Trinity isn’t just 3 dimensional and The Three Musketeers ain’t just for 3 but for all
Published on TAUS Review Issue 2
Preface or: How I Learned to Stop Worrying and Love the Black Hole
The author is fully aware of the confusion and inaccessibility of this article, from the topic to the cheesy metaphors. The intension is to show how ideas and visions can be too open to interpretation and then resulting in almost no way to be sure except through contacting the person who is really responsible in order to explain the whole story with context. Readers may find it hard to relate unfamiliar technical terms or imagine uncommon phenomena such as high dimensional theory , and then ask for definitions or names. For what it is worth, the author recommend readers to give the movie “Interstellar” a try, and think about how nameless and virtual high dimensional space-time could be. The quest from now one would be just like going where no one has gone before, such as a black hole, which somewhat carries highest possible entropy. Or putting it in the way of an anecdote: entropy was a fancy term suggested by John von Neumann to Claude Shannon in order to replace “information” because “nobody knows what entropy really is.” If readers are kind enough to bear with this writing style, there’s one thing more related to translation. The writing style is somewhat similar to East Asian way of rhetoric that tends to produce obscure texts. Imagine how articles in classical Chinese and Japanese can be so Zen-likeish. The exotic scent might be long gone unless the author had tried to imitate J. R. R. Tolkien. Since that’s not the case, the author will still provide a plain explanation as much as possible.
Quality, speed, and cost are usually considered as the three most important factors for translation services. To be general, effectiveness, efficiency, and satisfaction are well known as the three goals forof the usability offor any service, and can probably be seen as a paraphrase of the former. To be specific, adequacy, fluency, and elegancy (or the pertinence of pragmatics) are often argued in Mandarin Chinese as the three criteria for translation quality. From the above three examples, one can easily notice how much people are in love with three-degree perspectives. However, just like Newton’s laws of motion are inappropriate on very small scales, at very high speeds, or in very strong gravitational fields, the Information Age’s translation services are facing new challenges with massive contents that require affordable time and money to localize. Due to these changes, it will be intriguing to see if the beloved 3-D aspects still hold. If I can insert a pun here, “When in doubt, C4!” This Mythbusters quote is similar to the idea behind this article. The difference here is that C4 is not necessarily a destructive explosive, but a complementary fourth character.
Similar to the Christian doctrine of the Trinity, i.e. The Father is not the Son is not the Holy Spirit but all the above is God, a translation job’s characteristics in terms of quality, speed, and cost are often described as at least one will be sacrificed if the other two are dominant. Unlike the Trinity, however, a Venn diagram is usually applied to indicate there is a small intersection of the characteristics, a so-called to be a promised land. The concern roseof by the editor of this magazine is, the recent trend seems to suggesting that the faith is fading away and many clients who are in doubt might be ready to just give up on the quality. This article would like to respond within an alternative view with a fourth character, or more precisely, any high dimensional perspective will do. For example, there could be more than one way to look at a Venn diagram of quality, speed, and cost. When an observer is allowed to seeing it from the side, there will be always an angle where thethat two circles of the diagram perfectly overlapped. Furthermore, if the observer has the liberty to change the view angles dynamically, all three circles of the Venn diagram are just the same thing. Although in the fields of physics and statistical machine learning, models such as string theory or support vector machine do not come with names of higher dimensions, readers might settle with a fourth character as the observer who sees everything relatively, just like the one from the theory of relativity. As for how an observer can gain as much as possible for a translation job, the following sections will try to sketch iit outt.
The algorithm-changed world changes algorithms
The algorithms of search engines enabled the world to perceive and conceive more information in a shorter and shorter span of time, but the language barrier feels higher and higher once people are escaping the ignorance. Not so many years ago, a global company may take months to be prepared for launching a new local branch. Now, a single new product from Apple, say the iPhone 6 plus, must have all localized content available online without a time-zone difference. For modern translation services, machine translation appears to be the only choice that is fast enough. Since the quality of machine translation is usually unacceptable, a typical compromise is to ask the customer to prioritize the content, sacrificing one part in a trade for another. The inconvenient truth is, even with the help of translation memory and a term base, once the amount of requested content reached certain threshold, the translation task will exceed reasonable delivery time and budget anyway. Under such pressure, new algorithms for computer programs and human operations are desperately needed.
