How Baby Mouth Hand Was Born?

Pretty sure, much of you have seen funny or, worse, annoying mistranslations on packaging making interpretation of what’s inside a big challenge. And you likely thought the product was of as bad quality as its accompanying packaging. Chances are you failed at this point: though it is reasonable to think the manufacturer paid the same level of attention to the quality of its product as it paid to the packaging, actually the two may have nothing to do with each other. But no matter if this holds true or not in a precious case, if you can buy something sold in a box speaking clearly, you will likely buy that instead—the market is just full of competent products in most segments.

Credit: By McKay Savage from London, UK [CC-BY-2.0]

It also features unparalleled cost-to-value ratio to assign translation tasks to those translators not investing time, energy and money to become experts of a domain or two, so they will also rely on dictionaries. Dictionaries, however, do not always detail the context in which a possible translation is suitable, and do not contain all the target terms popularized by native speakers in one or other domain for each source term. “Slip carefully” is indeed an extreme and rather funny example, especially if you speak English natively, since you will likely find out, watching the pictograph, what it is about, and you will recognize the stairs are slippery. But look back to the hero image on the top of this post, the completely non-sense set of words: will you ever find out what they wanted to tell you? Will it make the mouth and the hands of the baby “fragrant”, to smell scented? Or will it make the mouth and hands of the baby free from scents to smell naturally? Do we speak about the mouth and the hands of the baby, at all? Impossible to decide until trying.

Should there anything worse than a translation made by an unprofessional exist, and it is called machine translation, for many languages. Machine translation (MT for short) engines rely on one of two possible approaches: morphological analysis and statistical translation. Morphological MT engines analyze the structure of the source text and the context of the source terms to try to understand relations between terms and clauses such as possessive and adverbial structures, suffixation and pre-positions, or subordinate and coordinate relations, and they are—theoretically, at least—capable of re-creating the same structures and relations in the target language. The key to the success for a morphological engine is, therefore, the need to accurately and deeply know the grammar of at least two languages (a source and a target). One problem, however, remained unresolved in the past decades spent by trying to create excellent morphological MT engines: the task of writing down a grammar in the form of algorithms or rules such that it provides consistently high quality translations proved not to be viable for many languages, including Hungarian, for example, because no matter how may grammar rules you observe and formalize, there are still an uncountable number of exceptions, and trying to create rules for all those would simply be inefficient, if not a never-ending story.

Inefficient, because the competitors of morphological MT engines, the statistical MT engines, preferring robustness over sophistication, are cheaper to develop to such extent that even if they provide generally worse results, such an engine becomes more effective and provide better ROI in real-life situation even if you add the cost of human post-editing. Most of today’s popular public machine translators (such as Google Translate and Bing Translator, to mention two), as well as those used internally by large enterprises, are basically statistical engines. To put it simple, a statistical engine takes a sufficiently large sample of pairs of sentences in the source and target language, and divides these sentences into smaller structures (called n-grams) consisting of one, two, three or a few more words. These fragment pairs (a corresponding source and target fragment) are then persisted to a database. When it comes to translation, the engine divides the source text to similar fragments and looks up its database to find the best match for the given fragment. Unless exact match is found, the closest match is selected, which may feature different relations, different suffixes and so on, things which an MT engine cannot consider and fix. If there is no exact match for “to the room near the kitchen”, and the best match is “from the room near the kitchen”, it will use this, even if we realize the direction of some movement or progress is reversed. For many languages only the law of averages guarantees a translation made this way will be a perfect translation—though the translation of the words present in the original text will likely be present in the target, the word order, the relations between clauses and the correctness of suffixes will mostly be casual, while proper use of these would all be required in order for the translation to make sense and represent the same concepts, ideas or facts.

As per the images published in this post, I can see equal chances that these are products of an unprofessional translator writing down dictionary entries one after the other, or the products of machine translations. Whatever happened actually, both examples are great proofs why clients are much better involving professional translators to protect their values, investments and brands at the cost of a few pennies.

_{(Hero image credits: Dan Gilles, Dan in China)}