Recent media coverage of developments in machine translation has led the world to believe that it is only a matter of time before computers will translate as well as humans. While neural machine translation (NMT) is a rapidly developing field, some developers’ claims of phenomenal improvement are debatable.(1) In 2016, Common Sense Advisory (CSA Research) found that growth in “pure” human translation has almost plateaued, with only modest growth forecast in the coming years. In contrast, huge increases in volume of machine translation will occur and survey respondents reported plans for double-digit annual growth in post-editing volume between now and 2019.(2)
However, since CSA Research conducted that research, a new, disruptive technology paradigm has emerged. This has the potential to lead to dramatic growth for human translators & to change their place in the international content lifecycle. To understand why, consider the position of translators today with post-editing: They sit at the end of a chain that seems purposefully designed to leave them powerless and unfulfilled.
Image Caption: Post-editing today leave translators at the end of the line
In this process, language service providers (LSPs) receive or generate raw machine translation output. If they are lucky, this can be quite good, but it often borders on useless, and the job of the linguist is to make the proverbial silk purse from a sow’s ear somehow. All this for a fraction – typically about 60% – of the price of full human translation, even if it takes longer to edit. This position engages linguists in low-value work and does not let them apply their skills effectively. In addition, their work has no real means to deliver improvement for the next cycle. No wonder many linguists dislike post-editing.
At the same time, CSA Research estimates that, world-wide, less than 1% of content is translated compared to an ideal world. This is because human translation takes too long and is too expensive, and MT is often not good enough. Addressing this gap for economically significant languages would require approximately two billion human translators.
So how can the world address the demand for language services with the limited supply of providers and the inadequacy of machine translation for use in cases with high standards for accuracy and fluency?
CSA Research finds the answer in a new form of translation, one it terms “augmented translation.” This term is based on “augmented reality,” as exemplified by new applications that overlay information on reality, rather than trying to replace it. Pokemon Go is perhaps the most visible example of this. However, it is increasingly important in many spheres of life. Augmented translation uses artificial intelligence as an aid to human effort rather than a replacement. It is an example of “intelligence amplification,”(3) in which machines handle data processing and help humans understand materials and make decisions. In this case, human translators sit at the center of the process and all translation is human translation.
In augmented translation, human linguists sit at the center and manage multiple resources
Rather than correcting MT output to make it marginally acceptable, humans handle all content, while drawing from five key technologies that amplify their abilities and speed the process up:
Lights-out project management lets linguists pick up work when they are available. It increases their flexibility and helps match jobs to the right people. AI components learn which translators like certain kinds of work and what they are good at, then assign those jobs to them.
Translation memory will continue to help translators see what they have done in the past. However, rather than requiring them to use exact matches with no review, it provides them with options that they are free to use or adapt.
Terminology management tools will suggest terms for translators. However, instead of relying on top-down definitions of terminology, the next generation of tools will help translators discover and document how terms are used “in the wild”, then provide this information at a click when they need it.
Adaptive machine translation will provide results for translators based on their writing style and history. Instead of translating entire sentences, it will suggest individual words and phrases that translators can use, modify, or ignore. Early adopters of this technology report that, by itself, adaptive machine translation can double translator productivity in some cases. Increasingly, this technology will use NMT to deliver better results than older phrase-based systems can.
Automated Content Enrichment (ACE) is a new technology that scans content to identify the concepts, dates, places, and other information in it, then link them to online resources. For example, if ACE encounters “London” in a sentence that also mentions “White Fang,” it would deduce that the text refers to the author Jack London, not the city. ACE then links to information about the author. Finally, this information is made available to assist translators and to improve MT by helping it determine which translation makes more sense in a particular context.
These technologies will communicate with one another and when used together, will provide their best results to translators, when and where they need them, without disrupting their work. Unlike older translation technology, all of this will be available in one environment and language workers will not need to move between various applications and paper resources. Instead, they will supervise the operation of these technologies. In cases where the technologies work well, the human will check the results and give approval. However, when difficulties arise, the translator can step in. Fortunately, issues that are difficult for machines are often the ones that can be the most interesting for humans, and contribute the most value. Augmented translation is not yet fully realized since developers are still working on the pieces. Some technologies already work well together, but the emergence of powerful architectures based on micro-services makes it much easier than before to combine them. For example, many tools today tie translation memory together with basic MT to provide MT in cases where no matches are found. Augmented translation will go further in the direction of seamless integration that adapts to and learns from individual linguists.
For translators, augmented translation offers a way out of constantly declining per-word rates. As they become dramatically more productive, word rates can fall much more, even as translators’ pay goes up. Pricing models will surely change, but cheaper translation prices will increase demand for human services and let them expand their market. Augmented translation, combined with translators’ superior ability to meet quality requirements, will give them a competitive edge over MT in situations where they are too slow or expensive today. It will also put them in a centralized position where they belong, one where their intelligence and insight receive well-deserved recognition.
Text Sources:
[1] Common Sense Advisory: “Is Neural MT Really as Good as Human Translation?”, Click: here
[2] Common Sense Advisory: “MT’s Journey to the Enterprise”, Click: here
[3] QUIRK’S Media: “A look at artificial intelligence vs. intelligence amplification”, Click: here
