Ancient Greek lemmatizer

Ancient Greek lemmatization and reading

greek.lemmatisation.com is a free digital philology tool. It lemmatizes Greek forms in context, gives their grammatical analysis and a French gloss, and lists less frequent vocabulary in textual order.

Parousia analyses each sentence locally as a whole: for every word it gives the LSJ headword, the part of speech, the morphology and its function in the sentence’s dependency tree, following Universal Dependencies. Pasted text remains in the browser. Unicode, elision and editorial noise are normalized before analysis. On first use the browser downloads Parousia and its runtime, about 16 MB compressed, then keeps them in its cache.

The LSJ, Pape, Bailly and Abridged Bailly dictionaries and the inflection lookup complement the analysis. Frequency ranks were counted before Parousia, by Sôtêr mini, an earlier model of the site, over an earlier state of The Greek Library corpus: 71,036,436 words across 5,974 editions. As with any automated analysis, suggested lemmas may require philological verification.

Where Parousia comes from

Parousia was trained on 87.6 million words of Greek from The Greek Library, from the Presocratics to the Byzantines, annotated under a single standard: LSJ headwords, AGDT positional morphology, UD dependencies. Those annotations come from a teacher model made of two larger parsers built on GreBerta and PhilBerta, themselves trained on 19.2 million words annotated by a large language model guided by the standard. Parousia is the pocket version: 3.8 million active parameters and 10.8 million in its tables (lexical memory and lemma rules), that is 17 MB once quantized to int8, small enough to run in a browser. It succeeds Sôtêr, which only gave the lemma.

How Parousia differs from a morphological analyser

Eulexis, by Philippe Verkerk (Biblissima+), is a morphological analyser: it looks each form up in a table of attested forms and returns every lemma that form can belong to. The same form always yields the same list, whatever the sentence. That is its strength — nothing is hidden, and the reader decides.

Parousia is a trained model. It reads the whole sentence, each word with its letters and its neighbours, and returns one lemma and one analysis. It therefore settles the ambiguity from context — αὐτοῦ as genitive of αὐτός or as the adverb “there”, ἦ as verb or as particle — where a table can only list the possibilities. In return, it can be wrong, and it does not enumerate: what it does not propose is not shown.

The two approaches complement rather than oppose each other, and this site uses both: the Eulexis table is consulted here, as a fallback, whenever Parousia returns a lemma that neither the dictionaries nor the frequency list knows.

Method and limitations

Parousia’s output is a prediction, to be checked as one checks a commentary. What you need to know to use it:

• One lemma and one analysis per form, with no alternatives; the Eulexis table is consulted only as a fallback, for a lemma that neither the dictionaries nor the frequency list knows.
• Words written entirely in capitals are analysed in lower case, without accents: the analysis is less reliable, and the headword is re-accented from the frequency list.
• Headwords are those of the LSJ; the French gloss comes from the Abridged Bailly, supplemented with translated LSJ short definitions. The two may diverge.
• Frequency ranks were computed by Sôtêr mini, another model: a lemma returned by Parousia may be missing from them.
• Trees are shown for the first 300 sentences of a text; lemmas, morphology and vocabulary cover the whole text.
• The analysis can be exported as CoNLL-U (UD and AGDT tags) and as CSV, to be corrected or reused elsewhere.

The figures below are percentages of Greek words, punctuation excluded, whose annotation matches the reference. The first table measures what the reader sees: the lemma returned by the site’s full pipeline, from text normalization to the displayed headword, compared with that of Sôtêr v2, the previous model. For the four AGDT texts, the sample is 300 forms per work, evenly spaced: one point of difference there amounts to three words.

The second table covers the whole analysis. Each treebank has its own conventions (choice of headwords, categories, relation labels), and Parousia’s standard matches none of them exactly: its output was therefore converted into each treebank’s conventions by correspondence tables learned on that treebank’s training split. These figures measure agreement with the reference after that conversion, not the site’s raw output: on Perseus, 93.1% of lemmas are correct after conversion, 92.5% as displayed. Part of speech: UD universal tag (UPOS); UAS: correct syntactic head; LAS: correct head and function.

Unedited Byzantine Greek remains the weak point: on the DBBE, the part of speech is correct for only 86.1% of words, and Sôtêr v2 returns the lemma slightly better than Parousia there. A disagreement with the reference is not always an error: two treebanks do not always use the same headword for the same word.

How to cite

Luc Pommeret, Jules Deret and Thibault Wagret, greek.lemmatisation.com: Ancient Greek lemmatization and parsing with Parousia, LISN (CNRS, Université Paris-Saclay) and HiSoMA (UMR 5189), version of 15 October 2026, https://greek.lemmatisation.com/en/ (accessed …).

@software{pommeret_greek_lemmatisation,
  author       = {Pommeret, Luc and Deret, Jules and Wagret, Thibault},
  title        = {greek.lemmatisation.com: Ancient Greek lemmatization
                  and parsing with Parousia},
  organization = {LISN (CNRS, Université Paris-Saclay) and HiSoMA (UMR 5189)},
  year         = {2026},
  version      = {2026-10-15},
  url          = {https://greek.lemmatisation.com/en/}
}

Sources and licences

Analysis

Parousia, by Luc Pommeret, run in the browser by ONNX Runtime Web. Annotation standards: LSJ headwords, morphology of the Ancient Greek Dependency Treebank (AGDT, Perseus), Universal Dependencies relations.

Dictionaries and inflection

Eulexis-web (Philippe Verkerk, Biblissima+), queried remotely: LSJ, Pape, Bailly 2020 and Abridged Bailly.

French glosses

Abridged Bailly, from the Eulexis data; supplemented with LSJ short definitions (Perseus Digital Library, CC BY-SA) translated into French. The fallback table of forms also comes from Eulexis.

Frequency list

The Greek Library corpus, lemmatized by Sôtêr mini (Luc Pommeret, Jules Deret, Thibault Wagret).

Typefaces

Lemmata, derived from Gentium Book Plus (SIL International), and Lemmata Sans, derived from Source Sans 3 (Adobe): SIL Open Font License 1.1 (Lemmata, Lemmata Sans).

Banner

Bellerophon, Roman mosaic, Musée Rolin (Autun); photo by Florian Hoffmann, cropped, CC BY-SA 4.0.

Privacy

This site does not measure its audience. The text you paste never leaves the page: Parousia runs in your browser. Only one setting (whether the parsing is shown) and your latest analyses, so that they can be reopened without being recomputed, are kept in this browser (localStorage); the “Forget them” link erases those analyses. Dictionary and inflection lookups, however, query the Eulexis service (Biblissima+) through a relay on this site: the requested lemma is sent to it, and this is the only case in which a cookie is set, the Eulexis session cookie, used only for these lookups and deleted at the end of the browsing session. All other requests merely download the site’s files (model, data, parts of the table of forms).

Feedback is welcome: