These tools are under development.
| Author | Source file | Greek word tokens |
|---|---|---|
| Basil | Basil_Homily_Final.txt | 35,268 |
| Chrysostom | Chrysostom_Genesis_hexaemeral.txt | 29,987 |
| Ps.-Eustathios | Eustathios_hexaemeron_complete.txt | 17,217 |
| Glykas | Glykas_complete_merged.txt | 41,372 |
| Gregory of Nyssa | Gregory_Nyssa_hexaemeron.txt | 14,384 |
| Manasses | Manasses_chronicle.txt | 3,068 |
| Mauropous | Mauropous_Etymologia.txt | 2,520 |
| Neophytos | Neophytos_Cyprus.txt | 13,497 |
| Philoponos | Philoponos_complete.txt | 84,382 |
| Ps.-Psellos | Psellos_hex 3.txt | 1,705 |
| Severian | Severian_of_Gabala.txt | 35,809 |
| Anastasios | anastasios_sinai.txt | 32,490 |
| Pisides | pisides_hexaemeron.txt | 10,681 |
| Portaites | portaites_hexaemeron.txt | 12,465 |
| Cosmas Indicopleustes | test_Cosmas_Indikopleustes_1-5.txt | 31,477 |
Displayed Greek word forms come from the corrected source files. Text segments are split at punctuation; their locations identify source-file lines. They are not edition page references.
Automated lemma assignment uses the local CLTK Greek form dictionary. A dictionary lemma is used only when the candidates agree after normalization. Ambiguous or unknown forms are retained as normalized word forms. Case, accents and final/lunate sigma differences are normalized for lemma searching. These assignments are provisional, and the working texts have not been fully collated against editions. Some OCR errors remain.
A new FastText skip-gram model was trained solely on the fifteen corrected texts: 100 dimensions, window 5, 30 epochs, minimum count 3, character n-grams 3–6, five negative samples, bucket 200,000, one training thread and subsampling threshold 0.0001.
The lemma map projects vectors for content lemmata and unresolved forms occurring at least three times into two dimensions with PCA. Colour indicates the author with the largest raw count. Author-distinctive terms have at least twenty occurrences and are ranked by smoothed log odds against the rest of the corpus.
The usage map averages vectors of content words within eight word positions on either side of each occurrence, excluding the target itself. Occurrences without a usable context are omitted. Each target receives its own PCA projection. These exploratory maps support inspection; distances do not establish textual borrowing.
The separate Glykas–Basil comparison preserves the archived September 2026 experiment and its source version.