RETRACTED: Application of Neural Network Machine Translation in College Translation Teaching

This article has been retracted by the authors.


Introduction
On July 20 th , 2017, the State Council of China published Development Planning for Artificial Intelligence in the New Stage, which laid out the development blueprint of artificial intelligence for the next 20 years. Latest progress in different fields of artificial intelligence, i.e. voice recognition, face recognition and machine simultaneous translation, have thus been massively reported by domestic media, while in the field of language service and translation studies, neural network machine translation also became one of the most significant topics under discussion. It was reported that neural network machine translation has already approached the translation level of human, which, combined with voice recognition technology, is competent in simultaneous translation. Under these circumstances, great concerns have been shown in such issues as to whether machine translation will soon replace human translators, whether graduates of translation major will face the risk of unemployment, and what should be taught in translation teaching and how to teach translation if machine can translate any language.
This paper aims to analyze the features of machine translation by reviewing the development of machine translation from technical perspective, especially machine translation based on neural network, and the challenges they face, followed by a discussion on the relationship between machine translation and translation teaching.

2
Review of Machine Translation Studies

Original intention
Electronic computer was designed originally to realize quick computation, which was soon endowed with the ability of symbol processing. In March, 1947, in a letter written by American scientist Warren Weaver to cybermetics expert Norbert Wiener, the concept of using computer to realize language translation was firstly mentioned, which was known as the famous translation memorandum [1]. The original drive in study of machine translation can be traced back to the code decoding between Russian and English during World War II, when people simply fed some translation rules and bilingual dictionaries into computer to realize machine translation [2]. However, the machine translation system constructed in this way functioned unfavorably. In 1966, American Automatic Language Processing Advisory Committee carried out a comprehensive assessment of machine translation between Russian and English, which revealed that machine translation studies cost too much while the translation quality turned out not satisfactory, and concluded that machine translation could never possibly achieve the level of human translators. Ever since then, machine translation studies met a grave recession.
After entering 21 st century, with the acceleration of economic and cultural globalization, the demands on translation showed a dramatic increase. The high demand in the translation market thus became the drive for machine translation studies. Compared with human translation, machine translation is quicker in time and lower in price. More and more practitioners in translation industry began to take machine translation as an aid to improve translation efficiency and quality.

Mechanisms of machine translation
Different mechanisms have been applied to machine translation in different stages of machine translation studies. At first, rule-based machine translation was applied, which was soon replaced by example-based machine translation [3]. Then the 90s of last century has witnessed the booming of statistical machine translation. In the past 5 years, neural network machine translation has become a hotspot in translation studies.
The above-mentioned different mechanisms applied in machine translation studies correspond with the different stages when different methodologies are observed to realize automatic translation. In the first stage, the language-rule-based method is adopted, which originates from an intuitive understanding of translation and holds that machine translation can be realized as long as the translation theories and methods are fed into computers. However, the language rules cannot cover all the complicated linguistic phenomena, which becomes a bottle-neck for the summary of knowledge. After the rationalistic method, machine translation enters into the second stage when the empirical method is adopted. Example-based machine translation or statistical machine translation and neural network machine translation, all aim to learn translation from the actual translation practices with the aid of computers. Though these iJET -Vol. 14, No. 19, 2019 three methods belong to the domain of empiricism, they differ from each other in the following aspects: Example-based machine translation, by finding most translation examples similar to the materials to be translated, tries to observe them to complete the actual translation; Statistical machine translation, by dividing sentences into smaller linguistic units (mainly words and phrases), tries to find the most possible translation equivalences (translation pattern) and then combine them into larger translation equivalences; Neural network machine translation adopts the end-to-end comprehensive translation mode. The development in the three empirical research methods demonstrates the route chart of reforms in machine translation strategies: from imitating human's translation rules to analyzing and reconstructing linguistic components, and then to adopting a comprehensive translation strategy.

Breakthrough in deep learning
Studies on artificial neural network were started as early as in the 40s of last century. In 1969, Minsky & Papert reported that neural network represented by sensor can only solve the first order predicate logic, unable to process even the simplest exclusive OR logic [4], which dragged the study of neural network into great recession for a long time. In August, 2011, Microsoft announced that the voice recognition system constructed by neural network could reduce the mistake rate by 33% in the real-time voice recognition, which, defined as a significant breakthrough, re-aroused researchers' passion for neural network studies. Deep learning is a multi-layer complicated neural network structure, where, besides the input layer and output layer, there are a lot of hidden layers with the structure of network nodes in continuous change. Artificial neural network realizes the functional simulation of human's intelligence through structural simulation.
Although deep neural network has made remarkable achievements in the fields of voice recognition and image recognition, difficulties have been met in the field of natural language processing. As late as in 2016, Google introduced the deep-learningbased multi-lingual neural network machine translation system (abbreviated hereafter as GNMT), which, compared with its previous phrase-based statistical machine translation system, could reduce the mistake rate in translation by 60% [5], approximately reaching human's translation level.
At present, with machine translation studies focusing on deep learning pattern, many researchers have made systematic studies in neural network machine translation [6] [7]. Compared with statistical machine translation, is neural network machine translation really better? In which aspects are the advantages shown? Why are there these advantages? What challenges will be faced? Solution to all these questions will help to further improve the performance of neural network machine translation and determine the ceiling machine translation could possibly reach in the future.

