The development of machine learning has changed many industries, including the language industry. We ask questions like “will machine translation replace human translators” from time to time. But what machine learning can bring to the language industry is more than a question of whether human will be replaced or not. This essay will examine the technologies that greatly impact the language industry, what is being changed, and how industry stakeholders can better prepare themselves for the future. 

Machine Learning Technology Impacting the Language Industry 

The main technology that has already exerted a profound impact on the language industry is the natural language processing (NLP) technology. NLP is a vibrant interdisciplinary field with the goal of getting computers to perform useful tasks involving human language, tasks like enabling human-machine communication, improving human-human communication, or simply doing useful processing of text or speech. For a long time, we’ve been seeing examples of these tasks like conversational agents, grammar checker, and machine translation (MT). Although MT has been developing for over 60 years, it was not until recently that human feels the real threat because of the great progress in machine learning with many evidences showing the near-human quality of neuro machine translation (NMT) output. 

Besides NLP, machine learning can predict users’ preferences based on a large amount of data, enabling customized user experience, simpler and more automated workflow. All these contribute to the transformation of the language field by bringing in new players, more streamlined production process, larger user coverage, and smarter tools. 

Changing Landscape: New Players, Streamlined Workflow, and Smarter Tools 

The widely applied machine learning technology in the language field – machine translation (MT) – has penetrated into every corner of language service. Nowadays, not only language service providers are branding themselves as AI/MT solution provider, tech companies are leading the way of MT development. There are more MT vendors in the field: Google, Microsoft, Amazon, DeepL, Alibaba, Tencent, etc. MT engines are easier to train, users can customize their own MT engines using tools like AutoML Translation from Google or Microsoft Translator Hub. MT aggregators like Inten.to, translate4eu, are emerging at a rapid speed. And we also witnessed the born and growth of LPSs like Unbabel and Lilt, with MT as its DNA. 

In addition to the high-level changes mentioned above, the rapid progressing technology also changed the workflow of translation/localization projects and the interaction between enterprises and language service providers (LSP). Traditionally, the localization workflow on the client side goes like this: the localization team receives resource files to be localized from engineering team and delivers the files to vendors through translation management systems (TMS), after the translation has been sent back, the localization project manager will check these files for quality control and then send them back to the repository for testing or build. With the improvement of machine translation engines, more and more products are going through a MT + post editing (PE) process rather than the traditional translation > editing > proofreading (TEP) process. 

Apart from the changes in workflow, the scope of MT also grows. Companies like eBay, Adobe, and Microsoft have been relying on MT for making their products more accessible globally for a long time. Take eBay for example, in the past, raw MT was mainly applied for user-generated contents (UGC), like search queries, item titles and descriptions, product reviews and descriptions, or member-to-member communication (M2MT) tools. But now, eBay is employing MT for all eBay-created content such as help documentations and UI. Applying MT to all products has already become the localization strategy for more and more companies. Combined with data analysis, companies can localize their products in a smarter and scalable manner, for example they can launch web pages in many languages which are translated by machine and polish those with significantly growing page views. In this way, companies can greatly expand the coverage of target users without increasing much cost. 

Machine learning technology also changes tools used in the language industry. On one hand, CAT tools or TMS tools are becoming more intelligent and personalized. Role-based personalization can result in pre-defined UI layouts aligned to the needs and tasks of given user groups. On the individual level, experience can be changed based on users themselves by utilizing collected customer data. On the other hand, we can see a growing trend of the integration of AI engines and these tools. Currently, most CAT tools are not pre-installed with connectors to cloud based AI engines. But we also witnessed the evolvement of new connectors acting as middleware to a bunch of machine learning engines. 

New Roles of Stakeholders: Machine Learning for Efficiency, Humanity for Creativity 

Like other technologies bringing both innovation and some disruptiveness, machine learning technology may squeeze out some narrow job titles while creating new roles like NLP algorithm specialist, MT solution architect, etc. For stakeholders who are already in the industry, we need to adapt ourselves to the transforming industry. 

For linguists (translators and interpreters included), parts of our job have already been taken over by machines because of the progress of machine learning algorithms, text-to-speech (TTS) technology, automatic speech recognition (ASR) technology, etc. In some fields, like literature, transcreation, human translation still plays an important role. As a translator or interpreter, instead of worrying about being replaced by machines, we should embrace the technology and try to find ways to benefit from it. Also, it is advisable to learn skills like MT post editing, data analysis and the training of cloud-based AI engines. 

