The failure of the AI-powered simultaneous translation system from Tencent, a Chinese internet giant, during the Boao Forum for Asia 2018 drew a lot of attention from the IT industry. It was the first time that machine translation (MT) was implemented in this forum. However, the gibberish spouted by the system has indicated once again that MT cannot easily replace human translators.
From the perspective of the global competitive landscape, China has become one of the most important players in the AI field and this wave of AI is actively promoted by the Chinese government. In July 2017, the State Council of China issued “A Next Generation Artificial Intelligence Development Plan” stating that “by 2020, the overall technology and application of AI will be in step with globally advanced levels, the AI industry will have become a new important economic growth point…” In the world of translation, AI also has the potential to reshape this sector.
International tech giants like Google, Microsoft, IBM and Amazon are all keen on developing MT. Especially after Google launched its neural machine translation (NMT) system in 2016, the improved quality of MT has gained more attention and interest from all sides globally. In China, using huge-crowd strategy, Baidu, one of the premier AI leaders in the world, recruited thousands of translators to translate English brochures, letters, technical manuals and other materials into Chinese for 15 hours a day. The mass translation event lasted for one month in December 2016, and helped Baidu obtain a large amount of parallel translation data. In the same year, Tencent also unveiled its first AI-assisted translation software, Mr. Translator.
In 2017, Sogou, one of China’s top three search engines, made a vast investment in UTH International, a Chinese information technology company that claims to have established “the world’s largest Multilingual Big Data Platform” to enhance the performance of their MT engines. NetEase Youdao has been studying MT for almost ten years and applied NMT to all of its translation products, reaching over 700 million users.
Since the application scenarios and markets for MT are very broad, a lot of companies and organizations in China are making attempts to build their own MT systems. Most of them focus on the economic, financial, patent and legal sectors. Very few of them have touched upon the medical and pharmaceutical industry, perhaps because of its high difficulty in translation and high requirement for information security. To the best of my knowledge Atman, an AI startup, is the only Chinese company that virtually works on MT solutions for the medical and pharmaceutical industry. They have already launched their own MT system, TransGod. In 2017, Atman signed a contract with Johnson & Johnson to privately deploy and custom-develop an MT system for them. Using the user’s own high-quality data to build a private MT system can improve translation efficiency and ensure the safety of user data translation.
Apart from IT companies, Chinese academics are also paying attention to MT. Northeastern University, Soochow University, Chinese Academy of Sciences, Harbin Institute of Technology, Tsinghua University, Fudan University and Xiamen University all have well-known MT resource and development groups. The China Workshop on Machine Translation (CWMT) will have been held for 14 successive years this October. Futhermore, Google has stated that MT errors between English and Chinese have been reduced by 60% through the use of their NMT compared to their previous phrase-based system.
Regarding the quality evaluation of MT, researchers have contributed a lot in this field. Meanwhile, end users are also conducting evaluations by themselves. In October 2016, an analyst at an investment bank and a translator, Michael Zhang, wrote an article about Google Translate (GT) on his blog. He tried GT with the same source text on two different dates: March 2016, when GT was using phrase-based MT, and in October 2016, when GT applied NMT. He compared the translation output of GT on the two dates with human translation output, concluding that the quality of MT from English to Chinese had a significant improvement after the introduction of NMT, reflected in two aspects: reduced errors in word order, and better treatment of terminology and punctuation. However, he also pointed out that omissions and consistency of terminology was still a problem.
In March 2017, another translator, Chaoxiong Li, tested the MT engines from four companies: Google, Baidu, Youdao and NewTranx. He used the MT engines to translate a short Chinese text into Russian. Results showed that NewTranx performed the best. The other three had problems in word order, omission and expression to varying extents.
A comparison of MT apps in China
To gain a deeper understanding of MT use in China, a comparison of six popular MT apps is conducted here. They are Google Translate, Microsoft Translate, Baidu Translate, Youdao Translate, Sogou Translate and Mr. Translator (Tencent Translate). With the exception of Google Translate and Microsoft Translate, all others are developed by Chinese local companies. In fact, the web page of Google Translate is always accessible in mainland China, despite other Google sites being blocked. However, the mobile app of Google Translate only became accessible in mainland China in March 2017, when the app was updated to version 5.8.
