Machine translation
Based on Wikipedia: Machine translation
The Ninety-Ten Problem
Here is a puzzle that has haunted translators and computer scientists for seventy years: "Japanese prisoners of war camp."
Read that phrase again. Is it an American camp holding Japanese prisoners? Or a Japanese camp holding American prisoners? In English, both interpretations are grammatically valid. A human translator encountering this phrase in a medical document about a World War II epidemic would need to pick up the phone and call the Australian physician who wrote it. The work might take hours.
A machine, confronted with the same ambiguity, will simply guess. And it will guess wrong roughly half the time.
This is the central tragedy of machine translation. About ninety percent of any given text consists of straightforward sentences that computers can handle reasonably well. It's the other ten percent—the ambiguous phrases, the cultural references, the subtle ironies—that consumes most of a human translator's time. And it's precisely this ten percent that machines handle worst.
Dreams of a Universal Language
The idea of automated translation is older than you might think. In 1629, the French philosopher René Descartes proposed creating a universal symbolic language where each concept would have exactly one representation, shared across all human tongues. If "tree" in English and "arbre" in French and "Baum" in German all mapped to the same symbol, translation would become trivial—just symbol lookup.
Descartes was working in a long tradition. Eight centuries earlier, the ninth-century Arabic scholar Al-Kindi had developed techniques for systematically analyzing and transforming languages. His work on cryptography—breaking secret codes—required him to study how languages function mathematically. He pioneered frequency analysis, the observation that certain letters appear more often than others in any given language. This same insight would eventually become foundational to modern machine translation.
But Descartes and Al-Kindi were working with pen and paper. The real story of machine translation begins in 1947, with the invention of electronic computers.
The Cold War's Unexpected Gift
In the years following World War II, two researchers independently proposed the same audacious idea. Andrew Donald Booth in England and Warren Weaver at the Rockefeller Foundation in America both suggested that these new electronic calculating machines might be taught to translate between human languages.
Weaver's 1949 memorandum became perhaps the most influential document in the early history of machine translation. He wrote with the optimism characteristic of postwar American science: if computers could break enemy codes during the war, surely they could handle the relatively simpler task of translation?
They could not. But nobody knew that yet.
The 1950s saw a flurry of activity. Yehoshua Bar-Hillel at the Massachusetts Institute of Technology became the field's first dedicated researcher in 1951. That same year, a team at Georgetown University began work on what would become the first public demonstration of machine translation. Japan and the Soviet Union launched their own programs in 1955. London hosted the first international conference on machine translation in 1956.
The most celebrated moment came in January 1954, when Georgetown University partnered with International Business Machines to demonstrate their translation system. The machine translated sixty Russian sentences into English. Headlines proclaimed a new era. One newspaper announced that a "thinking machine" had mastered foreign languages.
It had not.
The ALPAC Report: A Reckoning
By the early 1960s, enthusiasm was giving way to frustration. The systems being developed could handle simple, carefully constructed sentences. But throw real-world text at them—newspaper articles, technical manuals, literature—and they produced gibberish.
In 1964, the United States National Academy of Sciences formed the Automatic Language Processing Advisory Committee, known by the acronym ALPAC. Their task was simple: evaluate a decade of machine translation research.
The 1966 ALPAC report was devastating. After ten years and millions of dollars, the committee concluded, machine translation remained essentially useless for practical purposes. The technology could not match human translators in quality, and the projected cost savings had not materialized. Funding dried up almost overnight.
The report killed machine translation as a mainstream research field for nearly two decades.
Survival in the Wilderness
Yet the work continued, quietly, in unexpected places.
The United States military never stopped caring about translation. During the Vietnam War, a system called Logos successfully translated technical manuals into Vietnamese. The quality was adequate for soldiers who needed to understand equipment specifications, even if it would never win literary prizes.
The French Textile Institute began using machine translation in 1970 to process abstracts in French, English, German, and Spanish. Brigham Young University started a project in 1971 to translate Mormon religious texts automatically. These were narrow applications with limited vocabularies and predictable sentence structures—exactly the conditions under which early machine translation could succeed.
The company that would dominate the field for decades, SYSTRAN, emerged from contracts with the United States government in the 1960s. By 1978, Xerox was using SYSTRAN to translate technical manuals. The system was slow, expensive, and required extensive post-editing by humans. But it worked well enough to be commercially viable.
Two Philosophies of Translation
Early machine translation systems broadly followed two approaches, each with its own elegant logic and fatal flaw.
The first approach was rule-based. Researchers would study a language pair—say, French and English—and encode everything they knew about how the languages worked. Grammar rules. Vocabulary mappings. Exceptions to those exceptions. The resulting systems contained tens of thousands of hand-crafted rules.
The problem was that natural language is, to use a technical term, absurdly complicated. Every rule has exceptions. Every exception has its own exceptions. And speakers constantly create new expressions, puns, and idioms that no rule book could anticipate. Rule-based systems required armies of linguists working for years, and they still couldn't handle casual speech or creative writing.