For PMs, no matter we’re with the client side or the LSP side, the continuously progressing technology will alter our responsibilities, workflow, even co-worker team. For example, the workflow of MT projects will be different from traditional projects, depending on conditions like MT resources, language resources, and client needs. Before a project is launched, scoping the work comes first: How many words need to be translated? Usually, they come in much larger volumes, like millions of words, than traditional translation projects. Will DTP, MT training, human editing be included? Which MT engine to employ? What’s the evaluation metrics for QA? Can we simply quote by number of words as we are not sure about the quality of MT’s initial output? 

The people to work with even changed when it comes to such projects. We may need a machine translation solution architect who has expertise in the process of an MT project and is familiar with MT tools, so he or she can pick out (or train) the best one for this specific project. We may even need to work with an NLP specialist or computational linguists for further developing our tools. Engineers may want to invest more time in connecting AI resources with CAT tools. Post-editors should be experienced linguists who are familiar with common error patterns of MT and know how to stick to the style guide. 

Given the above possible changes, the role of a PM may evolve towards a resource management direction, he or she must make sure that each person in the team can access the needed resource at the right time and can provide the right solution to clients by combining resources in an effective manner. The PM should also have a relatively strong technical background, and communication, organization and coordination skills, as an effective project plan needs the cooperation of so many parties. 

Developers’ tasks are also changing. By combining AI technologies with local resources like TM, terminology base, linguists, developers are able to build tools with automation plugins, MT pre- and post-processing features, customizable dashboards, etc. They must be aware of the state-of-art machine learning models being developed and know what are needed for their application and deployment. 

For the manager, we need to realistically think about the tasks that will be disappearing over the next few years and start planning for more meaningful, more valuable work that should replace it. Some easy but repetitive tasks are actually a subtle encouragement for people to make narrow and boring job contributions. Machines do not get frustrated or annoyed, and they certainly don’t imagine, they’re more efficient in handling such tasks. But we, as human beings we feel pain, we get frustrated. And it’s when we’re most annoyed and most curious that we’re motivated to dig into a problem and create change. Our imagination is the birth place of our new products, new services and even new industries. So, why not bring more humanity to the language industry? 

Machine learning has brought more possibilities and will definitely affect the language industry in many ways, however, it is important to note that, human efforts still play an important role in many scenarios and many locales where there’s a long way to go to educate people about language technology. Most importantly, humanity is what makes all these possible. 

Launching your Android App with support for different languages is essential for Apps going global. The internationalization and localization of Android Apps is easy when working with Android Studio.

Pre-localization

Before we get our App ready for localization, the first thing to do is internationalization. Basically, we need to externalize all the hard-coded UI strings into an xml file and put it under the “res/values” folder of our project. Next, we could create alternative strings.xml files, each stored in a locale-specific resource directory, based on our target languages. What we’ve done so far can be redeemed as the internationalization engineering part if our App is developed with all the strings hard coded. However, Android developers with internationalization mindset would not build the App this way but will put all the strings into the resource folder for future localization. More detailed information on how to do pre-localization engineering can be found in the Android Development Documentations.

Once we have finished the externalization, our App is ready for localization. Basically, there are two simple ways to do this. First, add new locales to externalized strings under the res > values folder and get these strings.xml files with locales transferred to translators or editors. The other way is localizing with plugins, which will be explained in detail in this blog post.

Why plugins?

Before we start our localization journey, I’d like to mention why localizing with plugins would be cool and in which way can it outperform the first localization method.

The best part about localization with a plugin is that localizers do not need to manually add the string files and send them out for translation. With a plugin installed, we can automate the process of translation and generating resource files within just a few clicks. The other cool aspect is that, with MT implemented with these plugins, users can sometimes customize the MT engine leveraged by the plugin, making the localization more relevant to a specific locale. Moreover, most plugins support a variety of locales which can meet your needs most of the time.

Tutorial

To illustrate how it works, I will use the Android App called Minimal Todo. It is a very light and useful app, allowing you to add todos easily and quickly. The code is very clean and well organized and has been partially localized. When we first downloaded the source file of this App, for localization purpose, we deleted most reference string files and de-localized it to make it more “raw”.

There is more than one plugin out there that can be used for localization purpose, in this case, I chose the one called “AndroidLocalize” and I’ll show you how to work with it in the video below.

Summary

As mentioned at the end of the video, this plugin is more like a black box by using which you cannot tell which MT engine it is using. This would be potentially problematic if you’re quite concerned about data security. Although there are plugins allowing configuration of MT engines, the basic idea is the same: leveraging MT for translation. This means, once the translation has been automatically completed and the App has gone through functional testing, the App cannot be released right away because of the possible issues with translation quality. By employing plugins during localization, we have to include human editors to control the final quality of translated strings.