3
Principle of Neural Network Machine Translation

Theoretical base of neural network learning
Artificial neural network uses the physically realizable system to simulate the structure and function of nerve cells in human brain. Activation functions are used to simulate different status of human neurons by artificial neurons which, when connected with each other, form neural network structure. The learning process of neural network is a process to continuously adjust the values of network connections. The main learning method is called mistake-correcting learning, i.e. the deviation between the expected output and actual output of neural network is taken as the reference for connection and the values are modified towards the direction for an ultimate reduction of deviation. Therefore, neural network requires data training with labeled answers. Once the network parameters are converged, it means the network has already fitted the training data. Only the well-trained neural network can process new data. Figure 1 demonstrates the structure chart of multi-layer sensor, the basis of neural network. The so-called deep neural network can be simply explained as there are a lot of layers in the middle-hidden layer. Besides, the structure of neural network nodes is also in continuous development, for example, the Long Short-term Memory Networks (abbreviated hereafter as LSTM) contains the complicated structures of three controlling gates, which can effectively solve the problem of long-distance dependence [8]. LSTM is nowadays frequently used in machine translation because a large amount of context should be referred to in the practice of translation.

Resources of deep learning
The outstanding performance of deep neural network machine translation is largely dependent on big data, deep network pattern and complicated calculating.
Firstly, the performance of neural network machine translation system can only be guaranteed by a large scale of training data. According to the study of translation system from English to Spanish by Koehn & Knowles in 2017, only when the scale of training corpus is above 15 million words, can the performance of neural network machine translation system be superior to that of statistical machine translation, which means that a training corpus of just several millions words is not big enough to support a favorable neural network machine translation system. However, to acquire dozens of millions of manually labeled data, its price is even higher than that of the system itself.
Secondly, the deep neural network structure must contain enough hidden layers to conduct effective learning. For example, the neural network of GNMT adopts 8-layer encoding LSTM and 8-layer decoding LSTM. However, it is not the case that more layers guarantee better performance. Complicated network structure brings a lot of problems in pattern training as well as dramatic increase in training time and calculating cost, while there is no obvious promotion in network performance. At present, there is no theory to support how many hidden layers and hidden layer nodes there should be in the network structure to achieve the optimal performance, which is still under the relentless experiments of researchers.
Thirdly, neural network machine translation relies on complicated calculating. Personal computers can no longer fulfill the training task of deep learning pattern. In order to acquire the translation pattern extracted from 36 million translation pairs between English and French, GNMT adopts the most advanced NVIDIA K80 GPU, whose floating-point arithmetic speed can achieve thousands of billion times per second, to conduct the training on 12 machines (there are 96 GPUs in total) for 6 whole days. Therefore, it requires powerful hardware support to acquire the translation pattern of deep learning.

Features and Challenges of Neural Network Machine Translation
It has already been verified in many researches that neural network machine translation is much better than phrase-based statistical machine translation [9][10] [11], where the languages involved include English, French, German and Spanish. Compared with the translation between English and French or German, which belong to the same language family, the translation between English and Chinese is more difficult.

Features of neural network machine translation
Apart from the general features of machine translation, neural network machine translation has also manifested the following features: Remarkable promotion in translation fluency: Evaluation of translation fluency is made with reference to such factors as the spelling, diction, syntax and semantics of translation [12]. Neural network machine translation can remarkably promote the readability of its translation, because it not only handles such syntactic problems as subject-predicate agreement, past participle, double object and complement in a better way, but also dramatically reduces the mistakes of word order, morphological change and functional words. However, the phrase-based statistical machine translation separates the relation between phrases and gives rise to the disfluency of translation by making the independence assumption on the relation between different phrases in a sentence, which is caused by the defect of translation pattern itself. On the other hand, neural network machine translation adopts the end-to-end translation pattern by inputting the whole sentence into the network during translation training, which guarantees a complete translation at the output end. Through tremendous training in translation practice, the neural network, based on complicated calculating, can realize the optimal translation to the natural sentences.
Realization of zero-shot translation: GNMT can realize the translation among dozens of languages with one translation pattern, which means, deep learning can realize transfer learning and thus zero-shot translation between two languages that have never been trained directly by translation corpus [13]. For example, researchers firstly use the translation corpus from Portuguese to English to acquire the translation pattern between Portuguese and English, and then use the translation corpus from English to Spanish to acquire the translation pattern between English and Spanish. Though there is no translation corpus from Portuguese to Spanish, the existing translation pattern can translate Portuguese into Spanish in a favorable way. Realization of multi-lingual translation by sharing one pattern can dramatically reduce the amount of work in multi-lingual translation studies and make it possible for automatic translation between languages which lack training data.
Mysterious translation mechanism: For rule-based machine translation, its translation mechanism is transparent and the translation result can be explained and for statistical machine translation, the translation result is also explainable because it always chooses the translation with the maximal probability. However, for neural network machine translation, its translation result is hard to explain, because it simulates the structure of nerve system in human brain with a different information processing mechanism. The value finally saved between neural network nodes is tightly related to the training data and network structure, which makes the translation mechanism of neural network obscure and hard to explain.

