Modern AI approaches are used by the software, yet straightforward testing reveal that it is still far from being able to understand.
One Sunday, my friend Frank brought a Danish visitor to one of our regular salsa sessions. Frank’s mother was Danish, and he had grown up there, so I was aware of how well he spoke the language. His acquaintance spoke English with ease, as is typical of Scandinavians. To my amazement, it turned out that the two friends frequently used Google Translate when exchanging emails during the evening’s small talk. In order to create a fresh text in Danish, Frank would write a message https://translate.google.com/?sl=en&tl=ru&op=translate in English and then run it through Google Translate. In contrast, she would compose a message in Danish and then let Google Translate translate it.
So strange! Why would two clever individuals who were proficient in each other’s languages act in this manner? I’ve always been quite dubious about machine translation software because of my personal encounters with it. But it was obvious that these two did not share my mistrust. It’s true that many thoughtful people find little to complain about translation software and are rather quite fond of it. That puzzles me.
I have followed efforts to automate translation for many years as a passionate translator, a person who loves languages, a cognitive scientist, and a person who has always admired the subtlety of the human mind. Midway through the 1970s, when I first became interested in the topic, I discovered a letter from the mathematician Warren Weaver to Norbert Wiener, a significant figure in cybernetics and an early proponent of machine translation, written in 1947. In this letter, Weaver made the following intriguing claim, which has since gained widespread notoriety:
“This is genuinely written in English, but it has been coded in some bizarre symbols,” I think when I read a Russian piece. I’ll now start the decoding process.
Years later, he had a different opinion, stating that “no sane person would ever believe that a machine translation can ever reach grace and flair. Pushkin doesn’t have to tremble. Whew! I find this remark by Weaver to be much more hospitable than his earlier remark, which reveals a strangely simplistic view of language. I have dedicated one unforgettable intense year of my life to translating Alexander Pushkin’s sparkling novel in verse, Eugene Onegin, into my native tongue (that is, having radically reworked that great Russian work into an English-language novel in verse). Yet, his 1947 theory that translation is a process of decoding evolved into a tenet that has long guided the development of machine translation.
Since then, “translation engines” have steadily advanced, and more recently, the application of so-called deep neural nets has led some observers to speculate (see “The Great A.I. Awakening” by Gideon Lewis-Kraus in The New York Times Magazine and “Machine Translation: Beyond Babel” by Lane Greene in The Economist) that human translators might be a declining species. In this case, human translators would eventually stop producing original new material and would instead serve as quality controllers and bug fixers.
My mental life would undergo a soul-shattering upheaval as a result of such a development. I am completely aware of how fascinating it is to try to programme robots to translate correctly, but I am not in the least interested in seeing human translators replaced by inanimate objects. In fact, the thought horrifies and disgusts me. In my opinion, translation is a really sensitive art that continuously calls upon the practitioner’s extensive life experience as well as their imaginative creativity. My esteem for the human mind would be severely damaged if human interpreters were to one day become a thing of the past, and the shock would leave me whirling with awful perplexity and immense, irreversible sadness.
Every time I read an article claiming that the guild of human translators will soon have to submit to the terrible, swift sword of some new technology, I feel the need to independently verify the claims. This is due to a combination of factors, including my long-held conviction that it’s critical to fight exaggerated claims, my sense of terror that this nightmare may be just around the corner, and a desire to reassure myself that it’s not just around the corner. After reading about how “deep learning” has enabled the outdated concept of artificial neural networks, which was recently embraced by a division of Google called Google Brain, to produce a new category of software, I made the decision to try out the most recent version of Google Translate, which claims to have revolutionised machine translation. Was it a game-changer for the ancient games of chess and go, like Deep Blue and AlphaGo were?
I discovered that although though Google Translate’s more recent iteration can handle a huge variety of languages, at the time it only supported nine of them. (It has been increased to 96.) * I therefore restricted my research to the languages of English, French, German, and Chinese.
I should point out that this situation takes use of the ambiguity in the adjective “deep” before revealing my findings. One cannot help but interpret the word “deep” to mean “profound,” and as a result “powerful,” “insightful,” and “smart” when one learns that Google purchased a business called DeepMind whose products feature “deep neural networks” boosted by “deep learning.” However, the term “deep” here only refers to the fact that these neural networks have more layers (12, for example) than Earlier networks might only contain two or three. But does that level of profundity entail that everything such a network accomplishes must be significant? Hardly. Verbal spinmeistery, this is. Does this level of profundity imply that everything a network does must be profound? Hardly. Verbal spinmeistery, this is.
With the publicity surrounding Google Translate, I am quite cautious of it. But despite my dislike, I am able to identify some startling facts regarding my pet peeve. It will translate text from any of about 100 languages into any other language and is available to everyone on Earth for free. That is very demeaning. How much more pleased should Google Translate be if it can call itself “bai-lingual” (bai is Mandarin for “100”)? If I am proud to call myself “pi-lingual,” which means the total of all my fractional languages is a little bit more than 3, how much more so should it be? A bai-lingual person finds pi-lingualism to be really amazing. Furthermore, if I A few seconds will pass after I copy and paste a page of text in Language A into Google Translate before I receive a page of text in Language B. And in dozens of different languages, this is always happening on televisions all around the world.
In spite of the undeniable practical value of Google Translate and related technologies—which is probably a good thing overall—there is still something fundamentally lacking in the strategy. That something is understanding. Language comprehension has never been a priority for machine translation. Instead, the field has consistently attempted to “decode”—to avoid concern over what understanding and meaning are. Could it be that effective translation can occur without understanding? Could a being—human or machine—perform a high-quality translation without considering the nature of language? I’ll now move to the studies I ran to help shed some light on this subject.
I started my research very modestly with the following brief observation, which, in a human mind,evokes a certain situation:
Everything in their home is paired. There is her library and his library, as well as his and her vehicles, towels, and linens.
The translation problem appears simple, but in French (and other Romance languages), the pronouns “his” and “her” refer to the possession itself rather than the owner. What Google Translate offered me was as follows:
Everything enters their home in pairs. There is his car and his car, his serviettes and his serviettes, his library, and the siens.
The software fell for my ploy, failing to recognise that I was describing a couple and highlighting the fact that for each thing he possessed, she had a comparable one. For instance, the deep-learning engine used the same word, sa, for both “his car” and “her car,” making it impossible to determine the gender of any car owner. The same was true for “his towels” and “her towels,” and in the third instance involving the two libraries, “his” and “hers,” it was confused by the final s and somehow thought that it signified a plural (“les siennes”). The French statement from Google Translate completely missed the message.