The history of machine translation in the 20th century needs to be rewritten

AI page

En français :

• Histoire actualisée de la traduction automatique (17 mars 2017)
• Réécrire l'histoire de la traduction automatique (21 octobre 2019)
• Histoire de la traduction automatique à base de règles (1er septembre 2021)

In italiano:

• Federico Pucci: riscrivere la storia della traduzione automatica nel XX° secolo (15 marzo 2020)

*

In light of the recent discovery of a new key figure in the history of machine translation, I thought it would be useful to update the "timeline" of this field, organized around the two ages of machine translation (MT):

I. The Iron Age: From Prehistory to the 20th Century – Before the Web
II. The Golden Age: 20th and 21st Centuries – After the Web

The structure will be as follows:

I. From Prehistory to the 20th Century – Before the Web 

Three major stages:

1. The 17th Century 
2. The 1930s: The Pioneers 
3. The Following Five Decades 

* 

1. The 17th Century

The "prehistory" of machine translation is primarily marked by two names: René Descartes and Gottfried Wilhelm Leibniz, who laid some of its conceptual foundations.

According to John Hutchins and Harold L. Somers, Descartes and Leibniz speculated on creating mechanical dictionaries using universal numerical codes (« Both Descartes and Leibniz speculated on the creation of dictionaries based on universal numerical codes », in An introduction to machine translation).

Descartes elaborates on the invention of a universal language in his correspondence:

For a language to truly be universal, it must arise from "true" philosophy and thus proceed from a reform that transposes into thoughts the same simple and natural order that exists among numbers. Thoughts would then become clear and simple, making it "almost impossible" to err. The first step, Descartes specifies, is not to invent the primitive words and characters of the universal language, nor to ensure rapid learning times, but to establish "an order among all the thoughts that can enter the human mind, just as there is a naturally established order among numbers." One could then invent "words" and arrange them as one arranges invented languages to represent numbers and as one learns "in a single day to name all numbers up to infinity and write them in an unknown language, which nonetheless consists of an infinity of different words," and "do the same for all other words necessary to express all other things that come to the minds of men." Thus, a true universal language would emerge, as it would be capable of representing ordered thoughts in the human mind, the simple ideas. Such a language would "soon spread throughout the world," and many would be willing to spend "five or six days" to make themselves understood by all people.

A universal language can therefore only emerge after ordering, distinguishing, and enumerating human thoughts to make them clear and simple. This is "the greatest secret one can have for acquiring true knowledge." Based on the knowledge of "simple ideas," such a language would become easy to learn, pronounce, and write: "And if someone had clearly explained what the simple ideas in the human imagination are, from which all that they think is composed, and if this were accepted by everyone, I would dare to hope for a universal language that is very easy to learn, pronounce, and write, and most importantly, that would aid judgment by representing all things so distinctly that it would be almost impossible to err."

A universal language is thus a language of ordered thoughts, but also of clear and simple thoughts. In contrast, the words available to humans only have confused meanings, which explains why almost nothing is perfectly understood.

Source: Letter to Mersenne November 20, 1629, B 24, pp. 92–97.. “This letter has been studied in Cartesian critical literature, particularly in relation to the project of an artificial language, sometimes even seen as a precursor to Leibniz’s universal characteristic…” in DESCARTES : TRADUCTION, VÉRITÉ ET LANGUE UNIVERSELLE
Giulia Belgioioso (University of Lecce)

*

Let us now move from the early 1930s to the Web, that is, from Federico Pucci’s first "mechanical translator" to modern "neural machine translation" (see a comparison here...).

2. The 1930s: The Pioneers 

With Federico Pucci, the history of machine translation in the 20^th century now needs to be rewritten from the 1930s onwards:

1929 (December): Federico Pucci presents his study on the “mechanical translator” for the first time in Salerno.

1930: Federico Pucci’s participation in the first National After-Work Arts and Crafts Exhibition of Bolzano – literary section, with his concept of “mechanical translator”, awarded a silver medal.

1931: Federico Pucci publishes in Salerno the first part of what we might consider to be the first book ever published anywhere on a “mechanical translating device”, called: "“The mechanical translator and the method for Europeans to correspond, knowing only their own language: Part I: Translating from foreign language).”"

1932: likely construction of a prototype “translating machine” by Georges Artsrouni, later destroyed. No document has been kept about it, except for a photograph that makes a description impossible. (Source)

1932: Warren Weaver becomes director of the Rockefeller Foundation.

1933: filing of patent and presentation to Soviet authorities of Petr Petrovič Smirnov-Trojanskij’s machine, probably at the design and description stage. (Source)

1933-1935: construction of Georges Artsrouni’s “mechanical brain”:

1935: presentation of Federico Pucci’s “mechanical translator” at the Inventors Competition, part of the Trade Fair of Paris, receiving a silver medal for a “a method for translating languages without knowing them”! (Source)

1937: Georges Artsrouni presents some machines at the National Exhibition of Paris, the principle of which received a Grand Prix award for mechanical data processing, according to the inventor himself.

