Among the many challenges facing translators is the extraordinary idiomaticity of all languages. Generally speaking, “idiomaticity exists when the inherent meaning of the words expressed does not match the meaning of the expression as understood by the speaker(s) and the hearer(s)” (Wikipedia). This can take many different forms and pose considerable problems for both human and machine translators.
Idiomaticity permeates language at all levels
At one end of the scale, there is the simple collocation of two terms such as verb and object. In English, for instance, you “break a record”, but in French you “beat a record” (battre un record). In English you “set the table”, while in French you “put forks and spoons” (mettre le couvert).
At the other end, there are thousands of expressions which are strictly non-compositional, that is, their meaning cannot be reduced to the sum of their parts, nor can the individual words that compose them be replaced by other words. Take for instance the difference in English between a verb followed by a prepositional phrase and a phrasal verb followed by a noun phrase: to [run [up a hill]versus to [[run up] a bill].
In the first case, the meaning of the expression is transparent if you understand each of the words that compose it and each constituent can be replaced by a different one: run up the street, run down the hill, walk up the hill, etc.
In the second case, the expression is fixed and non-transparent, and you need to have memorized its idiomatic meaning: “To accumulate a large bill or debt that one is obliged to pay”. You don’t have to know where this expression comes from to translate it into French, or any other language, but you do need to be extremely sensitive to context. Look for instance at all the different translations in French given for “run up” on the site Linguee [1].
Metaphors
Metaphors are another case of idiomaticity that challenges both human and machine translators. Take for example the following excerpt from Mary Shelley’s Frankenstein and its French translation.
It was already one in the morning; the rain pattered dismally against the panes, and my candle was nearly burnt out.
Il était déjà une heure du matin ; une pluie funèbre martelait les vitres et ma bougie était presque consumée.
Literally, patter means “to speak in a rapid or mechanical manner”. It derives from the medieval idea of monks mechanically reciting their “pater noster”. The French translation evokes another metaphor to express the same idea, marteler, meaning “faire des petits bruits secs comme si on tapait avec un marteau”.
There are hundreds of other examples of idiomaticity in English and other languages, such as compounds, proverbs, collocations, formulaic expressions, and so on[2]. The few examples given here give some idea of the extent to which culture permeates language as well as the degree to which the skilled translator needs to be sensitive to differences between literal and metaphorical meaning at all levels and to the complex relationship between an expression and its context.
How should the human translator proceed?
To quote a recent talk on translating idioms: [3]
“To navigate these linguistic and cultural barriers effectively, translators must act not only as language experts but also as cultural mediators. They must decide whether to:
- retain the original idiom with footnotes or explanations,
- replace it with a culturally and functionally equivalent idiom in the target language,
- paraphrase the expression to preserve the meaning while sacrificing form,
- or, in some cases, omit the idiom if it adds confusion without contributing meaningfully to the text.
The author gives several examples of these techniques. For instance, using a functionally equivalent expression. “The English idiom when pigs fly is used to express impossibility. Its Russian counterpart is когда рак на горе свистнет (‘when the crayfish whistles on the mountain’), and in Uzbek, one might use qushlar suvda suzsa (‘if birds swim in water’). Though the images differ, the intended meaning is preserved. Likewise, the French idiom avoir le cafard (literally, ‘to have the cockroach’) means to feel down or depressed. Translating this idiom literally would confuse English speakers. The closest English equivalent in both meaning and tone might be to feel blue or to be in a funk, depending on the context.”
When there is no functional equivalent, a paraphrase may be used. Again, quoting Tulkinov, “the phrase to spill the beans, meaning to reveal a secret, might be translated simply as to reveal a secret in languages where no colorful idiomatic equivalent exists. Though this strategy loses the imagery and informality of the original, it maintains clarity and avoids confusion especially in legal, academic, or journalistic contexts where precision is prioritized.”
Idiomaticity is therefore a serious challenge to human translators. What about Machine Translators like Deep L Translator, Google Translator or Bing Microsoft Translator?
The greatest challenge to MT has always been the natural ambiguity and flexibility of human language and the possibility of different meanings depending on the context. Translating ambiguity in a specific context requires deep background knowledge of the domain and familiarity with the speech habits of a particular community of users. This is the point where MT begins to show its limitations.