80–20 rule, dynamic programming, and divide-and-conquer
Without a loss of generality and confidentiality, based on personal experience, a 100-million-character project in 2 months could be the toughest situation. When it was as modest as 10,000 characters per week a project, 3 Japanese yen per character with a simple estimate-execute-evaluate waterfall approach is reasonable. The operation can even take advantage of crowd sourcing for proofreading for an affordable cost. But when it comes to 100 million characters, all of sudden the perception is beyond Hooke’s law, and is not that linear any more. Yet somehow, a typical misconception is that when the demand is so high, wouldn’t it be appropriate to show some price elasticity? To make the most optimistic prediction with 20% repetition rate in terms of character, there will be still 80 million characters to go. Luckily, such a huge task can always have a silver lining: it’s big enough to provide statistical and linguistic insights. Sometimes this even comes with behavioral psychology hints to bring the stimulus-response back to a bearable range. Otherwise, randomly distributed translation jobs will wander in the market forever without anyone to take it since the high demand also implies low-hanging fruit will not be short in supply.
Let’s revisit the workflow with a computer assisted translation (CAT) system. It is intuitive to have a text translated from top to bottom in respect of context. However, a program should remember everything for you, back and forth. What if an algorithm could gather 20% of the most similar items together with context for you? What if an algorithm could analyze the critical path to complete as much translation as possible, like the artificial intelligence of chess game? What if the former two if’s could ultimately create positive feedback that would resolve the translation project in an exponential fashion based on the power of 20%? More importantly, its real objective, i.e. the 4th and the one essence, should be to guarantee each segment of text to be tractable and remain intact once it is translated properly. In this sense, a divide-and-conquer strategy is to find manageable jobs from a larger, seemingly intractable mission, where effectiveness/quality and efficiency/speed are merely due diligence embedded in its very nature.
For instance, after analyzing the 100-million-character project with recent advanced natural language processing tools such as word2vec, similar phrases can be clustered and sorted by frequency, by length, by semantic distance, etc. One may find “semantic distance” incomprehensible, hence a simple analog is invented: if the distance between “man” and “woman” are somewhat similar to the distance between “king” and “queen,” it is feasible to form a simple algebra like “king — man + woman = queen” for fun and for real world problem solving. Once the project manager is equipped with all the above well-organized phrases, it is just the matter of strategy to find the right person or the right system to digest each job, based on their specificity and the sensitivity, in terms of statistics. So the next significant question is who is going to be this wise guy?
D’Artagnan of the team
Computer programs and the ones who develop them may have their potentials, but a good team, which can survive with low cost and high satisfaction, is always more than that. It has already come to some, if not all, translation service providers’ attention that agile translation, lean localization, and many other buzzword combinations, introducedo something for those entire waterfall vs. iteration comparisons to the translation industry. It is easier for the service provider to say no to more one-shot estimate-execute-evaluate waterfalls and welcome an era of growing and reinforcing iterations. Yet, the stakeholder’s, i.e. the customer’s point of view, remains unclear and deserves emphasis. A highlight here will be the 4th and the one character: chicken. In the metaphor of Scrum, a famous methodology for agile/lean software development management, a chicken asks a pig to be a business partner in a restaurant venture, Ham-n-Egg. When the pig (developer) feels committed, the chicken (client) only appears involved. In spite of how weird it sounds, at least the chicken is involved. In reality, the relationship between a customer and a translation service provider is no better than the above story. Assuming there is one global business owner submitting several texts to a crowd translation platform, both the customer and the service provider may or may not notice whether crucial information is or not going along with the task to faceless translators. A funny true story is that, there was an app development company requesting localization for a short commercial, which contains another organization’s name, while neither the customer nor the translator realized that the organization name looked so like two common nouns with one colon in between, e.g. “X: Y.” The resulting translation is fluent and appropriate to be an advertisement, but the organization name is totally ruined. What if the chicken, or to name it with more respect, D’Artagnan, i.e., the customer, can involve in more than just the content and the paycheck?