Although we can see some downsides working with plugins, the advantages of a plugin like AndroidLocalize are obvious too. It helps to save the cost and time and removes many tedious steps in the localization process, such as copying string.xml files, adding locales to each file, translating manually, etc. It is so easy to pick up that even if you’re not an engineer you can still complete the task by yourself.

For the overall information of our Android App Localization project, please click here.

For companies wishing to achieve sustaining growth in the global marketplace, language support alone is not enough for new market adoption. The solution is making the product geo-fit when globalizing it, which means to integrate the cultural and regional factors in your product experience strategy.

Only 5% of companies can sustain a growth globally. If we take a look at some big companies, we may find out this is never an easy process. It took Windows 25.8 years to reach 1 billion daily active users. What’s App spent 6.8 years to reach 1 billion daily active users, and Uber spent 9 years before operating in 65 countries. Along with the international expansion is the idea to develop the global-first product.

How to develop a global-first product?

We may first want to think about what we are optimizing for and how we define success. That is to say, keep the goal clear. First, regarding the market readiness, we want to establish solid new markets entry strategy by doing TAM analysis, evaluate the competitive landscape, look into potential local strategic partnerships, etc. Then, in terms of product readiness, we want to optimize product experience on key funnel, for example the onboarding and core engagement of users. Additionally, we also should take organization readiness into account. We need to figure out the international team positioning within the structure because horizontal cross-functional effort requires alignment on the corporation core objectives for effective impact execution.

As global product strategy is actually the local product strategy on a global scale, after clarifying our goal, we also need to know local customers and local market. To take local customers into account, switching from US-first to User-first is essential.

By saying that, let’s image we are users. When introduced to a foreign product, what are some questions we may ask ourselves before we move forward? We may first want to identify whether this is something we’ve already know. Then, we want to know if we can use it and whether we care about it. If we can use it and we do care, we may get interested. The user-first model requires such thinking from users’ perspective. However, there are some common misconceptions in global product strategy: One size fits all global product experience can be created; we can get the product right for EN_US first, then optimize for international; EN market is the core because that’s where the money is. Sadly, these are ALL WRONG. In fact, the emerging markets have grown 2-3 times the pace of developed markets and are expected to account for >50% of global GDP by 2030.

Cases: Integrating Regional & Cultural Factors

Now, you’ve made up your mind to put your product on the global scale and started mapping out your globalization strategy. Taking an international brand to a local market is complicated and needs coordination between many different parties. But you always want to integrate regional and cultural factors into this process to make your product more locally relevant. Different companies may have different approaches, but I’d like to share some cases that I really like and hopefully can give you some new ideas.

Chitu (赤兔)：LinkedIn’s local App for Chinese users.

When LinkedIn decided to develop a brand new localized product to Chinese users, it grasped the idea that the importance of networking or “Guanxi” for career development is deeply rooted in Chinese culture. Inspired by the name of a horse of a Chinese ancient hero, the App brands itself as a social media targeting on young people without strong background, who mostly work at second-tier cities. Chitu knows what local customers want: they are eager to learn how to promote their career path and connect with each other.

Besides the difference of user targeting, Chitu involves more popular features like Live Mode and knowledge monetization.

It also launched three branding campaigns to establish trust from users. It also established partnership with local companies like Alibaba, allowing users link their Alibaba credit rating with LinkedIn profile.

Starbucks: Brewing brand trust in Mexico by hiring employees over 60

Starbucks Mexico opened its first shop staffed by people over the age of 60 this year. This not only contributed to the elderly community in Mexico but helped Starbucks gain local trust and good reputation.

Most of the times, the localization I’m exposed to is mostly about language localization, however in the there is a lot more to consider besides language: put user at the first place, integrate regional and cultural fact, and stay relevant and add value.

(This blogpost is inspired by Talia Baruch’s speech at Middlebury Institute on Thursday, September 13, 2018.)

Introduction: This blog post is about a SMT training project using Microsoft Translator Hub. In this portfolio, you will read about the description of the project, lessons learned from the project, my ideas on how we can improve translation quality by leveraging the power of MT, and my experiment of other techniques to improve the overall localization workflow.

SMT Training Project: Developing a MT engine for translating Chinese Government Work Report

Project Overview: This pilot project is designed to establish a Statistical Machine Translation (SMT) engine from English to Chinese in order to support the Chinese government initiative to provide public documentation in both Chinese and English. Our data source will be extracted from Chinese government public report websites.

In order to be considered “fully trained”, the post-edited machine translations (PEMT) from this engine must meet the following target criteria for efficiency, cost savings and quality:

• Efficiency: PEMT 20% faster than human translation
• Cost: PEMT 25% savings over human translation
• Quality: PEMT with an acceptable score of less than 30 based on the Multidimensional Quality Metrics (MQM). The acceptable score was increased from 10 points to 30 points due the length and complexity of the product review compare to the government reports.