Challenges of neural network machine translation
Although neural network machine translation has made remarkable progress, it has not yet achieved human's translation level. In 2017, Isabelle et al. constructed a translation challenge set between English and French to test GNMT, the result of which showed that the correctness rate of GNMT translation was less than 70% [7]. In the following, the challenges of neural network translation will be discussed in four aspects: Applicability in different literary fields: Different literary fields feature different literary styles and require appropriate translation techniques. When tested by the materials from the same fields with training corpus, machine translation can usually show favorable performance in translation results, which, however, will fall dramatically when the translation materials come from a different field. Concerning the applicability of literary fields, neural network translation is even worse than statistical translation. Besides, machine translation can perform very well in translating materials from such literary fields as news, science, patent and law, which are based on description of facts, knowledge and information, but much worse in translating materials from literary fields such as novels and poetry, which contain a lot of subjective comments on emotions and opinions.
Creativity in translation: It is generally admitted that one machine translation system can only output one optimal translation to the same translation material. Irrespective of the number of times a machine translation system is operated, its outcome of translation to the same material always stays unchanged. Machine translation can neither generate diversified correct translations, nor tell the subtle variations between different expressions. The outcome of machine translation is related to the method its developer adopts, but irrelevant to its materials to be translated. Unlike human translators, machine translation system has no creativity in itself.
Error tolerance and stability: Error tolerance refers to the ability to modify the minor mistakes in the original materials and provide the correct translation. Unlike human translators who can naturally correct the mistakes in the original materials, machine translation system merely translates whatever is inputted, no matter whether it is correct or not. Stability refers to whether the performance of translation can stay unchanged in the process of translating different materials. The stability of human translation is mainly dependent on the linguistic ability and knowledge scope of translator, while that of machine translation is quite unpredictable, which is proven by the fact that in the practice machine translation may conduct a very good translation to one sentence, but generate a much poorer version to another one of similar structure.

Machine Translation and College Translation Teaching
The development of machine translation will not deprive human translators of their working opportunities, and high-level translators are still in short supply. Therefore, college translation teaching, under the background of artificial intelligence era, needs the aid of technology to cultivate qualified translators. In practice, college translation teaching can be reformed in the following aspects:

Advocacy of coordinated translation between humans and machine
In college translation teaching students should be well acquainted with the features of machine translation and the division of labor between humans and machine. For example, concerning the terminologies in science and patent translation, machine can help translators to keep the consistency and accuracy of terminologies, while in the translation of literary works, human should always take the dominant role. Machine translation and human translation are not contradictory to each other but supplement and promote each other [14]. College students should be taught how to construct an efficient coordinated environment between human and machine translation.

Introduction of post-translation editing
The popularization of machine translation has created a brand-new language service industry: post-translation editing, which refers to the manual reviewing and editing after translation so as to improve the quality and appropriateness of machine translation. Since the mode of post-translation editing has been applied in the translation industry, translation teaching in universities and colleges should also add posttranslation editing course so as to expand the new field of translation service and cultivate professional post-translation editors [15].

Cultivation of software application ability
Traditional college translation teaching focused on the teaching of translation techniques, while nowadays, machine translation, computer-aided translation and translation management system provide the manual translation with more efficient translation tools. Therefore, courses should be added to introduce how to apply related software and tools in translation teaching. Competence in applying translation software has gradually become one of the necessary skills for high-end translators.

Conclusion
Neural network machine translation uses deep neural network to realize the end-toend translation mode, which remarkably promotes the fluency and accuracy of translation, but its translation mechanism is harder to explain. Although the machine translation based on deep learning has made substantial progress, it does not mean that neural network is the only way to address issues concerning machine translation. With the continuous development in technology, the quality of machine translation will keep improving in the future. However, there is still an obvious gap between neural network machine translation and that of human translators. Machine translation is competent in literal translation, but also shows its weakness in translating the thoughts, emotions and opinions behind the words, where only the human translators can show the unparalleled advantages.
Machine translation will not replace human translators as a translation technology means, but create new service field in the future. College translation teaching should also follow the improvement of technology in order to use them to reform the teaching content and teaching modes.