About Vocab.rs

Why this dictionary?

To read in Serbian, and to learn the language through reading.

According to the linguist Stephen Krashen, we acquire a language when we understand what we read and hear, especially when it is slightly above our level. Speaking can't be taught directly: fluency "emerges" on its own once enough comprehensible input has built up.^[1] That idea is where Vocab.rs comes from: the more you read and understand, the faster the language sinks in.

Reading is especially convenient for this: unlike speech, a text can be paused, reread, and an unknown word looked up. And much of what we see in a foreign language every day, from news to social media to messages, is text.

But a text does not become comprehensible right away. To understand it without help, you need to know about 98% of its words, that is, meet no more than one unknown word in 50. Minimally adequate comprehension starts at about 95%.^[2] For English, 98% of ordinary written text is covered by 8,000–9,000 word families.^[3] So a learner keeps running into unknown words and needs a way to understand them quickly and keep reading.

That is what Vocab.rs is built for: a word opens in a couple of seconds, with meanings, examples and forms, and you can get back to reading. Looking words up also helps you remember them: words from a text are retained noticeably better when the reader pays attention to them, by seeing a gloss in the margin or looking them up in a dictionary.^[4]

This is why the dictionary starts with the most frequent words: they are the ones you meet most often in texts. Their order comes from Serbian subtitles: word frequencies in subtitles reflect everyday language well, as a study of Greek showed.^[5]

How is the dictionary made?

Every entry goes through the same steps.

1. Choosing words. Words are taken from a frequency list of Serbian subtitles (OpenSubtitles), from the most frequent to rarer ones. So the dictionary is first of all about living, everyday vocabulary.
2. Facts from Wiktionary. For each word, the English Wiktionary provides the forms, stress, pronunciation, gender or verb aspect, and short definitions and labels. We cross-check these facts against other reference works and look into any discrepancies separately.
3. Writing the entry. AI receives these facts and uses them to write the meanings in English and Russian, examples with translations, set expressions and notes.
4. Automatic checks. A program rejects synonyms that are not in Wiktionary, set expressions that don't contain the word itself, words that mix Latin and Cyrillic letters, and answers in the wrong format. An entry that fails the checks is not published.
5. Review by another model. A different AI model, not the one that wrote the entry, reads it and looks for wrong meanings, inaccurate translations and unnatural examples.
6. Fixing. A model fixes the entry based on the review. The fixed entry goes through the automatic checks again.

The bottom of every entry shows the date it was last updated.

How can we improve the dictionary together?

Spotted an inaccurate translation, a clumsy example, a wrong stress mark or a recording that sounds off? Suggest an edit.

Every entry has a " Suggest an edit" button. Choose what to fix: a translation or meaning, an example, stress, the recording, word forms, a typo or a missing meaning, and describe it in your own words. No name or email needed. If you'd like a reply, write to contact@vocab.rs.

What happens next. Once a week we go through all suggestions. A model rechecks meanings, translations and examples against the entry and fixes them if the suggestion holds up. Stress, word forms and recordings we check ourselves. A suggestion doesn't change the entry by itself: the fix appears only after it has been checked, with the next site update, and the "Updated" date at the bottom of the entry changes.

Where does each part come from?

Word forms (declension and conjugation), stress, pronunciation in the International Phonetic Alphabet, noun gender, verb aspect and aspect pairs almost always come from Wiktionary and other sources. If a word's forms were written by AI, the bottom of the entry says so. Wiktionary is written and edited by people. Its data is open: you may freely copy and adapt it as long as you credit the source (CC BY-SA 4.0 license).

AI doesn't write the Cyrillic either: a program converts Latin to Cyrillic by rule.

From these facts, AI writes everything else: meanings, translations, examples, set expressions and notes. It also suggests synonyms and antonyms, but only those found in Wiktionary make it into the entry. The examples are not quotes from books or films but typical sentences that show how the word is used.

Entries are written by Claude Opus (Anthropic), reviewed by Claude Fable (Anthropic) and GPT-6 Astra (OpenAI), and fixed by Claude Fable. The model that writes them was chosen through tests: six models wrote entries for the same 45 words, and two other models rated the results blind. Without the facts from Wiktionary, even the best model made noticeably more mistakes: it missed meanings and made up synonyms.

How are entries checked?

Checking costs more than writing an entry: we use stronger models for it. So entries are checked by frequency, from the most common words to rarer ones.

In September 2026 we measured quality: GPT-6 Astra (OpenAI) reviewed 100 random entries.

• 54 entries had no issues.
• 44 entries had issues. Most often a missing meaning, unnatural wording in an example's translation or an imprecise synonym. For example, the entry nakaza lacked the figurative sense "a morally repulsive person".
• 2 entries had a serious error: a wrong translation that could teach you something incorrect. Both are fixed.

The fixed entries were checked again. None of the 43 had a serious error, and 33 passed with no issues.

A model's review is no substitute for a human proofreader: different models can share blind spots. No native speaker has proofread the dictionary yet. If you are a native Serbian speaker and willing to look at a few entries, write to us at contact@vocab.rs.

Can you trust the dictionary?

For study, reading and everyday conversation, yes. But it is not a dictionary proofread by editors, and you shouldn't rely on it the way you would on an academic dictionary.

• Most reliable: word forms, stress, pronunciation, gender and aspect. They almost always come from Wiktionary and other sources, not from AI.
• Usually right: the main meanings of common words and their translations.
• Inaccuracies happen: a rare or figurative meaning is missing, an example's translation sounds unnatural, a synonym doesn't fit every context, a note is worded too categorically.
• Serious errors are rare: a wrong translation of a meaning or an example.

If a lot depends on the translation (documents, official letters, work), check with a native speaker or an academic dictionary, such as the Matica srpska dictionary (Rečnik srpskoga jezika). And if you spot a mistake, suggest an edit with the " Suggest an edit" button: that's the fastest way to get it fixed.

Which words are in the dictionary?

The most frequent words of Serbian, plus set phrases such as dobro jutro and zato što. The dictionary keeps growing. Word forms (je, su, ide) don't have entries of their own: search finds them and takes you to the entry for the base form (biti, ići). Vulgar words and words that Wiktionary labels as Croatian or Bosnian only are not included yet.

Which variety of Serbian?

Ekavian, as spoken in Serbia: mleko, dete, reka. The main script is Latin. The "Latinica | Ћирилица" switch in an entry turns on Cyrillic. Ijekavian forms (mlijeko, dijete) don't have separate entries yet.

What's next for the dictionary?

Vocab.rs grows together with its readers: every suggested edit makes an entry more accurate.

Here's what we plan:

• Grow the dictionary. New words will be added based on how the language is actually used.
• Add real examples. For now, examples are written by AI. Over time, examples from real texts will be added.
• Keep up with new words. Every year Serbian gains loanwords, tech vocabulary and new slang. We want to spot such words through real usage and publish a yearly selection of the most notable ones, with explanations and examples.