How AI Translation Began with Machine Translation Technology in the 1970s

Machine Translation is the foundation of today’s AI translation and emerged as early as the 1940s. In this reflective piece, Ralph Strozza explores how the technology took shape, how it influenced the evolution of AI in the translation industry, and how it intersected with his own journey as a language professional.

The Machine Translation Landscape of the 1970s and 1980s

Many people think that the research and development into Machine Translation (MT) systems began with the onset of Google Translate in 2006. They may be surprised to learn that MT was sometimes referred to as Computer-Assisted Translation (CAT) so as not to scare human translators (the term “AI translation” might spook them as well😉). Research and development for CAT tools actually began in the mid-20th century, shortly after World War II.

Today’s AI translation is powered by Large Language Models (LLMs) and deep neural networks. But decades of MT research and experimentation contributed to today’s AI translation models. While most people associate AI translation with recent advances from companies like Google or OpenAI, the roots of these breakthroughs reach back to the early pioneers of MT. Ralph Strozza witnessed this evolution firsthand, beginning his career during a pivotal era when the translation industry began merging human expertise with computational power.

Here’s a brief timeline of key milestones:

A visual timeline charting key milestones in the evolution from human-only translation to AI-powered translation. Highlights include the conceptual birth of MT in 1949, the first public demo in 1954, the ALPAC report in 1966, the rise of rule-based and statistical and neural MT, and the emergence of large language models (LLMs) by 2023. The timeline illustrates the transition from rule-based methods to today’s hybrid AI models.

The Birth of Machine Translation: When Computers First Learned to Translate (1940s–1950s)

• 1947–1949: An early draft of the idea of using computers for translation was proposed by Warren Weaver, a mathematician and one of the pioneers of information theory. In 1949, he wrote a famous memorandum suggesting that cryptographic techniques used during WWII could be applied to language translation.
• 1954: The Georgetown-IBM experiment marked the first public demonstration of MT. It translated 60 Russian sentences into English using a rule-based system. Though limited, it generated significant optimism.

Rule-Based Systems Era (1950s–1980s)

• MT systems during this period were rule-based, relying on linguistic rules and dictionaries. These systems were labor-intensive and struggled with ambiguity and idiomatic expressions.
• The ALPAC Report (1966), commissioned by the U.S. government, concluded that MT had not met expectations, leading to a significant reduction in funding in the U.S. for over a decade.
• Statistical Machine Translation (SMT) (1990s–2010s)
• In the late 1980s and early 1990s, rule-based and statistical MT  emerged, pioneered by IBM researchers and several privately-owned companies. SMT used large bilingual corpora to learn translation probabilities.
• This approach led to more scalable and data-driven systems, such as Google Translate, which initially used SMT.

Neural Machine Translation (NMT) (2014–present)

• Around 2014, Neural Machine Translation began to outperform SMT. NMT uses deep learning models, particularly sequence-to-sequence architectures with attention mechanisms.
• Google, Microsoft, Facebook, and others adopted NMT, leading to significant improvements in fluency and accuracy.
• The 1970s and 1980s marked a transformative era in the history of MT. Despite early skepticism (especially following the 1966 ALPAC report, which criticized MT’s progress), several companies and institutions forged ahead, developing systems that laid the groundwork for today’s translation technologies.

The Lasting Legacy: How Early Machine Translation Systems Inspired AI Translation

The MT systems of the 1970s and 1980s introduced foundational concepts such as building language models, controlled language input, creating structured dictionaries, and iterating based on human feedback with domain-specific translation engines. These innovations paved the way for statistical and neural MT systems that emerged in the 1990s and beyond.

What was once rule-based and labor-intensive has evolved into probabilistic, data-driven, and now neural systems capable of learning patterns autonomously. In essence, today’s AI translation is the direct descendant of those early MT frameworks.

Ralph’s Experience from Lexicographer to Language Technologist During MT’s Rise

While a graduate student in French and Italian at Northwestern University, I was contacted by Weidner Communications Corporation (who later rebranded as WCC – Worldwide Communications Corporation) to work part-time as a lexicographer. Basically, my job was to input French and Italian terminology into the MT dictionaries, without which the software was basically useless.

Ralph Strozza is Interpro’s Founder and Consigliere, leading Interpro from the early days of Machine Translation while adopting localization workflows to adapt to technology advances in eLearning authoring tools, software, websites, and more.

As this was a rule-based system, specific grammatical information was required to enable the MT software to generate as correct a translation as possible. Data related to nouns required gender (masculine, feminine, neuter (in the case of German), and different models were set up in the system to select the plural versions of the nouns. So, if inputting the term water in French (eau), the model for the plural version would have been something like “bateau” (the plural of eau is eaux; the plural of bateau is bateaux). The lexicographer would thus assign the model “bateau” to “eau”, so that the system would generate the correct plural form for “eau”. The same concept applied to verbs, adjectives, and other parts of speech.

