ROME — "Wo zui xihuan de shiwu shi shousi, my favorite food is sushi..."
In a recent video, U.S.-based journalist Louise Matsakis can be seen and heard expressing herself in perfect Mandarin. Having only been studying Chinese for a few years, Matsakis is still far from fluent. But in the video, she pronounces every syllable flawlessly and in the right tone, without errors or awkward pauses, just as a native speaker would. The voice was soft but also "slightly alien," she herself acknowledges in an article last month in The Atlantic.
Matsakis had used the HeyGen software, a Los Angeles startup that makes it possible to create deepfake videos, that is, to use artificial intelligence to make real people say almost anything. All it takes is to upload a picture of one's face and some text, which is then matched with an artificial voice and can be translated into more than 40 languages. Matsakis writes that the tool works so well she wonders if all her efforts at learning Mandarin were a waste.
Automatic translation was not always so convincing. The early tools (Google Translate is from 2006) were rather poor, only able to give a general idea of, for instance, of a French or Portuguese website. In 2010, in the Netherlands, a subpoena translated from Dutch to Russian using Translate instructed a defendant not to show up in court when he should have gone. The big leap forward came in 2015, when Baidu (China's leading search engine) put its large-scale neural machine translation service into operation. In just a few years, neural networks, the machine learning systems behind programs like ChatGPT, have improved the quality of machine-made translation, making it significantly more reliable.
For the latest news & views from every corner of the world, Worldcrunch Today is the only truly international newsletter. Sign up here.
Such progress, however, is accompanied in some countries by another phenomenon: a collapse in the number of students taking up foreign languages. In Australia in 2021 only 8.6% of high school seniors had chosen to learn another language, a record low. In South Korea and New Zealand, universities are closing French, German and Italian departments. At U.S. colleges between 2009 and 2021, enrollment in non-English language courses declined by 29.3%, while it had grown steadily in the previous 30 years.
Last September, after a heated debate, West Virginia university decided to eliminate the foreign languages and literatures department, replacing it with an online app. In some cases, even English proficiency is declining: in France, half of children finishing secondary school do not reach the minimum level required (A2) by the Common European Framework of Reference for Languages, a system developed by the Council of Europe.
Beyond the factors that could explain this disaffection -- the pandemic that has fragmented education at all levels, the cuts suffered by the disciplines of the humanities -- it is clear that the younger generation is giving up on learning languages just as machine translation becomes ubiquitous on the Internet (in social media apps, messaging services, streaming platforms).
In the very near future, it may enter the daily lives of billions of people.
Limits of technology
Technology of course has limitations. It works well with English, Mandarin, Arabic and French -- for which there are huge amounts of digitized texts and recorded speeches (almost always transcribed) -- and less well with languages with little online presence such as Swahili and Urdu.
According to most studies in which native speakers have been involved, it does not do well when confronted with literature either: it offers acceptable translations of only about 30% of excerpts from novels (usually simple passages). This is because it is less adept at finding creative solutions that preserve aspects of a book that are difficult to quantify, such as style, rhythm, wit, sensibility.
Teachers' attention should shift from grammar exercises to understanding practices and cultures.
In addition, it may be questionable to employ technology in sensitive situations, for example, to translate interviews of asylum seekers and testimonies coming from conflict zones.
There is also a risk of exaggerating its successes. To promote itself on social media, Jumpspeak, a language learning app similar to Duolingo and Babbel, has made advertisements in which an AI-generated person reads computer-translated sentences: "I struggled with languages all my life. Then I learned Spanish in six months. I was offered a job in France and learned French. I learned Mandarin before I visited China." With each sentence, it switches from one language to another.
Leaving aside the weaknesses and far-fetched promises, we nonetheless can take it for granted that machine translators will far exceed the technical skills of the average language graduate. Then the conclusion shared by many experts in education and linguistics is that teachers' attention should shift from grammar exercises to understanding the practices and cultures embedded in various contexts.
Angles of slang
Valerie Trapp, another Atlantic reporter, says her approach was roughly this. Spanish was the first language she learned, spending her early childhood in the Dominican Republic, but when she moved to the United States as a child, she practically abandoned it for English.
Back in Santo Domingo as an adult, she found that her Spanish sounded "a little outdated" and that she struggled to express a personality with her peers. To catch up, she dived into slang and self-imposed the study of reggaeton. Quickly, she saw a new level of comfort and colloquialism in her Spanish that she might not get from an app. "Language is not a rigid algorithm to decode. It is fluid, changeable and heterogeneous."
Slang connects us to the current version of a culture.
The creative intricacies of slang challenge machine translation models but uniquely attract our interest. In the essay Slang: The People's Poetry Michael Adams, professor of English at Indiana University, cites a study that measured the brain activity of some people as they read Shakespeare's Coriolanus. The researchers found that when language was used in the play in a more original way, for example if a noun was turned into a verb, the brain became more activated. This expedient falls under what linguists call functional shifting, and is also how slang works.
Slang, in other words, can connect us to a current, if impermanent, version of a culture. For others who may be confronted with learning their family's language of origin, the experience helps to discover unexplored parts of themselves. But all of us should remember what may be most valuable about learning a foreign language is simply how much fun it is.