The second approach was statistical. Instead of encoding rules about how languages work, why not analyze millions of sentences that humans had already translated? If "maison" usually becomes "house" in translated documents, the system could learn that pattern without any explicit rules.
Statistical machine translation got a major boost from an unlikely source: the Canadian parliament. Because Canada is officially bilingual, every word spoken in parliamentary debates is transcribed in both English and French. This created a massive corpus of parallel texts—the same content expressed in two languages, aligned sentence by sentence. Researchers called it the Hansard corpus, after the British tradition of recording parliamentary proceedings.
The European Parliament provided an even larger resource. Every document produced by the European Union exists in multiple languages, creating what researchers call EUROPARL—millions of aligned sentences across more than twenty languages.
The first major statistical system was CANDIDE, developed by International Business Machines. The results were promising but not revolutionary. Statistical methods required enormous amounts of parallel text, and such corpora simply didn't exist for most language pairs. What if you wanted to translate between Hungarian and Vietnamese?
The Google Moment
In 2005, Google did something that changed everything. The company trained its internal translation system on approximately two hundred billion words harvested from United Nations documents. The scale was unprecedented. And the results, while still imperfect, represented a quantum leap in quality.
Two years later, in 2007, researchers released MOSES, an open-source statistical translation engine. For the first time, anyone could build and experiment with machine translation systems without massive corporate resources.
By 2012, Google announced that its translation service was processing enough text daily to fill one million books. The technology had escaped the laboratory.
The Neural Revolution
Then came deep learning.
Neural machine translation, which emerged in the mid-2010s, represented a fundamental shift in how computers approach language. Instead of learning statistical patterns between words, neural networks learn to represent entire sentences as points in a mathematical space. Translation becomes a matter of finding the corresponding point in the target language's space.
The results were immediately and dramatically better. Sentences that sounded stilted and mechanical under statistical methods suddenly flowed more naturally. The systems could handle longer-range dependencies—references and pronouns that pointed back to earlier parts of a text.
A German company called DeepL Translator emerged in 2017 and quickly gained a reputation for producing the best machine translations available. As of the early 2020s, many professional translators consider DeepL their first choice for generating rough drafts that they then polish.
But neural translation has not solved the fundamental problem. The ninety-ten split remains. Machines handle routine text adequately; they stumble on ambiguity, context, and nuance.
Large Language Models Enter the Arena
The most recent development is the application of large language models—systems like the Generative Pre-trained Transformer, commonly known as GPT—to translation tasks. These models are not specifically trained for translation. They learn language in general, absorbing patterns from billions of words of text across many languages and domains.
When you prompt such a model with "Translate the following French text into English," it can often produce serviceable results. The approach is promising precisely because these models have such broad knowledge. They can draw on context and world knowledge in ways that specialized translation systems cannot.
But the approach is expensive. Running a large language model requires significant computational resources. For high-volume translation work, specialized systems remain more practical.
And the quality problems persist. Studies comparing translations produced by systems like ChatGPT against those produced by human professionals consistently find that humans outperform machines on terminological accuracy—getting the technical vocabulary right—and clarity of expression.
What Machines Cannot Do
Claude Piron spent decades as a translator for the United Nations and the World Health Organization. He understood, from long experience, exactly where machine translation succeeds and fails.
Why does a translator need a whole workday to translate five pages, and not an hour or two? About ninety percent of an average text corresponds to simple conditions. But unfortunately, there's the other ten percent. It's that part that requires six more hours of work. There are ambiguities one has to resolve.
Piron cited the "Japanese prisoners of war camp" example. The ambiguity is invisible to a machine—both readings are grammatically valid English. Resolving it requires research, possibly a phone call to the author on another continent. No translation algorithm can make that phone call.
If a machine simply guesses—perhaps based on which interpretation appears more often in its training data—it will be wrong with uncomfortable frequency. If it instead asks the human operator to resolve every ambiguity, it automates only about twenty-five percent of the work. The hard seventy-five percent remains human labor.
The Named Entity Problem
Consider the sentence: "Smith is the president of Fabrionix."
Both "Smith" and "Fabrionix" are what linguists call named entities—proper nouns referring to specific people, organizations, or places in the real world. They behave differently from ordinary words.
Ordinary words should be translated. The English word "president" becomes "président" in French or "Präsident" in German. But "Smith" should remain "Smith" regardless of the target language. Names are transliterated—converted letter by letter—not translated.
Except when they're not. Consider "Southern California." The word "Southern" should be translated—"Californie du Sud" in French. The word "California" should remain as is, though it might be respelled following the target language's conventions—"Californie" rather than "California."
Machines frequently get this wrong. They treat "Southern California" as a single unit and transliterate the whole thing, producing nonsense like "Sauzèrne Kalifôrnia." Or they translate both words, which might produce something interpretable but wrong—suggesting a place called "Midi Californie" that doesn't exist.