A brief history of MT
The first use of “translating machines” goes back to the mid-1930’s[4], when the term was introduced by Artsrouni and Troyanskii, who built what was basically a mechanical multilingual dictionary with rules for coding grammatical roles. In 1947, the work done on code breaking during World War 2 led to new developments in information theory and speculations about the universal principles of natural languages. The first scientific conference on MT was convened at MIT in 1951 by Yehoshua Bar-Hillel. Three years later, Garvin and Sheriden, two American researchers, presented an MT system which had a vocabulary of approximately 250 words and a restricted grammar.
During the next decade, after an initial flurry of enthusiasm, there was considerable skepticism about the usefulness of further investment in research on MT. According to the report published in 1966 by the National Academy of Sciences[5]:
“[We] do not have useful machine translation [and] there is no immediate or predictable prospect of useful machine translation”
Nonetheless, many research groups sprang up around the world in the 1980s and 1990s and dozens of new programs appeared, such as the Météo prototype developed to translate weather forecasts from English into French, SYSTRAN, developed to translate Russian scientific and technical materials into English or EUROTRA (EURopean TRAnslation) established and funded by the European Commission from 1978 until 1992.
Since 2000, MT has reached a new dimension with the advent of Google Translate, Deep L and other commercial programs, not to mention dozens of domain specific systems, like Cap Volmac Lingware Services in the Netherlands, Cap Gemini Innovation in France, Digital Sonata in Australia, or Meaningful Machines in New York. Most of these systems have now switched to neural machine translation.
What is neural machine translation?
MT has gone through many phases since the 1980s: Rule-based MT, Phrase-based MT, Example-based MT, Statistical MT, to mention only a few, and, since 2016, Neural MT. Rule-based machine translation relies on explicit linguistic rules and dictionaries to perform translations. Statistical machine translation involves the use of statistical models and algorithms. Neural machine translation uses artificial neural networks to model the translation process. It has replaced earlier systems due to its superior performance and its ability to capture contextual nuances.[6]
How well do Machine Translators translate idioms?
In 2024, M. Ben-Yahia tested the ability of 3 commercial translation systems – Google Translator, Bing Microsoft Translator and Deep L Translator – to translate 10 American idiomatic expressions into Arabic, Spanish and French.[7] He chose the ten expressions most often used in the American media – in the long run, on the sidelines, all out, behind the scenes, big deal, across the board, around the corner, follow suit, over the top, and in full swing – and submitted them to Google Translate, Bing Microsoft Translator, and DeepL Translator. Here are the results of his experiment:
- “Concerning the translation of the idioms from English to Arabic, the results indicate that Google Translate results are 30% accurate. The translations are mostly literal ones; this is proven to be as not useful with regards to idiom translation. … Bing Microsoft Translator, on the other hand, shows an accurate translation of idioms with a surprising percentage of 100%.
- “Shifting focus towards the Spanish translations, the results indicate that Google Translate’s accuracy … is 90%. However, it should be noted that some of these translations could be deemed not as accurate in certain contexts. … As for Bing Microsoft Translator, the results show an almost similar accuracy of 80%, with the strategy being translation by paraphrasing in most cases. … For DeepL translator, the level of accuracy when translating idioms from English to Spanish is 80% as well
- “With regards to the translation to French, Google Translate is 60% accurate. … The translation accuracy of Bing Microsoft Translator, on the other hand, is found to be 70%. It seems to be capable of identifying the actual meaning of an idiom and of avoiding literal translation that would render the translation inaccurate. As for DeepL Translator, its level of accuracy for translation into French is 40%. In most cases, DeepL Translator opts for literal translation. … Even worse, it sometimes changes the meaning or omits part of it.” (p.53)
His conclusion: “Based on the results, it is concluded that machine translation is currently capable of translating idioms with an average accuracy rate of 68.75%, with Bing Microsoft Translator being the most accurate online machine translation service, compared to Google Translate and DeepL Translator.” (p.54)
What does this mean for a professional translation service?
While neural MT has made significant progress and continues to evolve, a number of issues still remain[8]:
- Idioms and Sarcasm: MT often fails to capture local expressions like “kick the bucket” or “spill the beans.”
