These dictionaries are built from real sentences — hundreds of thousands of them per language. Every word is extracted from natural text, so the definitions, examples, and frequency rankings reflect how words are actually used.

Sentences are analyzed using Lexide, a set of custom machine learning models trained for language learning. They handle things like tokenization, lemmatization, and part-of-speech tagging across all supported languages.

Patterns across sentences are also analyzed using a unigram language model to automatically discover phrases, also known as multiword expressions. This matters because a lot of meaning lives in phrases — "ice cream" means something completely different from "ice" and "cream" on their own, so a good dictionary should give "ice cream" its own separate entry.