The evidence says translate. Your teacher says don't

You have been told to think in Spanish. To let the word attach itself to the thing rather than to its English partner, because the English partner is a crutch and crutches are for people who don't want to walk.

I believed it for years. And there is a number for how far it reaches: Hall and Cook surveyed 2,785 English teachers in 111 countries for the British Council, and for each classroom function they asked about, between 20 and 35 percent said they use only English.

So here are the studies. They cut against the rule, and they fall well short of a clean win for translation. One of the two results this post rests on got much smaller when someone re-ran it properly, and that arithmetic is the most useful thing on this page.

Fair warning. Verbamor's cards put a target-language word next to English, and one of its tabs is called Translate. A finding that translation helps flatters my product. That conflict gets its own section below.

Where the rule came from

Graham Hall and Guy Cook surveyed this question for Language Teaching in 2012, across 38 pages. The belief, they write, has been that students "should be discouraged or even banned from making any use of the language(s) they already know."

Then they go looking for its origin. The 19th-century Reform Movement usually gets the blame, and Hall and Cook say it shouldn't: Henry Sweet, one of its British leaders, "explicitly advocated the use of translation in the teaching of vocabulary."

Their credible source is Maximilian Berlitz, and the story is an accident. Berlitz hired a Frenchman named Nicholas Joly to cover a French class for English speakers, not realising Joly spoke no English. Joly taught French in French with gestures and pictures because he had no other option. The students enjoyed it and seemed to progress, and Berlitz built a method on it.

It's still on sale under the same name. Berlitz's German site, checked 3 September 2026, says "in your Berlitz course you only speak and hear your target language." A staffing accident from the 1870s is still the marketing copy.

200 Berlitz schools existed by 1914. And Hall and Cook name the incentive that carried the principle outward:

A monolingual course sells in 40 countries. A bilingual one has to be rebuilt for each. That's a business reason for a rule that arrived dressed as a pedagogical one.

None of which makes the rule wrong. Bad reasons and true conclusions coexist. It just means the confidence has to come from somewhere else.

What 42 studies found about English glosses

A gloss is the little translation in the margin next to a hard word. Yanagisawa, Webb and Uchihara meta-analysed them in Studies in Second Language Acquisition: 359 effect sizes, 42 studies, 3,802 learners.

Words learned from reading, glossed vs. unglossed

0

glossed, one week later

0

unglossed, one week later

0

the first as a share of the second

Two things in there are mine. The ratio holds almost exactly across the week: glossed reading is 1.70 times unglossed immediately and 1.69 times a week later. Glossed learners keep 73.7 percent of what they had, unglossed 74.4 percent. So the gloss buys more words at the door and then forgets them at the same rate as everything else. An encoding advantage, and it doesn't decay away.

Invert it into reading. To keep 100 words a week later at 33.4 percent, you meet about 299 unknown ones. At 19.8 percent, about 505. The unglossed reader does 1.69 times the reading for the same vocabulary.

Now the part that touches the rule directly, in six words: "L1 glosses yielded greater learning than L2 glosses." Your own language beat the target language for explaining a word, across 42 studies.

The authors also found "no interaction between language (L1, L2) and proficiency (beginner, intermediate, advanced)." Advanced learners didn't outgrow it. That is the exact claim the training-wheels framing makes, and it isn't in the data.

An independent meta-analysis published this May puts glossing at d = 0.935, CI [0.780, 1.089], across 123 comparisons from 26 studies. Trust its direction more than its size: I² is 89.2 percent, and the authors' own trim-and-fill imputed 36 missing studies on the low side of the funnel, in a subsample sitting at d = 0.503.

One more thing about that L1-over-L2 result, and it cuts against me. The next section takes apart a translation study for testing people in the format they practised, and the same objection lands here. A learner given English glosses is often tested by being asked for the English meaning, the very thing the gloss handed them, while a learner given target-language definitions is tested further from their study. How much of the L1 advantage that explains, I can't tell you: the moderator tables are behind the paywall and I haven't read them. The authors do report that gains were "moderated by test formats," which is the variable that would carry such an effect. Treat this as the weaker of the post's two legs.

The famous translation study shrank

The result usually cited for translation is Laufer and Girsai, Applied Linguistics 2008. 75 Israeli high schoolers, three groups, time on task held constant: meaning-focused work, English-only vocabulary exercises, or translation both ways plus explicit comparison of the two languages. The translation group won every test, immediately and a week later.

Ali Jahangard re-ran it in 2022 with 105 Iranian university learners and fixed one thing. In the original, the translation group practised translation and was then tested by translation, while everyone else was tested on a format they hadn't practised. Jahangard built a mixed test so no group had the home advantage.

Here is what happens to the effect. I computed these Cohen's d values from the means and SDs both papers publish, so you can check them.

Translation vs. English-only form focus, as Cohen's d

2008, matched test d = 0.69 to 0.94 Translation beats form-focused work on immediate passive (0.69), immediate active (0.85) and delayed passive (0.94) recall.

2022, mixed test d = 0.09 to 0.25 Same comparison, same design, fair test: 0.09, 0.25 and 0.14. Jahangard reports all three as non-significant. Only delayed active recall, a fourth test, stayed significant.

