A colleague asks you a simple question in your target language. You hear it, you understand it — but the understanding arrives late, in two steps. First the words land, then something underneath quietly converts them: chotto matte becomes wait a moment, hablar becomes to speak. You answer, and the same machine runs in reverse. You build the sentence in your own language, then translate it word by word, and hope the version that comes out of your mouth is not the version that was in your head.

Learners describe this as their biggest obstacle to fluency, and they usually describe it as a bad habit — something they could switch off if they tried harder. The research says something more interesting. Mental translation is not a habit you forgot to break. It is a property of how a second language is stored in your memory in the first place, and it fades on its own schedule, not yours.

Here is what is actually happening in your head, what the evidence says you can and cannot switch off, why translation is a genuinely good study tool and a genuinely bad speaking strategy, and a six-step protocol for building direct routes to meaning.

Two Kinds of Link in Bilingual Memory

The best-known account of this problem was published in 1994 by Judith Kroll and Erika Stewart in the Journal of Memory and Language. Their revised hierarchical model proposes that your two languages are connected in two different ways.

  • Lexical links run directly between words: the Spanish word libro and the English word book are tied to each other as forms.
  • Conceptual links run between each word and meaning itself — the shared store of what a book is: paper, pages, cover, reading.

The model's key claim is about the strength of those links. For the language you grew up in, the conceptual link is strong and does most of the work. For a language you learned later, that conceptual link is weak at first, while the direct word-to-word bridge to your first language is relatively strong. So when you meet a new word, the easiest available route to meaning is through your mother tongue: the word routes to a translation, and the translation routes to meaning. That detour is mental translation, and it is not laziness. It is the shortest path your memory currently has.

An earlier experiment by Michael Potter and colleagues, published in 1984, showed the same architecture at work in beginners. They compared two ways of getting from a picture to a word: naming the object in the second language, or translating a word from the first language. Beginners performed as if they were translating — the idea was reached through the first language's word. Proficient bilinguals performed as if they were naming the concept directly. The route changes with proficiency; it does not change by willpower.

You Cannot Simply Decide to Stop

The most striking evidence that translation is automatic comes from a 2007 study by Guillaume Thierry and Yan Jing Wu in Proceedings of the National Academy of Sciences. Chinese-English bilinguals were asked to judge whether pairs of English words were related in meaning. They read only English. But hidden inside the experiment, half of the word pairs contained a repeated Chinese character when the English words were translated — a covert pattern the participants were never told about and never noticed.

Behaviorally, nothing happened: the hidden pattern did not change how fast or how accurately they answered. But their brain responses did change. The electrical signal that indexes semantic processing was modulated by a factor the readers were unaware of, and — the decisive detail — the same modulation appeared in Chinese monolinguals reading the same words in Chinese, when the character repetition was plainly visible. The pattern even survived when the words were heard instead of read.

"These findings demonstrate that native-language activation is an unconscious correlate of second-language comprehension." — Thierry & Wu (2007), PNAS 104(30), 12530–12535

Read that sentence again, because it is the whole point. When you comprehend a foreign sentence, your first language is active whether you want it to be or not. You did not fail to stop translating. You were never asked.

Why Speaking Is Harder Than Listening

The asymmetry every learner feels — understanding is manageable, speaking is a construction site — is one of the most replicated findings in bilingual research.

In a longitudinal study published in Second Language Research in 2002, Kroll, Michael, Tokowicz and Dufour had learners and fluent bilinguals name words and translate them in both directions. Learners were slower and more error-prone when producing the second language, and there was a consistent asymmetry inside translation itself: forward translation, from first language into second, was slower than backward translation, from second language back into the first. The same asymmetry shows up in the oldest demonstrations of the revised hierarchical model.

The reason is structural. Translating into your second language requires you to pass through meaning and then find the right form, which is exactly the step beginners are weakest at. Translating out of it only requires the form-to-form bridge, which is exactly the step beginners are strongest at. So the direction you find hardest is the direction that actually builds the skill.

There is a second cost, and it explains why switching between languages feels like weightlifting. In a 1999 study in the Journal of Memory and Language, Meuter and Allport had bilinguals name digits in whichever language a cue told them to use. The costs of switching were asymmetrical: switching back into the dominant language was more expensive than switching into the weaker one. The standard explanation is that the dominant language must be actively suppressed for the weaker one to be used — and paying that suppression off is what you feel. A later framework, the adaptive control hypothesis of Green and Abutalebi (2013), describes language control as a set of control processes tuned to whatever conversational context you are in. Using one language only is not the absence of control. It is a specific, effortful control setting.

Some Words Escape the Detour — and Some Never Do

Not all vocabulary is equally trapped behind translation, and knowing which is which lets you spend your effort where it pays.

Van Hell and de Groot (1998) ran a word-association experiment with Dutch-English bilinguals who gave associations twice to each word, once inside one language and once across languages. Translations of one another were produced far more often for concrete words than for abstract words, and far more often for cognates — words that look and sound alike across the two languages — than for non-cognates. Nouns produced translations more often than verbs did. In other words, concrete nouns and cognates share a conceptual representation across languages, and can be reached without the trip through the first language. Abstract words and non-cognates have to be mediated.

De Groot and Keijzer (2000), in Language Learning, found the same three factors — concreteness, cognate status, and frequency — predicting both how easily a word was learned and how easily it was forgotten.

Then there is the finding that tells you the detour can persist for years. In a 2004 study in the Modern Language Journal, Nan Jiang had Korean speakers of English judge whether pairs of English words were related in meaning. Some pairs shared a single Korean translation; others did not. The Korean speakers responded significantly faster to the pairs that shared a Korean translation, while native English speakers showed no such pattern. The conclusion was uncomfortable: the second language words were still carrying first-language semantic structure into an all-English task.

