How the dictionaries are made
tapdict is a small, independent project from Germany. It started with a simple wish: to read real books in a new language on an e-reader, with a dictionary that knows the words that are actually on the page.
The content is written by AI
Writing a full dictionary by hand keeps a team of lexicographers busy for years. We use AI language models instead, mostly small models that run on our own computer, and a set of automatic checks around them. That is how we can offer good dictionaries for free, also for language pairs that publishers don’t cover.
It also means that the dictionaries contain mistakes. Most entries are right, but now and then you will find a wrong translation, a missing meaning or odd grammar. We say this on every page, and we publish our measurements with every version.
From texts to entries
- Words from real texts. We collect the word forms that occur in a large amount of text in the language, with their frequency and the sentences they appear in. That is why the dictionaries know forms like “fjorden” or “went”.
- Analysis. For every word form, a language model works out the base form, the part of speech and the grammar, based on the sentences in which the word appears.
- Translations. For every base form, a model writes short translations, the most common meaning first. Dictionaries into other languages are usually translated from the English entries.
- Word forms. Missing forms such as plurals or past tenses are added, so you can also look up forms that were rare in the texts.
How we check the results
- Agreement. Small models make random mistakes. Important decisions are asked several times in different ways and only accepted if all answers agree.
- Second opinion. The most frequent entries, the ones you will look up most, are reviewed by a larger AI model, which corrects them where needed.
- Spot checks. Before a version is published we check a random sample of entries by hand and measure the error rate.
- Coverage. We measure how many words of a book you can look up, using a book that was not used to build the dictionary, and how much of the basic vocabulary is included.
You find these numbers on the page of each dictionary under “How good is it?”.
What works well and what doesn’t
Common words, inflected forms and the main meanings are the strength of these dictionaries. Rare words, idioms, words with many meanings and very new or regional words are harder. If a translation seems strange in context, trust the context.
Written by AI, so read with a little care
The entries were created with AI language models and checked automatically, partly by a second, larger model. Most of them are right. Some are wrong or incomplete.
Free, without tracking
The dictionaries are free for personal use. We set no cookies, show no ads and use no analytics. We only count how often each file is downloaded, as a plain number without any information about you. The details are in the privacy policy.
Contact
Found a mistake, missing a language or just want to say hello? Write to [email protected]. We read every message.