Table 1 shows a comparison of the six translation apps in China. It has to be noted that the web version and mobile version of the same translation app may have a few differences. For example, Mr. Translator can support 15 languages for the web version, but only six for the mobile version. In addition, the number of languages supported for different modes of translation may vary as in the instant camera translation that usually supports fewer languages than text translation. For example, Google Translate supports over 100 languages via text translation, 37 via photo, 32 via voice in “conversation mode” and 27 via real-time video in “augmented reality mode.” In this study, the mobile version of the six apps is compared. “Number of languages supported” refers to text translation.
All six MT apps are free to use, but users need to pay for the API service. According to Table 1, we can see that the apps have all applied NMT to their systems. Voice translation and instant camera translation are both available in all apps. Google Translate and Youdao Translate support the most languages. Mr. Translator can only support six languages, but an important feature is that it supports Chinese-English “simultaneous interpretation,” which means speakers can see real-time translation on the screen while speaking to the app.
As two relatively young apps, Sogou Translate does not support offline translation and Mr. Translator only has a free trial offer for Chinese-English offline translation. Figure 1 shows the mobile interface of the six apps from the latest iOS version in April 2018.
In Figure 1, Microsoft Translator stands out due to its dark color. Its interface is simple and straightforward, with four bubbles floating in the air, representing text translation, voice translation, dialogue translation and instant camera translation. The interfaces of the other five all have a white background. Under the input box of Baidu Translate, Youdao Translate and Sogou Translate, there are articles to help people learn language and culture. The novelty of the dialogue style interface of Mr. Translator may be favored by young people. When the user opens the app for the first time, Mr. Translator automatically sends a video guide on how to use the app. Another highlight is that Mr. Translator is not simply a translation app, it also aims to help users learn languages. For example, there is an icon of a headphone beside the dialogue bubble, which means users can read the English translation and the system will mark their pronunciation.
Users hold different opinions on the quality of these apps. Shiyibao (www.shiyibao.com), an automatic grading system of translated text, regularly holds contests between human translated text and machine translated text. The rules are as follows: a piece of news is selected from the Financial Times, an experienced translator provides the human translated text, while some MT tools provide the machine translated texts. Participants select the text that they believe to be the human translated one. If they select the correct one, human translation receives one star, if they are wrong, machine translation receives one star. Up to November 30, 2017, Shiyibao has held 19 contests and 353 people have participated in the contests. Results show, perhaps surprisingly, that human translation received 892 stars and machine translation received 1,061 stars. More specifically, Youdao Translate and Google Translate both received 290 stars, Baidu Translate received 192 stars, Sogou Translate received 173 stars and Bing Microsoft Translator received 116 stars.
A comparison of Chinese portable translation devices
Following the trend of developing MT apps, Chinese IT companies are racing to launch portable translation devices. iFLYTEK was the first to dip its toes into the water and introduced Xiaoyi Translator to the world in their annual conference in 2016. Other rivals are stepping up to get into the game. As a result, in 2017, Baidu, NetEase and Babel Technologies all managed to bring their products to the market. In 2018, Sogou Travel Translator also put its finger in the pie. Not far behind, Xiaomi, the Chinese smartphone startup, just launched its AI Translator that supports 14 languages and continues the company’s usual style of budget products with a surprisingly aggressive price of 299 yuan (approximately 48 dollars). Table 2 briefly shows a comparison of five Chinese portable translation devices.
We can see that all five portable translation devices adopt NMT. Most devices support over ten languages. Among them, AIcorrect Translator supports the most with 39 languages. Xiaoyi supports only six languages with the most expensive price of 2,999 yuan. AIcorrect, Youdao and Sogou are equipped with touch screens, which can display text synchronously. This feature helps users correct the text if the voice recognition is inaccurate, so as to obtain better translation output. With regard to offline translation, Xiaoyi and Sogou Travel Translator are the only two that have this feature, and both of them support ZH-EN only. As the name suggests, Baidu Wi-Fi translator works both as a Wi-Fi hot spot device and a translator. With a built-in automatic internet connection configuration, Baidu Wi-Fi translator can be connected without a SIM card or Wi-Fi. The translator has been in cooperation with operators across over 80 countries so far. Besides, it has incorporated the popular concept of sharing and launched a rental online service in collaboration with Ctrip, a Chinese provider of travel services. The rental price is 29 yuan (approximately $4.60) per day, almost the same as that of a pocket Wi-Fi device. Except for Youdao Translator Egg, all the other four devices cost more than 1,000 yuan. At almost the same price as a middle-end smartphone and overlapping functions with mobile translation apps, the market for portable translation devices remains a question. For most Chinese people, traveling abroad is not something they do often. Compared to buying a portable translation device that may be rarely used, the rental mode seems to be more flexible and feasible. Meanwhile, iFLYTEK reported that since the launch of Xiaoyi Translator in March 2017, it has sold more than 200,000 units, covering over 130 countries, indicating the translator has already gained market recognition.
