Cognate checker, with the guessing labelled

You walked into Spanish knowing thousands of words already. Hospital, animal, familia, importante. English took in an enormous Latin and French vocabulary after 1066, Spanish grew out of Latin directly, and the overlap is your unfair head start.

A handful of those words are lying to you, and that handful is what this tool is really for.

How often is the cognate checker wrong?

About 3 times in 10 on a generated form, and every one of those rows is labelled. The tool has two sources of answer and never blends them.

Attested. A list of 44 English headwords whose four translations I looked up by hand before typing them in. Those rows say "attested, checked by hand" and you can lean on them.

Generated. Everything else runs through a suffix rule: swap the English ending for the Romance one and print what comes out. Nothing checks that the result is a word. Those rows say "generated by the -tion rule, not checked", and here is the part most tools like this leave out.

I ran 20 rule-triggering English words, 80 generated forms across the four languages, and checked every one by hand:

So a generated form is wrong about 3 times in 10, and the misses are not random. French swallows the accents the rule can't predict (célébration comes out as celebration). Italian and Portuguese assimilate consonants English keeps (ottimismo, not optimismo). And an English stem that drifted takes the whole guess with it: tourist gives tourista, and the word is turista.

A 3-in-10 miss rate is what a letter-level rule can do, and I would rather print the number than describe it as "approximate". Use the generated forms the way the post on cognates suggests using them: guess boldly in conversation and let your tutor catch the misses. Do not put one on a flashcard without checking it first.

The false friends, and where the list came from

Now the lie. Some pairs look related and drifted, or never were related. The classic is embarazada: it looks like embarrassed and it means pregnant, and generations of learners have announced a pregnancy to a dinner table while trying to apologise for one.

False friends are nasty for a memory-specific reason. You're fighting an existing, deeply automatic association, and under pressure the old habit wins. A brand-new word has no competition. Éxito has 30 years of "exit" pushing against it.

The false-friend list is mine. 48 entries, compiled from the cognates post and the standard teaching examples, typed by hand. I'm releasing it under CC0: take it, use it, no attribution needed. It comes from no third-party word list, because a licence I can't verify is a worse problem than a shorter list.

Two honest weaknesses in the false-friend list. It leans Spanish, 20 of the 48, because that's where the post's evidence is and where I can check my own work fastest. And 48 is small. The good news buried in that: common false friends genuinely do number in the dozens, against tens of thousands of honest cognates, so the tax rate on the inheritance is low. You just have to know which relatives to watch.

The suffix rules, written out

Fifteen suffix rules, applied to the stem after the English ending is stripped. The order in the file is the order they're tried, and the first match wins. Twelve are below; the three left out are -ary, -ate and -ct, which are the rough ones.

Two rules I deleted rather than shipped, which is worth saying because the deletions are the interesting part.

No general -ly rule. English adverbs build on an adjective the rule can't recover. "Famously" isn't famous plus mente, it's famosa plus mente, so the rule returned famousmente in three languages out of four. A rule that's wrong every time is worse than no rule.

No -ment rule. Same problem in the stem. "Movement" is movimiento, mouvement, movimento, and the rule produced movemento and got none of them.

What survived on top of the endings is two spelling swaps, applied to Spanish, Italian and Portuguese and not to French: ph becomes f, th becomes t, and a y in a stem those two touched becomes i. That's how physical reaches fisico instead of physico. The real Spanish word is físico, with the accent the rule cannot predict, and that near-miss is the single commonest way a generated form is wrong. French keeps its ph, which is why it gets none of this.

What this does not tell you

A generated form may not be a word. Repeated here because it's the whole risk. Three in ten are wrong as printed, and the tool has no dictionary to check itself against. Every row tells you which kind it is; believe the label.

Gender, articles and plurals are all missing. The tool returns a bare form. Whether it's el or la, whether the adjective agrees, and what happens in the plural are all things you still have to learn, and a card built from a bare noun will teach you the wrong habit.

The drift zone is not covered. Between honest cognates and outright liars there are pairs that share most of a meaning with edges that moved. Asistir usually means to attend. Discutir leans toward arguing. Those are trickier than false friends because they're right often enough to keep your guard down, and the tool only flags the clean cases.

Cognates pay out on the input side. Recognising información while reading costs you nothing. Producing it correctly, right stress, right vowels, no English -tion sneaking in, still takes reps. The head start makes your reading vocabulary enormous on day one while your speaking vocabulary stays honest.

And the pronunciation trap. Hospital the Spanish word shares letters with hospital the English word and almost none of the sounds. A cognate is the word most likely to be read with an English mouth, which is why it especially needs native audio attached from the first encounter.

Sources

The research behind why cognates are worth chasing is in cognates and false friends: de Groot & Keijzer (2000) for cognates being learned faster, recalled better and forgotten more slowly, and Costa, Caramazza & Sebastián-Gallés (2000) for the cognate facilitation effect in picture naming. The attested list and the false-friend list on this page were both compiled by hand for it, and the false-friend list is CC0.

If you'd rather check a whole page of text than one word, the coverage checker tells you what share of a text sits inside the first few thousand words of a language.

Questions people actually ask