- Domain-Specific Jargon: Legal, medical, and technical terms can be mistranslated, especially without domain adaptation.
- Mixed Languages (Code-Switching): Sentences that mix two languages (e.g., “Spanglish” or “Chinglish”) confuse models.
- Cultural Context: Humor, tone, and intent are difficult for machines to interpret.
There are different solutions to these problems. One is what is called “adaptive MT”. In this approach, human translators “teach” the machine the appropriate terminology and style specific to a given industry and brand, in that way finely honing a multi-tool generic translator to a specific translation context. This is done by “steering” the system, that is, giving it explicit instructions that tell it to replace its output if it doesn’t match what you want. Another way to overwrite and steer an MT model is through user feedback or by feeding it relevant data.[9]
Another solution is to combine human and MT competence in a “hybrid” approach. In terms of cost and time efficiency, this may be the best solution for certain translators. The following is a comparison of the three models, that is, pure human, pure machine and hybrid.[10]
Pure Human Translation
This model delivers the highest quality and cultural accuracy, making it ideal for legal documents, marketing content, and sensitive communications. Human translators ensure deep contextual understanding and cultural nuance, reflected in a high COMET quality score of 0.82.
- Cost per word: $0.12
- Turnaround time: 48–72 hours
- Cultural sensitivity: High
Pure Machine Translation (Neural Machine Translation – NMT)
If speed and cost are your top priorities, machine translation offers rapid results at a fraction of the price. However, it’s best suited for internal documents or low-stakes content due to its lower quality score of 0.67 and limited cultural awareness.
- Cost per word: $0.05
- Turnaround time: Within minutes
- Cultural sensitivity: Low
Hybrid Model (Machine Translation + Post-Editing)
This model strikes a balance between cost, speed, and quality. Machine-generated translations are refined by professional linguists, improving both accuracy and cultural nuance. With a COMET score of 0.76, it provides a solid option for most business needs where quick delivery is important, but quality cannot be compromised.
- Cost per word: $0.08
- Turnaround time: 24–48 hours
- Cultural sensitivity: Medium to High
Conclusion
Machine versus human translation is no longer a pertinent question in 2025. It all depends on the translator’s specific needs, competence, budget and time limitations. The field is evolving so rapidly that any statement made today will be out of date tomorrow. I hope this article at least gives you some perspective on the problem.
[1] https://www.linguee.fr/anglais-francais/traduction/run+up+a+bill.html
[2] To learn more about this, go to sections 5.2 and 9.3 in A Linguistic Handbook of French for Translators and Language Students.
[3] Tulkinov Abdulaziz, “Challenges of Translating Idioms and Fixed Expressions”, Conference on the Role and Importance of Science in the Modern World, Volume 02, Issue 06, 2025: 113-119.
[4] See Idiom Treatment Experiments in Machine Translation, Dimitra Anastasiou, Doctoral dissertation, Saarbrücken, 2010, for a complete discussion.
[5] Pierce, J. R.; Carroll, J. B.; Hamp, E.P.; Hays, D.G., Hockett, C.F., Dettinger, A.G., Perlis, A., (1966), Language and Machines. Computers in Translation and Linguistics, Report by the Automatic Language Processing Advisory Committee (ALPAC), Division of Behavioral Sciences, National Academy of Sciences, National Research Council, Washington, D.C.: National Academy of Sciences, National Research Council, Publication 1416, page 66.
[6] See the Wikipedia article on machine translation for more information on these different processes.
[7] Moad Ben-Yahia, “Examining the Efficiency of Machine Translation in Translating English Idioms used in American Media”, Journal of Translation and Language Studies, 2024 Volume 5, Issue 2: 43–55, https://jtls.sabapub.com, DOI: https://doi.org/10.48185/jtls.v5i2.1070
[8] According to the AI Phone Team blog published in July, 2025: “What machine translation can and cannot do in 2025.
[9] « Tailoring AI-powered machine translation for your business”, Ollie Stott’s Blog post on “AI powered MT for your business”. Published in April 2024.
[10] See the Blog post “The AI revolution in translation: human expertise vs machine translation in 2025”. Published on June 13, 2025.
La Boîte à Langues 
Laisser un commentaire