From Mechanical Translation to AI: Martinet’s Legacy and the Evolution of Linguistic Mediation

 Introduction

Machine translation (MT) has evolved dramatically from its early rule-based methods to today’s AI-driven models. Initially, translation systems relied on word-for-word substitution, an approach that often led to nonsensical or misleading results due to the complexity of linguistic structures. In his article "The Word," André Martinet criticized this method, arguing that treating words as discrete, self-contained units of meaning created an artificial "screen" that obscured the structural and relational nature of language.

This article argues that while contemporary AI translation has moved beyond simple lexical substitution, it still struggles to fully account for linguistic interdependence. Despite advancements in deep learning, modern algorithmic models continue to approximate rather than truly understand syntactic and morphological dependencies, making his critique as relevant as ever. By examining Martinet’s structuralist insights, the historical development of machine translation, and the promise of neuro-symbolic AI, this article explores the extent to which modern translation systems align with or diverge from Martinet’s vision of linguistic mediation.

Martinet’s Critique of Early Mechanical Translation

Martinet observed that early machine translation systems treated written words as independent units of meaning, disregarding their relational role within sentences. He warned that such an approach obscured the true nature of human language:

"Certain applications of linguistics, such as the research applying to mechanical translation, by the emphasis which they place on the written form of the language might seem to lend importance to spaces in the written text and lead us to forget that it is from speech that one should always start in order to understand the real nature of human language. The fundamental traits of human language are frequently to be found behind the screen of the word."

Martinet’s critique targeted the assumption that words could be directly mapped between languages without accounting for syntax, morphology, or pragmatics. Idiomatic expressions, verb valency, and variations in word order all illustrate the failure of a purely lexical approach. By emphasizing language as a structured system rather than a mere collection of words, he anticipated the limitations of early MT systems and the need for a more holistic approach to linguistic mediation.

The Failure of Early Machine Translation

In the 1950s and 1960s, early MT systems attempted translation by using bilingual dictionaries and rigid grammatical transformation rules. However, these models were unable to handle linguistic complexities such as polysemy, idioms, and syntactic variation. The Automatic Language Processing Advisory Committee (ALPAC) report (1966) confirmed these shortcomings, concluding that machine translation was largely ineffective. The report discouraged further funding for MT research focused on direct translation applications and recommended instead that resources be allocated to computational linguistics, shifting emphasis from translation to the broader study of language processing.

This shift validated Martinet’s critique, reinforcing the idea that language cannot be reduced to mechanical word-pairing. His structuralist perspective anticipated the difficulties faced by early MT, particularly the necessity of contextual awareness and structural depth in translation.

The Shift Towards AI and Contextual Understanding

With the advent of statistical models and neural networks, modern AI-driven translation systems have surpassed early word-for-word approaches. Systems like Google Translate and GPT-based models leverage deep learning and vast corpora to process entire sentences and discourses, capturing contextual meaning rather than simply replacing words. These models utilize techniques such as tokenization (breaking text into subword units) and word embeddings (mapping words into high-dimensional vector spaces) to better approximate relational meaning. Through transformer architectures, intelligent systems process language bidirectionally, considering both preceding and succeeding words to infer meaning more accurately.

Does AI Overcome the "Screen" Effect?

Despite its advances, AI translation still grapples with many of the challenges Martinet identified. While AI models approximate linguistic relationships more effectively than early MT systems, they do so probabilistically rather than structurally. This means they generate text that appears fluent but does not always preserve deep syntactic accuracy.

Although modern machine intelligence no longer treats words as rigid, isolated units, its reliance on statistical correlations rather than intrinsic linguistic understanding suggests that the "screen" he identified persists in a new form. Rather than being constrained by word boundaries, computational language models are now constrained by the patterns they extract from vast datasets. This limitation is particularly evident in morphologically rich languages or syntactically flexible languages, where automated translation systems may fail to maintain the intended meaning. Additionally, algorithmic language processing struggles with domain-specific language, idiomatic expressions, and culturally embedded references, further highlighting the enduring relevance of Martinet’s concerns.

The Future of AI and Linguistic Mediation

Recent developments in AI suggest that the field is moving toward models that better integrate linguistic structure rather than relying purely on statistical learning. One promising direction is neuro-symbolic machine intelligence, which seeks to combine deep learning with explicit grammatical rules. Unlike purely statistical models, neuro-symbolic systems attempt to incorporate linguistic constraints, potentially addressing some of the structural challenges identified by Martinet.

Research into multimodal computational linguistics, which integrates textual, auditory, and visual data, also points to a more holistic approach to language processing. Efforts such as IBM’s Neuro-Symbolic AI framework and hybrid language models aim to merge pattern recognition with symbolic representations, potentially bridging the gap between statistical fluency and structural fidelity. While these advancements offer hope for overcoming some of automated translation’s limitations, the question remains: Can AI ever fully transcend the "screen effect" and engage with language in the way Martinet envisioned?

Conclusion

Martinet’s critique of early mechanical translation remains a valuable lens for assessing modern machine-learning-driven translation systems. While computational models have advanced beyond simplistic word substitution, they still struggle to fully grasp linguistic interdependence. Although deep learning has enabled automated systems to generate more contextually appropriate translations, they operate within probabilistic constraints rather than structural understanding.
As machine translation continues to evolve, Martinet’s insights serve as a crucial reference point. Language is not merely an inventory of words but a system of interdependent elements, and achieving true linguistic mediation requires more than surface-level fluency. Whether future AI models will fully overcome the "screen effect" remains an open challenge—one that demands continued dialogue between linguistic theory and computational innovation.

Bibliography

Martinet, André, and Victor A. Velen. "The Word." Diogenes 13, no. 51 (1965): 38-54.

Saussure, Ferdinand de. "Course in General Linguistics." Translated and annotated by Roy Harris. With a new introduction by Roy Harris. Bloomsbury, 2013.

Saussure, Ferdinand de. Cours de linguistique générale. Edited by Charles Bally and Albert Sechehaye, with the collaboration of Albert Riedlinger. Arbre d’Or, Genève, 2005.

Automatic Language Processing Advisory Committee. Languages and Machines: Computers in Translation and Linguistics. Washington, DC: National Academy of Sciences, National Research Council, 1966.

Poibeau, Thierry. Machine Translation. Cambridge, MA: MIT Press, 2017.

Almahasees, Zakaryia. Analysing English-Arabic Machine Translation: Google Translate, Microsoft Translator, and Sakhr. New York: Routledge, 2022.

Steiner, George. After Babel: Aspects of Language and Translation. Oxford: Oxford University Press, 1998.