So most of the translation advantage over ordinary English-only exercises was the test rewarding the practice. Plenty of people cite the 2008 paper as proof that translation is the best vocabulary method, and that number did not survive a fair measure.

Now look at what did survive, because it dwarfs what vanished.

Both groups that pointed at the words at all crushed the group that only read for meaning, which is a separate quarrel about whether input alone does the job. In the replication, translation beat meaning-focused instruction at d = 2.81, 3.71 and 4.14 across the three tests, and English-only form-focused work beat it at 2.42, 2.97 and 4.26. Effect sizes above 1 are uncommon in this field. These are two to four times that, and both studies agree on them.

Jahangard's summary is that the two "yielded roughly equal outcomes except for the delayed active recall of words." That exception is the one place translation stayed ahead, and it happens to be the hardest test in the set: producing the word from cold, a week later, with no prompt.

That's a smaller and more defensible claim than the one on the tin. Translation is a legitimate way to make a learner attend to a word, and the ban has no result of this size pointing its way.

The case for staying in the language

Here is the strongest version of the argument against everything above, and parts of it are right.

Glosses teach meaning and they don't teach spelling. Kang, Kweon and Choi eye-tracked 81 Korean university students reading with L1 glosses, L2 glosses or none. Verbatim from the abstract: "L1 and L2 glosses failed to significantly enhance the acquisition of visual word forms, whereas both types of glosses were significantly effective in consolidating form–meaning associations."

So a translation pair gets the meaning into your head and leaves the shape of the word outside, which is why you can recognise desafortunadamente instantly and still not spell it.

Second, English can substitute for the work instead of supporting it. Swain and Lapkin's position, as Hall and Cook summarise it, is neither prohibiting nor encouraging own-language use, precisely to avoid it "substituting for, rather than supporting, new language learning." A learner who reaches for English at the first friction never builds a tolerance for not-quite-understanding.

Third, the ban has a function unrelated to memory. A rule keeping 20 teenagers in Spanish buys target-language minutes, and it beats a subtler rule nobody follows. Duff and Polio recorded 13 university foreign-language classrooms and found teachers using the students' own language for 32.1 percent of their utterances on average, with individual classes running from 0 to 90 percent. If you're the teacher at 90, a blunt rule is an improvement.

Fourth, and this one constrains my own argument most, the profession is less absolutist than the folk rule. The same 2,785-teacher survey reports that "teachers generally agree that English should be the main language used in the classroom, most do not try to exclude completely the learners' own language, but allow its use only at certain parts of the lesson," and that "only around one third of survey respondents reported that they felt guilty" about other languages in class. So the strict ban is a minority practice held by roughly a quarter to a third of teachers, and it's something learners absorb from marketing more than a position the field defends. If you read that as me knocking over a target of my own building, that's fair, and it's why the claim I end on is about your private study rather than anyone's classroom.

Fifth, none of this is about your inner monologue. Glossing studies are about how a word gets introduced. Whether you should be mentally translating while someone talks at speed is a different question, and I have no study for it. My guess is that advice is about processing speed and is basically sound, but it's a guess.

My commercial interest in this finding

Every Verbamor card pairs a target-language word with English. Translate is a whole tab. Evidence that bilingual pairs work is evidence my design was right, which is exactly when you should get suspicious of the person showing it to you. Three things on the record.

The replication cuts against me. The clean version of this post is the 2008 result with translation winning everything, and I could have written it without mentioning Jahangard, whose paper sits in the MEXTESOL Journal and is barely cited. The honest version puts translation's edge over ordinary form-focused study at 0.09 to 0.25 of a standard deviation, usually not significant.

The gloss evidence is about reading, not cards: a translation printed beside a word in a text. A flashcard is a different object with a retrieval step in it, and that transfer is an inference I'm making rather than a result I can hand you. The card evidence I'd lean on is the testing effect, which is not about translation at all.

And the eye-tracking limit applies to my product by name. If the English side of a card does all the meaning work, the card teaches recognition and leaves word form behind. Verbamor's prompts are fill-in-the-blank in the target language with the sentence audio playing, so what you produce is the word. Whether that closes the gap Kang and colleagues found, I don't know. Nobody has run that study on this app.

What to do with your next unknown word

Look it up in English. The 42-study synthesis says a plain L1 gloss beats a target-language definition, and the advantage doesn't fade as you improve. Take the shortcut and get on with reading.

Then do the part the gloss doesn't. Write the word, type it, say it. Meaning arrives from the translation and form doesn't, so form is your job. Same case for hanging an image on it, and for spotting when a word is close enough to English to be nearly free.

Don't stop reading to translate whole sentences unless you're doing it deliberately as practice, where both papers here call it a reasonable exercise with a small edge in producing words from cold.

And if a teacher tells you to stay in the language during a lesson, do that. They're buying speaking minutes, and those are worth more than the 13.6 percentage points a gloss buys. The rule worth retiring is the one about your private study, where nobody is buying anything and the English word is the fastest route to the meaning.

Verbamor sits in that private-study half. It takes words from your own lessons and builds cards pairing them with English, native audio and a blank where the word goes, then schedules them. $19.99 a month, iOS only. Reading a lot teaches more words than any deck will. The deck is for the ones you met once and would otherwise lose.

---

Sources

Every load-bearing claim Verbamor makes is traced to its paper on the research page.