A total of 13 rounds of successful training were completed with an initial data set of approximately 56,916 segments used as training data, approximately 1,270 segments used as tuning data, and approximately 1,559 segments used as testing data. A BLEU score will be given after each round of training by the Microsoft Translation Hub. The initial round of training achieved a BLEU score of 12.6 and over the next three weeks, 12 more rounds of training were done with a result of 18.61 (~48%) as the best BLEU score achieved. Finally, the MT system with the highest BLEU score was deployed and the translation of this system was compared with the human translation team to reach our final conclusion of this project.

Data: If you feed the machine with garbage, your machine will output garbage. To maximize the quality of the input data, the official bilingual files from Chinese government were used and each file was carefully segmented and aligned using CAT tools.

• • Testing Data: Chinese Government Work Report (2014/2016/2018)
  • Tuning Data: Chinese Government Work Report (2013/2015/2017)
  • Training Data:
    • Chinese Government Work Report before 2013
    • UN Translated Dataset downloaded from OPUS
    • Chinese Government White Paper collected from The State Council Information Office of the People’s Republic of China
    • Chinese Government Bilingual Documents from China Academy of Translation

1. Timeline

*GWR: Government Work Report; WP: White Paper

2.Conclusion

A 500-word excerpt from Government Work Report was used to test the deployed engine. The MT output was post-edited by editors and errors from the output was identified. On the other side, the same file was human-translated. After comparing the time, quality and cost, we reached the following conclusion.

(Comparison of errors found in HT and SMT using MQM as the criteria)

The outcome of our SMT did not meet any of our goals for this project. Nevertheless, the team has put tremendous effort into the allotted time and completed the entire process from initial planning to deploying the trained SMT.

Lessons Learned

The experience and knowledge gained from this pilot project resulted in the analysis to accurately project the requirements to fully train this SMT to support the government work report initiative.

Training Data Requirement:

Based on our pilot project result, we project at least 1,000,000 segments are required to fully train the SMT engine. The major quality issue resulted from our SMT engine are grammar, more specifically sentence structure problem. We conclude that increasing the quantity of training data and tuning data will significantly improve the grammar quality of the SMT engine.

Additionally, well-aligned source documents are key to the success of the SMT training. We only accepted the official government translations and conducted multiple alignment and reviews of our data. The result was less quantity but better quality, the result was evident in our 9th round training result. Our BLEU score increased from 13.4 to 18.6. We believe that adding a dictionary and monolingual data would have significantly improved our BLEU score and the overall quality of our SMT engine.

Time and Cost:

We project 783 hours and about $35,000 are required to fully train the SMT engine. A team of at least 4 full-time members would ensure the SMT can be fully trained to meet the goals. Our team significantly increased the rate of SMT training rounds after the fourth round, which has helped us quickly identify a deficiency and the direction we should focus on.

Tools:

Initially, our team only used the TMXMall as our primary tool for favoring its simplicity and accessibility. But, quickly we found the quality of our aligned documents have suffered from our tool or choice (see first four rounds). By implementing TRADOS Studio and Okapi Olifant, we were able to align all of our documents into sentence segments and drastically improved the quality.

In conclusion, we do not recommend using the SMT instead of the human translation for the government work report initiative. We believe, with the proposed plan outlined above, the fully trained SMT engine would meet the established goals in order to be utilized for future government work report initiative.

Meet the future: Customizing a Neural Machine Translation engine?

When identifying error patterns of MT outputs, we captured some causes of these errors. Word order is a typical error of the MT output. The MT system tends to translate word by word, which ignores the fact that following the same word sequence in the target language may actually doesn’t make any sense. This can be partly improved if we create a dictionary or term database that is linked to this system. But the NMT system may be better at delivering a more accurate translation as NMT better captures the context of full sentences before translating them.

By the time we finished this project, Microsoft has not opened its NMT system for customized training. But the good new is, THEY JUST DID THIS a week ago. Click here to learn more.

Personally, I’m quite interested in training an NMT engine and can’t wait to see the results. What I do believe is that no matter how unsatisfied we are with the output by existing MT, being able to utilize various techniques is a basic skill for any ambitious translator.

Technology is not always about some unreadable codes or algorithms, sometimes utilizing a simple tool like this will make your translation work much easier.

https://youtu.be/hau8DHpKoVY

There’s much more to be explored. Always being curious, eager to try, and open to latest technology is perhaps something most valuable I get from the CAT course at Middlebury Institute. Thank you, professor Adam Wooten.

(You can access the files of our SMT project here.)