法研究与发展综述[J].计算机工程与应用,2019,v.55;No.929(10):30-35+66.

[6]罗季美,李梅.机器翻译译文错误分析[J].中国翻译,2012,v.33;No.215(05):84-89.

[7]李亚超,熊德意,张民.神经机器翻译综述[J].计算机学报,2018,v.41;No.432(12):2734-2755.

[8]杨南. 基于神经网络学习的统计机器翻译研究[D].中国科学技术大学,2014.