Every Shape of Every Word

Tap an unfamiliar word on a Kindle and a definition appears. It feels like nothing, a solved problem. We wanted the same small magic for the oldest books in our library: pages of Greek printed four centuries ago. Building it turned out to require something unexpectedly large: a map of every shape of every word in the language.

The problem is that Greek words refuse to hold still. A dictionary lists προσέχω, “to attend,” but a real page almost never shows that form. It shows προσέχειν, or προσέσχομεν, or one of several hundred other shapes the verb takes as it changes for tense, person, and case. Dictionaries list each word exactly once, under its lemma, the headword. Between the page and the dictionary stands a gap, and on a book from 1590, no tap in the world could cross it.

The lifetime index

Scholars have always bridged that gap the hard way. Learning to read Greek is, in large part, years of training the eye to see προσέχειν and know to look under προσέχω. And when a scholar wants to study a word rather than a page (every place an author uses it, how its meaning shifts between authors or centuries), the traditional instrument is a concordance: an index of every occurrence of every word in one author’s work, compiled by hand, sometimes over a lifetime. The concordance to one poet. The concordance to one philosopher. That was the bargain: a career of labor bought word-level access to a single author.

What dissolves the gap, for the student and the scholar alike, is a table: every shape of every word, mapped to its dictionary entry. This week we built one for our Greek collection.

Built from our own shelves

We swept all 1,040 Greek books in the library (471,544 scanned pages) and collected every distinct word form that appears in them: 3.86 million, once early modern spelling and printing quirks have their say. The well-attested forms then went through Morpheus, the venerable open-source Greek analyzer developed by the Perseus Project, which can tell you that a given shape is the present infinitive of προσέχω. Joined to the great Liddell-Scott-Jones lexicon, the result is a table of 568,918 word forms, each pointing to its dictionary entry.

Because the table was generated from our books rather than a classical syllabus, it knows the Greek our readers actually meet: the spellings of Renaissance printers, Byzantine and patristic usage, the Greek that surfaces inside Latin scientific books. And because the analyzer, its grammatical tables, and the lexicon digitization are all openly licensed, the result is fully open too.

Is it any good? We checked.

A table like this is only worth sharing if you can say how accurate it is. So we tested it against the strongest gold standard there is: the Perseus treebanks, in which classicists hand-verified the dictionary form of every word across 132,156 words of ancient text. Our table covers 88.8% of those words, and where it offers an answer it contains the scholars’ answer 96.6% of the time. Most remaining disagreements are not errors but conventions: two reference works choosing different headwords for the same word; the genuine gaps are mostly proper names.

The Latin table passes the same kind of exam. Tested against the Index Thomisticus treebank (333,281 words of Thomas Aquinas, hand-verified, and medieval Latin rather than classical, so much closer to what our books contain), it covers 91.4% of words and agrees with the scholars 95.1% of the time where it answers.

Against our own pages (real OCR of real early modern printing, not clean modern editions), a word tapped at random resolves to a dictionary entry 80% of the time for Greek and 76% for Latin, with misses dominated by names and abbreviations rather than ordinary text.

What the dataset is, exactly

One row per word form. Each row carries the form as it appears (normalized), the dictionary headword(s) it can belong to with their Liddell-Scott-Jones entry references, and how many times the form is attested across our books. 568,918 rows for Greek; the same build for Latin (where our collection is six times larger) carries 1.47 million forms mapped to Lewis & Short. Both inherit only open licenses (MPL-2.0 machinery, CC BY 4.0 lexicon data), so the tables can be republished, embedded in other tools, or used to train whatever comes next. Both are published as a citable dataset: doi.org/10.5281/zenodo.21884364.

What it makes possible

For a reader or a learner: tap any word on an original page and get its dictionary form and meaning: the beginning of reading Greek with the original in front of you, rather than about it.

For a scholar: run the table in reverse and it becomes a concordance machine, not for one author, but for a whole library at once. Every attestation of a word across two centuries of books, regardless of the shape it takes on the page: when a term enters the literature, when it fades, how it was actually used, each passage one click from the scanned page it lives on. The kind of question that once cost a career to answer for a single author becomes an afternoon’s work across thousands of books. That is what we are building next.

Built on the Perseus Project’s Morpheus analyzer (maintained by the Perseids Project, MPL-2.0), the lsj9 digitization of Liddell-Scott-Jones (CC BY 4.0), and validated against the Ancient Greek Universal Dependencies treebanks. Gratitude to the Alpheios Project, whose reading tools first showed what tap-a-word Greek could feel like.