The previous story actually came with a regrettable twist. The one who finally spotted the issue was a reviewer responsible for vetting translators periodically. Since the reviewer is not the proofreader or any other role on the project, the damage was already done. If one is willing to listen to the lesson learned from software development, test-driven development could be one effective attempt to fix it. Although many CAT tools or translation management platforms provide basic quality assurance such as length check, tag/symbol/punctuation validation, whitespace normalization, etc., there are more semi-automatic approaches worth pursuing. For example, under certain circumstances, many different Japanese expressions can be translated into the same English phrase. It sounds like good news when a translation service provider is recycling translation memory, but the whole document may disagree. Since they can be easily detected, a test of contextual relevancy can be conducted. The best part of test-driven development is that it is not only a reusable safeguard, but also a clue to enlighten the team towards a deeper understanding of a certain subject. Eventually, test cases will be patterns, templates, exemplars, and feeding back into the algorithm of computer and the people around it. Another example in terms of Mandarin Chinese is about the Simplified-Traditional and the China-Taiwan-Hong Kong-Singapore situation. If the target languages are both Simplified and Traditional Chinese, different characters and terms can be looked up in a carefully crafted table. Ultimately, it is up to the customer to decide how to proceed: converting Traditional Chinese into Simplified to keep the ambiguity low? Taking the opposite direction to secure the higher supply of translators? Or like what Apple did: for the China market oriented website, “Bigger than Bigger” used to be a blunter translation than Taiwan and Hong Kong. However, after receiving end users’ complaints, the former one turned to follow the same way of the latter one, just with different characters. There’s probably no way to tell if Apple has them tested internally, but certainly many companies do not want a slogan out there and have it tested the hard way.
A job is a story
Yet another unfortunate lesson learned is from a static HTML page translation job. After a the application of the fashion of estimate-execute-evaluate waterfall approach, the customer realized that the texts of Open Graph Protocol (OGP) properties that are usually used for Facebook optimization were never extracted and translated, not to mention they were not included as part of the quote in the first place. Since the customer was not involved until the evaluation phase, it is easy to imagine the awkwardness between every party of the job. At this point, quality, speed, and cost are not relevant anymore, if the whole intention was to maximize the customer’s exposure on Facebook. The software development had many catastrophic experiences concerningabout that, so finally it came out with aa treatment: job story. In terms of translation, counting words/characters of a given job and multiplying with unit price (e.g. 3 yen) and estimated time (e.g. 3,000 character per day per translator) is seemingly a universal practice. The quality is then projected to the former two factors with educational guesses. When it comes to the HTML page translation job, the quality might be just fine but the purpose is lost, even though the character-count of OGP is typically small or even zero — if they are generally duplications of other common HTML texts came with the page. So a new paradigm of job story is trying to prevent mistakes that might come from it: just write the true purpose down in this way: when (situation)…, I want to (motivation)…, so I can (expected outcome)…. The involvement of a customer can be as minimal as the above simple statement, and soon strangles many tragedies in their cradle.
There are always more dimensions outside of the box
The trinity of translation is quality, cost, and speed, but quality is neither cost nor speed, it is the box. Using machine or crowd translations to increase speed or to decrease cost might be good ideas inside the box, but the box always comes with boundaries. One of the final issues (if not only one) is that, how much costs will an extra dimension demand? When a 100-million-character project comes to you, will it be comfortable to pay for research and development first, and then reduce the size of work in return? Admittedly, the profit might not cover the loss on the very first try if the magic came with an outrageous price. Even worse, every once in a while a chicken may throw in the chicken or egg question: since you have not done it before, how should any customer trust you? Well, in the long run, if we are not dead yet: to be fading away, or not to be?
To boldly go where no one has gone before
Recently a Medium post reminded us of the great accomplishments of Margaret Hamilton, who is famous for being the “one in which computer science and software engineering were not yet disciplines; instead learning was done on the job with hands on experience” and also invented the Universal System Language (USL) that has “taken on multiple dimensions as a systems engineering approach.” Looking at her portrait of the Apollo program might shed some light: programming at that time was more like a human translation task thaen programming today. Such an incredible amount of cards for the Apollo computers had helps from USL and this eventually let people land on the moon. Before we all get there, the author will just admit that he does notn’t know what will be the silver-bullet as the fourth character in a higher dimension. However, the beauty of scientific approach is that at very least we all can try, and every falsifiable hypothesis will help us to improve by negation. In the end, even theology needs via negativa to define what is not included in the Trinity.