A word game needs more than a list of legal strings. It needs to say what each one means, and meaning is the hardest thing to store.
WordChess plays on a 25×25 board with a 148,941-word English dictionary6, entries averaging 8.6 letters, some running to 25. That is the easy part. A list of legal words is just a set of strings the game will accept. The interesting part is telling a player what the string on the board means, and doing it across twenty-two languages at once.
Lexicographers draw a sharp line between a lemma and its inflected forms. The lemma is the headword you look up, run, house, be. Around it orbits a paradigm of surface forms: runs, ran, running. Each carries the same core sense, dressed for a different grammatical role.3 A dictionary spends its prose on the lemma and leaves the shadows to point home.
How many shadows a word casts depends on the language. English is analytic: it leans on word order and little helper words, so a verb rarely has more than a handful of forms. Finnish is agglutinative, it builds meaning by gluing suffixes onto a stem. A single Finnish noun inflects for some fifteen cases, in singular and plural, before possessive endings and clitics stack on top; the practical count of distinct forms runs into the thousands.4 Hungarian packs a whole English phrase, in my house, into one word, házamban.5
The definitions come from Wiktionary, the free dictionary that anyone can edit. It is enormous and multilingual: the project spans well over a hundred active language editions and tens of millions of entries, written collaboratively by volunteers rather than a paid editorial board.1 For a game that runs in 22 languages, no other free lexicon comes close on coverage.
But coverage on paper is not coverage you can read out. Wiktionary stores inflected forms in a way that saves human editors from writing the same gloss ten thousand times. Instead of spelling out a definition, a non-lemma entry carries a form-of template, a small pointer that says, in effect, "this is a particular grammatical form of that lemma."2 The entry for smoulders is not prose; it is a template that renders as "third-person singular simple present of smoulder."1
These pointers are not a rounding error. On the English Wiktionary, of roughly 1.27 million definitions, some 478,000, well over a third, are "form of" definitions rather than written-out senses.1 The data is there. It is just folded up.
So the challenge that mattered was never "the word is missing." It was that the word's meaning sat inside a template a naive parser would throw away. WordChess's definitions were built by reading the raw Wiktionary dumps and expanding the inflection templates language by language, teaching the parser what each pointer means and rewriting it into a plain gloss.6
Every language hides its forms behind different machinery. Spanish tucks verb forms behind {{forma verbo}}, resolved into "verbal form of X." German uses {{Grundformverweis Dekl/Konj}} for declined and conjugated forms. Finnish leans on {{taivutusmuoto}}. Russian often stores no form template at all, the inflected word is a bare redirect (собаки → собака) that must be followed to recover a "form of" gloss.6 Handle each convention, and the coverage jumps.
| Language | Before | After | Gain |
|---|---|---|---|
| German | 45% | 92% | +47 |
| Dutch | 58% | 90% | +32 |
| Finnish | 54% | 74% | +20 |
| Russian | 20% | 70% | +50 |
| Greek | 15% | 57% | +42 |
Measured from this project's design and build notes. Percentages are share of dictionary entries carrying a usable definition or resolved "form of" gloss.
The data was never missing. It was folded into a pointer, and giving a machine the word meant learning to unfold it.
It is worth being honest about what the game now holds. When WordChess shows that собаки is a form of собака, "dog," it has not understood anything. It has mapped one string to another string, a gloss, by following the same pointers a human editor left behind. This is lemmatization, the workhorse of natural-language processing: reducing a surface form to its base so a machine has fewer distinct things to track.7
That is a real and useful kind of knowing. It lets the game answer "what is this word?" in Athens or Amsterdam without a lexicographer on staff. But it is knowing-as-lookup, not knowing-as-meaning. The gloss is a promise that a person, reading it, will recognize the word. The machine holds the promise; the reader supplies the sense.
Some languages hit a ceiling that no parser can lift, because the data simply is not there to unfold. Hungarian entries frequently carry only an etymology and a table of translations, no definition text at all, so even a perfect reader comes away empty. The hardest cases cluster where the linguistics is hardest: agglutinative languages such as Finnish and Hungarian, which spawn enormous paradigms, and the Cyrillic and Greek scripts, where community coverage is thinner to begin with. Greek climbed from 15% to 57%; that it stops well short of German's 92% is a fact about the source, not the code. 6