1939-1945 : World War Two

➽ Federico Pucci’s publishing activity is interrupted between 1931 and 1949, a time corresponding to the pre-war, war and post-war periods, during which little is known about Federico Pucci, apart from his participation in some Exhibitions and his work as a censor, about which he writes:

Then the war came, and I attempted to steer my studies towards a military use. I managed to create mechanical translating devices “C” and “D”, a mechanical solution, attempting to create a new mechanical-based language, with device C working as a transmitter, and D as a receiver device. They were to be submitted to the 1940 Engineering Exhibition, but the War Ministry opposed its participation. I was called to Rome to explain the invention. It was approved, and I was authorised to build and try out the device, at the State’s expense, since I had informed them, I could not afford to build it on my own. Obviously, I was obliged to keep everything secret. However, as I was not a mechanic, I thought that I would need the assistance of other persons, who might not be able to keep the secret. I did not want to run this risk, so I turned down the assignment, and left the invention in the hands of the War Ministry, so that it might do whatever it wanted with the idea.

This marks a big gap in our story that it would be very interesting to fill...

*

3. The Following Five Decades 

•  The First Decade (≅1945-1955) : The first steps
•  The Second Decade (≅1955-1965) : From enthusiasm to disappointment 
•  The Third Decade (≅1965-1975) : The quiet period 
•  The Fourth Decade (≅1975-1985) : The revival
•  The Fifth Decade (≅1985-1995) : Maturity 

There is no need to repeat here the developments of machine translation according to the timeline proposed in 1994 by Jacques Anis in dans « Ordinateurs et traduction : survol d'un demi-siècle » [In: Langages, 28ᵉ année, n°116, 1994. Le traducteur et l'ordinateur. pp. 111-122; doi : 10.3406/lgge.1994.1699]. I refer the reader to the source document.

It should be noted that, according to the author, his work was primarily based on John Hutchins’ 1986 book, “Machine Translation: Past Present Future”, the same researcher in whose work I first found mention of Federico Pucci. Yet, Pucci wrote at least 12 books on languages over 35 years, 7 of which focused on the "(dynamo-)mechanical translator" from 1931 to 1958, and so far, there has apparently been no trace anywhere of either the inventor or his inventions, despite his alleged participation in the Lépine Contest! These are mysteries I hope to unravel...

The fifth decade finally overlaps with the advent of the World Wide Web, starting in 1990, a year sometimes considered a turning point for the renewal of machine translation.

*

II. 20th and 21st Centuries – After the Web 

I still need to develop this section, which is undoubtedly the richest (and thus will require time, although I have already laid the first groundwork), likely structured as follows:

1. The Decade 1995–2005
2. From 2006 to 2010

These 5 years, of course, coincide with the maturity of the Google-machine translation duo.

       3. Neural Networks (2010–2020)

This era marked the rise of neural machine translation (NMT). Encoder-decoder architectures, like those introduced by Sutskever et al. (2014), became foundational, encoding input text into a fixed representation and decoding it into the target language. The introduction of the attention mechanism (Bahdanau et al., 2015) allowed models to focus on relevant parts of the input, improving translation quality for longer sentences. The Transformer model (Vaswani et al., 2017) revolutionized NMT with its self-attention mechanism, enabling parallel processing and better handling of long-range dependencies, leading to more accurate and fluent translations.

        4. Generative AI (2020–present)

Large language models (LLMs) like GPT and BERT derivatives have pushed machine translation further. These models, trained on vast datasets, exhibit emergent capabilities, such as zero-shot translation, where they translate without specific training for certain language pairs. Their ability to understand context and generate human-like text has improved translation quality, especially for low-resource languages, while also enabling tasks like multilingual dialogue and cross-lingual knowledge transfer.

        5. Likely Evolutions

Looking ahead, foreseeable developments in machine translation and related fields revolve around several key axes:

• Continued Model Scaling: The ongoing increase in the scale of models (e.g., larger language models with billions of parameters) is expected to lead to the emergence of new capabilities, such as improved translation accuracy, better handling of nuanced contexts, and enhanced zero-shot or few-shot performance across diverse languages.
• Architectural Optimization: Innovations like Mixture of Experts (MoE), sparse attention mechanisms, and retrieval-augmented generation (RAG) will enable more efficient management of computational resources. These approaches reduce the computational cost of processing large models while maintaining or improving performance, making translation systems faster and more scalable.
• Multimodal Integration: Incorporating multimodal data (e.g., text, images, audio) will enhance contextual understanding. For example, translation systems could leverage visual or auditory cues to disambiguate meanings, improving accuracy in context-sensitive translations.

To be continued...

*

This evolution in less than a century reflects a shift from rule-based and statistical methods to neural and generative approaches, drastically improving translation accuracy and versatility. Anyway, computing power and IT resources have been critical enablers, arguably accounting for 50–70% of MT's progress. Without GPUs, TPUs, and large-scale data infrastructure, neural and generative models would be infeasible due to their computational demands. However, algorithmic innovations (e.g., attention, Transformers) and curated datasets are equally crucial, as raw computing power alone cannot achieve high-quality translations without sophisticated architectures and training strategies. The synergy of hardware advancements and algorithmic breakthroughs has driven the drastic improvements in machine translation.

I wish to conclude this post by calling for the intervention of a University or any Authority in the field of Machine Translation, in order to highlight the unique role played by Federico Pucci in the history of MT, and to realize his dream of building prototypes of his manifold “translating machines”, for which he himself provides all the elements needed in his books.

In the final analysis, apart from the well-known machines of Georges Artsrouni and Petr Petrovič Smirnov-Trojanskij, which in fact have never had any practical implications in the field, John Hutchins dates the nascent years of machine translation back to the time 1947-1954.

So we can assert without fear of being denied that Mr Federico Pucci is the very first precursor of machine translation as we know it today!

This work is licensed under a
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.