Only two translators, AIcorrect and Sogou, are equipped with instant camera translation. Sogou Travel Translator has a big screen and looks like a smartphone. It should be noted that Sogou has launched different products for different consumers. While they target businessmen with Sogou Travel Translator, Sogou focuses on students with another product, Sogou Transcription Pen, which can turn recordings into text and realize simultaneous interpretation at an affordable price of 299 yuan. As with Sogou, iFLYTEK also has turned attention to these two groups. While Xiaoyi is aiming at businessmen, YiBei Portable Translator is designed just for students with more colors available, a cute shape and cheaper price (1,199 yuan). It has some special features for learning English, such as shadowing exercises, dialogue practice and numerous embedded teaching materials. In April 2018, iFLYTEK rolled out Xiaoyi Translator 2.0, which costs the same as the previous version. The major upgrade is the new translator now supports 33 languages and several Chinese dialects. Besides, it has a screen and is equipped with instant camera translation. It is rumored that Youdao Translator Egg is also going to be updated to its 2.0 version with more languages and features soon.
Attitude of end users
There is no doubt that Chinese IT giants have made significant strides in machine translation. However, is the attitude of language service providers toward MT improving in pace with it? To figure this out, four Chinese translation companies were randomly selected: Tianyi Times, LocaTran, Transn and Easy Translation. On their websites, all clearly state computer-assisted translation (CAT) tools are included in their workflow. However, whether they are using machine translation remains unknown. On the website of another Chinese translation company, Ecyti Translation, there is a sentence stating that “we reject the use of machine translation or other online translation systems.” Nevertheless, since many CAT tools have incorporated machine translation into their systems, the workflow tends to be more like translation memory + machine translation + post-editing now, so it is difficult to know if translators are using machine translation or not. In order to know more about the attitude of Chinese end users toward MT, a brief online investigation was carried out.
Sina Weibo, a Chinese micro-blogging website, is one of the most popular social media sites in China. Microblogs including the word “机器翻译” (meaning: machine translation) that were posted from Jan 1 to Dec 14 2017 on Sina Weibo were searched. For the purpose of the investigation, news and academic posts had been eliminated. The remaining 238 posts were analyzed for negative, positive and neutral comments. Figure 2 shows that the majority of the posts are negative (54%), 24% of the posts are positive and 22% of the posts are neutral.
Examples including the original post and its translation were:
Translation: Today in class, the teacher asked us to translate a business contract. I picked up my phone immediately and used Youdao Translate for instant camera translation. I thought I had to polish the translation output, but it turned out the translation looked so professional that I didn’t need to do anything. Compared to the Youdao Translate a few years ago, this year it had a qualitative upgrade. Translators may lose their job in the future; luckily I am not a translator.
Translation: The product description page on Amazon Shopping Overseas website is simply a disaster. Who says that machine translation can replace human?
Translation: The accuracy of translation is questionable. The sentences used in the video are quite simple, there is no problem to use them for traveling purpose. However, it may be hard to use machine translation for further communication. Because of its limited level, machine translation is still distant from human translation.
There are not a large number of microblogs about machine translation posted by end users on Sina Weibo (only 238 posts for almost the full year of 2017), indicating that machine translation may have not drawn much attention among end users in China, although a great quantity of news items on machine translation were posted on Sina Weibo. Results of the online investigation show that Chinese end users’ attitudes toward machine translation is predominantly negative, though it should be acknowledged that posts on social media are not 100% reliable. Nevertheless, it can be seen that some end users were really surprised about the high quality of MT, making them either happy about working more efficiently or worried about the future of translators.
It is a shame that there is a big gap between the fast-growing MT industry and the stagnant unawareness of end users. Based on my own experience, when asked about machine translation, some undergraduates in China even thought it was the same thing as an online dictionary. While focusing on the development of MT products, Chinese IT giants need to strengthen publicity and popularize knowledge to the end users. After all, as Peter G. W. Keen put it in his book Shaping the Future: Business Design Through Information Technology, “it is not the software but the human side of the implementation cycle that will block progress in seeing that delivered systems are used effectively.”
Ke Hu would like to give special thanks to Dr. Sharon O’Brien and Prof. Dorothy Kenny for their advice to this article. This work is supported by the Science Foundation Ireland (Grant No.: P31021).