Immersion alone does not automatically fix it either. Sunderman and Kroll (2009) studied learners of Spanish and found that some learners returned from study abroad with seemingly no change in how they accessed Spanish words. The pattern pointed to a threshold: without enough working memory resources to exploit the immersion, the environment delivers nothing extra. Time in the country is an opportunity, not a mechanism.

Translation Is a Study Tool, Not a Speaking Strategy

Here is the part most articles leave out, and it matters because half of what you have been told is wrong in the other direction.

Translation as a learning activity has real experimental support. Laufer and Girsai, writing in Applied Linguistics in 2008, compared three groups of high school learners of English working on ten unfamiliar words and ten collocations, with time on task held constant. One group did meaning-focused work; one did form-focused vocabulary tasks; one did translation tasks in both directions with a teacher-led contrastive analysis of the target items and their first-language equivalents. The translation group significantly outperformed both other groups on every test, including a week later.

But the evidence is not unanimously flattering, and honest writing about it should say so. Hummel (2010), in Language Teaching Research, tested 191 French-speaking students on three conditions — translating into the second language, translating into the first language, and rote copying. All three produced significant short-term recall, and the two translation directions did not differ from each other. The rote-copying condition significantly beat both translation conditions. So the defensible claim is not "translation is the best way to learn vocabulary." It is: translation is a legitimate, effective study technique, roughly on par with other active methods and clearly better than passive reading when the material is contrastive and the task is demanding.

The distinction that resolves the contradiction is mode. Translation is an excellent way to study the relationship between two languages. It is a terrible way to use one of them. When you are analyzing a tricky collocation, comparing how two languages carve up a concept, or deliberately pushing yourself to produce a word you half-know, translation earns its place. When you are holding a conversation, every trip through your first language spends time you do not have and hands your attention to the wrong task.

The Six-Step Protocol for Building Direct Routes

1. Separate study mode from use mode

Do not try to eliminate translation from your whole life — you will fail and conclude you are untalented. Instead, set explicit windows. In study mode, translation, contrastive comparison, and analysis are allowed and encouraged. In use mode — conversation, listening, reading for pleasure — the rule is different: keep moving, even badly. Fluency practice is practice at not stopping.

2. Rebuild the flashcards that teach the detour

If your card asks for the translation of a word, you are training the lexical bridge and nothing else. Replace most cards with three types that force the conceptual route:

  • Image or situation to target word — the cue is a picture, a gesture, or a one-line scene description in the target language, never an English word.
  • Monolingual definition with a context gap — the card front is a sentence in the target language with the word missing; the back is the word plus a definition written in the target language (you already know enough words to write one).
  • Target to target — synonym, antonym, or a wrong-use example the learner has to fix.

Keep translation cards for exactly two cases: abstract nouns and false friends, where the point of the card is the mismatch in how the two languages divide a concept. Also keep them for technical and professional vocabulary, where the accurate equivalent is the entire goal.

3. Move to a monolingual dictionary

Looking a word up in your first language guarantees one more trip through the detour. Look it up in the target language whenever you can survive the definition, and keep a bilingual dictionary for the moments you cannot. The habit itself is training: you are practicing reaching meaning without leaving the language.

4. Practice paraphrasing on purpose

Circumlocution is the skill that makes translation unnecessary. Twice a week, take five words you know and describe each one in the target language without naming it, as if your listener did not know the word. Then take a sentence you can say easily and say it three different ways. This builds the network of alternative routes that speakers actually use when a word will not come — the same network that lets you stay inside the language when your memory fails.

5. Run timed output bursts with no translation pauses

Set ninety seconds. Speak or write about a familiar topic and do not stop moving. If you cannot retrieve a word, go around it — the previous step is what makes this possible. Afterward, mark two things rather than every error: the number of times you stalled to translate, and the phrases you patched a gap with. Those two numbers are your real progress metric. Translation stalling should fall before your grammar improves.

6. Shadow and read aloud at speed

Read a paragraph aloud at natural speed with a recording, keeping pace even when you lose a word. Fast reading aloud creates pressure to process meaning on arrival instead of after arrival, and the lag you are trying to remove is precisely a timing problem. Ten minutes, four times a week, is enough to feel the difference in a month.

What Not to Do

  • Do not ban your first language from study. The contrastive translation group beat the other conditions in Laufer and Girsai. Analyzing two languages side by side is one of the highest-yield things you can do.
  • Do not go fully monolingual at beginner level. Beginners use the word-to-word route because it is the only one available. Forcing concepts prematurely creates guessing, not fluency.
  • Do not confuse the two directions. If forward translation into the second language is the slower, harder direction, that is where the training value sits. Practice producing, not just recognizing.
  • Do not expect zero mental translation. The goal is not silence. It is a delay short enough that it stops interrupting you, and a fallback route — meaning first — that is available when the words fail.
  • Do not measure progress by vocabulary size alone. Jiang's speakers had the words. The question is how they were stored.

The Bottom Line

Mental translation exists because a late-learned language is stored with a strong lexical bridge to the language you already had, and a weak conceptual bridge to meaning. Thierry and Wu caught that bridge activating when participants were not looking. Kroll and her colleagues measured its asymmetry, and showed the route into the second language is the slower one. Van Hell and de Groot showed which words can leave the detour early — concrete ones, cognates, nouns. Jiang showed the detour can still be visible years later in an all-English task.

The detour is not defeated by resolve. It is displaced by use: thousands of encounters where meaning arrives with the word instead of after it, and where the fastest route your memory has happens to be the direct one. Every monolingual definition you look up, every card whose cue is a picture rather than an English noun, every ninety-second burst where you talk around a forgotten word instead of freezing — each one lays a little more of that direct route and a little less of the bridge you were born with.