Even when systems include specific lists of names that should not be translated, they first must correctly identify those names in the text. Miss the identification, and the error cascades.
The Vernacular Gap
Machine translation systems learn from parallel texts. And the vast majority of such texts come from formal sources: government documents, corporate materials, international organization proceedings.
This creates a profound bias. Machines learn to translate the way diplomats and bureaucrats write. They do not learn to translate the way people actually speak.
Slang, dialects, regional expressions, internet-speak, text message abbreviations—all of these fall outside the training data. A system trained on United Nations documents will struggle with a teenager's casual messages.
This limitation matters especially for mobile applications. People don't write formal prose on their phones. They dash off quick messages full of abbreviations, missing punctuation, and casual phrasing. Machine translation handles this input poorly.
The Parity Illusion
In recent years, researchers have occasionally announced that machine translation has achieved "human parity"—performance equivalent to professional human translators. These claims should be treated with skepticism.
The studies demonstrating parity have consistently involved narrow conditions: specific language pairs, specific text domains, specific evaluation methods. A system that matches human quality on news articles may fail spectacularly on poetry. A system that handles German-to-English well may struggle with Korean-to-Arabic.
The evaluations themselves are often problematic. The standard automated measure, called BLEU (Bilingual Evaluation Understudy), compares machine output against reference translations. But BLEU doesn't understand meaning. A translation that mangles a named entity—rendering "George Washington" as "George Washing Machine"—might score well on BLEU while being obviously wrong to any human reader.
When professional literary translators evaluate machine output, they consistently find problems. The errors may be subtle—a slightly wrong connotation, a missed cultural reference, a phrase that sounds unnatural to native speakers—but they are pervasive.
The Future of Human Translators
So where does this leave the human translator?
The profession is changing, but it is not disappearing. The most common workflow today involves machine translation producing a first draft that human translators then edit and refine. This is called post-editing, and it has become a standard practice in the industry.
Post-editing is faster than translating from scratch. But it requires different skills. A traditional translator reads the source text and creates the target text. A post-editor reads both the source and the machine's attempt, then fixes the machine's mistakes while preserving what it got right. The work is cognitively demanding in a different way.
For some types of text, machines have indeed taken over. High-volume, low-stakes content—product descriptions, technical specifications, routine business correspondence—can be handled by machines with minimal human oversight. The results are good enough for the purpose.
But for anything requiring precision, nuance, or creative judgment, humans remain essential. Legal contracts. Literary works. Marketing campaigns. Medical communications. Anything where a mistake could cost money, reputation, or lives.
The ninety-ten split endures. Machines have gotten better at the ninety percent. The ten percent remains stubbornly human.
Deep Versus Shallow Understanding
Yehoshua Bar-Hillel, who pioneered machine translation research at MIT in the 1950s, saw the fundamental problem clearly. He argued that without a "universal encyclopedia"—complete knowledge of the world—a machine would never reliably disambiguate between multiple meanings of a word.
Consider the English word "bank." It might refer to a financial institution, the edge of a river, a airplane's turning maneuver, or several other concepts. Humans effortlessly pick the right meaning based on context. We know that rivers have banks and money goes into banks, that pilots bank their aircraft and pool players bank their shots.
Researchers distinguish between "shallow" and "deep" approaches to this disambiguation problem. Shallow approaches simply look at the surrounding words. If "bank" appears near "river" and "water," it probably means a riverbank. If it appears near "money" and "account," it probably means a financial institution. This works surprisingly well for common cases.
Deep approaches would require the system to actually understand what banks are—to know that rivers flow between banks, that fish swim near banks, that people sit on banks during picnics. This level of understanding remains beyond current technology.
Modern large language models fall somewhere in between. They have absorbed so much text that they can mimic understanding. But whether they truly understand, or merely perform sophisticated pattern matching, remains one of the deepest questions in artificial intelligence research.
A Seventy-Year Project
Machine translation began in 1947 with Warren Weaver's optimistic memorandum. More than seven decades later, the dream of fully automatic, high-quality translation remains unfulfilled.
This is not a story of failure. The technology works vastly better than it did even a decade ago. Travelers can point their phones at foreign signs and get approximate meanings. Businesses can process foreign-language documents at a fraction of the cost of professional translation. Researchers can get the gist of papers published in languages they don't speak.
But it is a story of humility. Natural language turned out to be harder than the early researchers imagined. The ambiguities, the contextual dependencies, the cultural knowledge embedded in every utterance—these resist computational approaches in ways that other hard problems, like chess and Go, did not.
Perhaps the most honest assessment comes from considering not what machines can do, but what humans can do that machines cannot. A human translator can pick up the phone. A human translator can ask the author what they meant. A human translator can research the historical context, consult subject-matter experts, agonize over word choice until the result feels right.
These capabilities—curiosity, persistence, judgment, taste—are not yet within reach of machines. And until they are, human translators will remain essential partners in bridging the world's languages.