Obviously, this meant that whoever was inputting terms into the MT dictionary needed a very good grasp of the target language grammar. If a term was input incorrectly, the MT system would consistently generate an incorrect translation, very unforgiving.

As a lexicographer, after inputting a glossary of terms, I would run content containing those terms through the MT system, analyze what needed to be tweaked, fix it, and then run the content through again to make sure the error was fixed. If fixed, it was fixed forever.

As mentioned previously, this was an extremely labor-intensive process, and our clients were warned upfront that for MT to fulfill its promise, there was a lot of up-front work to be done.

From the onset of MT R&D in the mid-1950s, the primary objectives were to:

• Improve the speed and scalability of producing usable translation
• Reduce the cost of translation
• Making information accessible to non-native speakers (we call this “inclusion” now)

Sound familiar? What companies and organizations are looking to MT to deliver in this first quarter of the 21st century are basically identical to what companies in the latter half of the past century sought.

So while the technology has evolved beyond what 20th-century MT developers could have only dreamed of, the end game is basically the same.

Need help implementing AI translation? Interpro’s experts are here to help. 

Major MT Developers of the Era

Technology and adoption rates depend on the teams embracing the technology. Here are some of the companies and milestones that strongly impacted MT’s rise.

SYSTRAN

Founded in 1968 by Dr. Peter Toma, SYSTRAN was a pioneer in rule-based MT. It gained early traction by translating Russian-English texts for the U.S. Air Force during the Cold War. In 1976, the European Commission adopted SYSTRAN to manage multilingual documentation, making it one of the first large-scale governmental uses of MT.

Weidner Communications Corporation (WCC)

Launched in 1977, Weidner’s CAT software revolutionized translation productivity. It reportedly quadrupled output and halved costs. By the mid-1980s, WCC was the largest translation company in the U.S. by number of installed systems. Siemens acquired a version of the Weidner system in 1980, which later influenced the METAL MT project.

ALPS (Automated Language Processing Systems)

Based in Provo, Utah, ALPS developed interactive translation tools that blended human input with automated dictionary look-up. It was used commercially and in academia, notably at Coventry Polytechnic in the UK. ALPS emphasized human control, making it a hybrid between an MT tool.

Logos Corporation

Founded in 1970 by Bernard “Bud” Scott, Logos initially developed an English-Vietnamese system for the U.S. Department of Defense. It later expanded to support multiple language pairs, including English-German and English-French. Logos was known for its rule-based system and longevity, surviving into the early 2000s.

The Canadian METEO System: A Meteorological MT Success Story

One of the most successful domain-specific MT systems of the era was METEO, developed in Canada to translate weather forecasts between English and French. It was created in response to the Official Languages Act of 1969, which required bilingual federal communications. TAUM-METEO, the prototype, was developed in 1975–76 by the TAUM Group at the Université de Montréal. The first operational version, METEO 1, ran on a Control Data CDC 7600 supercomputer starting in 1977. In 1982, John Chandioux created Gram R, a linguistic programming language that enabled METEO to run on a Cromemco microcomputer with just 48Kb of RAM and a 5Mb hard disk.

Companies and Institutions That Adopted MT Systems

• Organizations that adopted MT systems included:
• the U.S. Air Force
• European Commission
• Siemens
• Coventry Polytechnic
• Pan American Health Organization
• Caterpillar
• Xerox
• CIA
• FBI
• Intergraph Corporation
• Michelin
• Interpol
• Occidental Petroleum
• Toyota of Canada
• Nissan of Canada and
• the US Army School of the Americas, to name just a few.

Computer Hardware That Powered MT Systems

MT systems in the 1970s and 1980s were powered by a mix of mainframes, mid-range computers, and early microcomputers. Mainframes like IBM and CDC systems supported SYSTRAN, Logos, and METAL. Weidner’s systems ran on Digital Equipment Corporation’s PDP 11/44 and VAX mid-range computers, and a microcomputer-based version ran on the IBM PC/XT. Microcomputers enabled systems like Weidner, ALPS, and METEO to become more accessible.

Leverage Today’s AI Translation Technology with Interpro

Today’s AI translation systems learn these rules automatically by analyzing massive multilingual datasets and refining their understanding over time. Yet even with all the advancements, the human element to train, guide, and apply these tools responsibly is essential.

At Interpro, we see this as the natural evolution of what began in those early decades. MT created the foundation; AI translation is the next chapter. Together, they represent a shared pursuit: empowering organizations to communicate clearly, globally, and intelligently.

Let Interpro help guide you in this next chapter of AI translation.

Explore Services

Consulting • Translation • eLearning • Video & Multimedia • Document