Translation innovations that are leading the change and may help us surviving the change
Published on TAUS Review Issue 4
Innovation is uneasy, especially about the differentiation. There are many different kind of innovations, ranging from improved customer satisfaction to invented market demands, and in terms of the recent trend, the “disruptive” kind is the most welcomed innovation. By decorating an adjective “disruptive”, however, doesn’t really weigh clearer degrees on the spectrum of innovation, hence it is common to transform the question to the degree of inefficiency: how efficient of the target market? Here we adapt this viewpoint for the market of translation, and then try to identify certain disruptive innovations that might be game-changers, in particular two factors of inefficiencies: human-resource and software ability.
Inefficient supply-demand match making
Crowd-sourcing translation services are quite popular in recent couple of years. On one hand, crowd-translation indeed increased supply, and potentially decreased the price. On the other hand, the information asymmetry between buyers and sellers remain almost the same. Customers may still find it difficult to locate the best translator who will yield the expected outcome, and vice versa. While it is a general question to most of businesses, translation can be even tougher when the work of conduct requires further “translation” by its nature. Typical tactics to tackle the problem include categorizing source texts, relying on translators’ reputations, assuring qualities by post-editing and/or proofreading, etc. What if a match making can be done systematically and semi-automatically? That is one of the questions DuoLingo is trying to respond.
Luis von Ahn, the creator of reCAPTCHA, started DuoLingo. According to the very first sentence of its Wikipedia page, DuoLingo is “a free language-learning and crowdsourced text translation platform.” To comprehend how language-learning and crowd translation align, it might be easier from understanding how reCAPTCHA works. Like the original CAPTCHA, reCAPTCHA asks web surfers to type words shown in images, to prevent bots accessing restricted web pages. Unlike CAPTCHA, reCAPTCHA presents two sets of word images, one is with known answer in the system, and another is not yet “translated” since optical character recognition (OCR) doesn’t work on such image. The idea is to utilize the wasted human effort on entering CAPTCHA, so the entered words can actually be the answer of failed OCR attempts. When it comes to the efficiency of digitalizing paper-based archives, reCAPTCHA eases the pain of normal validation that is usually labor-intensive. One may ask if the hard-to-OCR text image didn’t come with an answer, how could reCAPTCHA be trustworthy? The trick is that if the known image got the correct result, it is usually true for the unknown one. Furthermore, same unknown image can be presented to different users, so a voting system could help as well. By replacing every appearance of image, word/text, and OCR on the above description with source language, target language, and translation, it shall start making sense to remodel reCAPTCHA into DuoLingo.
One obvious problem to remodel reCAPTCHA into a translation service is that a bilingual web surfer is not as much as a free lunch than a monolingual one. So DuoLingo began as a free language-learning platform with gamification traits, which successfully significant amount of learners. The top learners would be introduced to translate some new texts and vote others’ translations. New texts could then be used as new learning materials. As DuoLingo revealed, the demand of translation is currently from CNN and BuzzFeed news. Despite this business model is still in its early stage, it is intuitive to see that eventually texts in different genres and domains will join this learning-translation feedback. Recently DuoLingo launched a language certification service, and in our opinion, this complimentary business model will strengthen the learning-translation feedback and build a healthy cycle by inversing the supply-demand dependency every once in a while, so buyers and sellers will both know what they can get. For example, BuzzFeed first earn some extra bucks by selling their translated news articles to DuoLingo as learning materials, and later DuoLingo may sell more translated news articles back to BuzzFeed with the quality no worse than the learning materials, while learners may get certified and then motivated to join the translation industry.
Inefficient corpus annotation for machine translation model training
If we swap reCAPTCHA’s image, text, and OCR elements with “machine” translation terms, would it be as promising as DuoLingo? One major issue lies on the preparation of learning materials, or annotated corpus in terms of machine translation. For DuoLingo, it is reasonable to leave the alignment problem, as in determining which source phrases correspond to which target phrase, to learners. For common machine translation training, however, it is required to include alignment as one of corpus curation tasks, because we want machine translation models to remember minimal phrases and their translated counterparts, to generate full translations out of them flexibility no matter the exact source sentence is seen or not, just like what qualified translators can do. One may notice that how similar this task sounds as in reCAPTCHA-to-OCR model training scenario. The unfortunate difference here between OCR and machine translation is that phrase alignment is not always straightforward. One component from a source phrase could be missing in its target phrase, not to mention that not every language in the world comes with whitespace delimited word boundaries. Typical machine translation modeling nowadays involves a module of automatic alignment, such as GIZA, which requires carefully handcrafted training data. Something needed careful handcrafting often implies inefficiency, and an even worse situation of alignment is that it has to be done by experts, and the market of this kind of experts are limited, so either reCAPTCHA or DuoLingo won’t help much.
What if machine translation modeling can get rid of the burden of alignment? Thanks to the advance of deep learning, there might be a solution. A series of research led by Yoshua Bengio has showed that recurrent neural network, also known as deep learning (at least one major type of it), may be a potentially good alternative. Instead of aligning source-target phrases strictly, this so-called neural machine translation performs “soft” alignment. If we try out best to describe what a soft alignment is, we could say that neural machine translation attempts to imitate how human translators do their jobs: searching for the best fit translation for a phrase in their minds, without trying to align every phrase in a strict one-to-one fashion. As one may see in the figure of the demo site, while old fashion one-to-one alignment is in black lines, neural machine translation can also take grey-scaled lines into account.
It could be somewhat confusing that more alignment possibility seems to be more complex. At first glance, it is. However, neural machine translation’s soft alignment doesn’t need human experts to annotate those multiple choices in grey-scaled lines. Neural machine translation doesn’t ask for a single standard answer of alignment. It just learns to weigh its options. If we rephrase this nice feature in translation industry’s way, one may say that the only thing neural machine translation asks for is translation memory.
If they are so magical, how come they haven’t taken over the world?
Both DuoLingo and neural machine translation have their weakness. DuoLingo doesn’t really care about every perspective of translation quality, it only cares about macro factors, such as consistency between learning material and translation demand. Acquiring learning materials of non-alphabetical languages to bootstrap the whole cycle are not a trivial task. Similar obstacle can happen to neural machine translation, too. All the fine-grained specifications one may encounter in translation business will be still there for quite a while. That being said, disruptive innovation probably shouldn’t be seen as savior of the industry. It is more like a shiny new armor or sword to help common mortals to survive in the constantly changing environment.