Character-level language-model analysis of the Iliad and Odyssey at Book resolution (Fasoi, Pavlopoulos and Konstantinidou, 2021) showed wide within-Book perplexity variation and a linguistically distinct block of Odyssey Books 9-12, but could not decide whether such patterns are genuine authorship signals or artifacts of the late, arbitrary Book division, and it examined only four of the thirty-three Homeric Hymns. The same group's follow-ups have since realized the sub-Book (600-character) and verse-level analyses and localized Odyssey Book 11's deviation to specific contested passages. The surviving open problem is to complete the fine-grained computational authorship analysis of the Homeric corpus: extend language-model proximity analysis to all thirty-three Homeric Hymns; determine, with sub-Book units and register controls, whether the Apologoi (Odyssey 9-12) divergence reflects distinct authorship or narrative and divisional artifacts; validate character-level perplexity as an authorship signal for multi-poet attribution within a single ancient Greek tradition; and re-establish the human-expert baseline under content-controlled conditions with Ancient Greek language specialists.

Background

The Iliad and the Odyssey, the two epic poems attributed to Homer (traditionally dated to the eighth century BCE), stand at the center of the oldest continuous authorship debate in Western scholarship, the Homeric Question: whether the poems go back substantially to a single poet or accreted over generations of oral performers before being written down. Modern scholarship agrees that both poems emerged from an oral, formulaic tradition — composition in performance using ready-made phrases fitted to the dactylic hexameter meter — and that the surviving text absorbed contributions from multiple poets, performers, scribes, and editors. Individual passages and even whole Books have long been suspected of being interpolations, i.e., later insertions into an earlier text; famous cases include the Doloneia (Iliad Book 10), the Catalogue of Ships (Iliad 2.494-759), and parts of Odyssey Book 11.

Two structural facts shape any computational attack on this question. First, the division of each poem into 24 Books (rhapsodies) is not original: it was imposed by later, probably Alexandrian, grammarians in the Hellenistic period and does not follow natural breaks in content or language. Second, the works once attributed to Homer also include the Homeric Hymns, thirty-three hexameter hymns to individual deities ranging from 3 to 580 verses, of which only four (to Apollo, Aphrodite, Demeter, and Hermes — the so-called great hymns) are long enough for traditional philological authorship arguments.

Fasoi, Pavlopoulos, and Konstantinidou (2021) approached the question with character-level statistical language modeling, a stylometry technique in which a language model is trained on a text at the level of individual characters rather than words; an excerpt that the model finds surprising is linguistically remote from the training corpus. The technical measure is perplexity (PPL), the exponentiated average negative log-likelihood per character, for example for a character trigram model:

PPL=exp⁡(−1N∑i=1Nlog⁡p(ci∣ci−2,ci−1))\mathrm{PPL}=\exp\left(-\frac{1}{N}\sum_{i=1}^{N}\log p(c_i\mid c_{i-2},c_{i-1})\right)

with lower values indicating closer match to the training distribution. Character n-gram models of this kind had previously supported authorship attribution of modern languages and genre classification of ancient Greek prose and verse (Gianitsos et al., 2019). Training one trigram model per Book on all other Books of the same poem (a leave-one-Book-out design) and correlating the resulting PPL profiles, the 2021 study found: (i) large within-Book variation of excerpt-level PPL — scores from about 2.5 to 18 across excerpts of Books whose means all lie between 6 and 9 — most pronounced in Iliad Books 14 and 16 and Odyssey Books 9, 19, and 24, indicating that the Book is too coarse a unit; (ii) in the Odyssey, Books 9, 10, and 12 — three of the four Books of the Apologoi (Odyssey 9-12), Odysseus's first-person account of his wanderings told at the Phaeacian court — form the most substantial linguistically divergent block, mutually strongly correlated and negatively correlated with the rest of the poem, raising the question whether a whole block was added by a different hand; and (iii) Book 11 (the Nekyia, the underworld episode interrupting the 9-12 run) is itself linguistically at odds with much of the poem, though it shows no particular affinity with Books 9 and 10. A classifier built from these models separated Iliad from Odyssey excerpts about as well as the average human expert, and of the four hymns studied, To Aphrodite proved linguistically close to both epics and To Hermes remote.

The authors' own follow-ups realized part of the program that this list of unresolved issues set out. Pavlopoulos and Konstantinidou (2022) re-ran the analysis with 600-character rolling excerpts as the unit of analysis, isolating 47 outlier passages, including much of the Catalogue of Ships and, in Odyssey 11, a run of about 100 verses (11.230-331) coinciding with the long-contested introduction to the Catalogue of Heroines — thereby localizing Book 11's deviance to specific, philologically disputed passages — and verified the models against Hesiod's Theogony and Works and Days. Pavlopoulos, Sandell, Konstantinidou, and Bozzone (2024) released HoLM, a verse-level corpus of linguistic-unexpectedness scores for both poems, showing that the scores correlate strongly with word and character-n-gram frequency and proposing a frequency-corrected ratio; Konstantinidou, Pavlopoulos, and Barker (2024) used this resource to study cross-poem affinity at verse level, flagging Books 1, 9, 10, and 24 of the Iliad and Books 11, 15, 22, and 24 of the Odyssey. What these follow-ups do not do is analyze the other twenty-nine hymns, settle the interpretive status of the Apologoi cluster, test the method on a genuinely multi-poet single-tradition attribution problem, or re-design the human-expert baseline.

Problem Statement

The problem — the surviving open core of the research program initiated by Fasoi et al. (2021), after the authors' own 2022-2024 follow-ups realized the sub-Book (600-character) and verse-level analyses and localized Odyssey Book 11's deviation to contested passages — is to complete the fine-grained computational authorship analysis of the Homeric corpus with character-level statistical language models, at units of analysis shorter than a Book. It comprises four open questions directed toward one objective, a validated fine-grained authorship map of the Homeric corpus:

(1) Hymn corpus. Extend the language-model proximity analysis from the four great hymns to the entire corpus of thirty-three Homeric Hymns, characterizing each hymn's linguistic affinity to the Iliad, to the Odyssey, and to the other hymns, at a resolution appropriate to hymn length, and testing all three grouping variables — length, addressee, and putative date — to assess whether the collection differentiates by any of them rather than forming one homogeneous group.

(2) The Apologoi. Determine, using sub-Book units together with controls for narrative register (the situational and stylistic variety of the language, such as a first-person tale told at a feast versus third-person epic narration), for direct speech versus narrative, and for thematic content, whether the linguistic distinctness of Odyssey Books 9-12 (in particular the 9-10-12 cluster) is evidence of distinct authorship — for example a separately composed poem incorporated into the Odyssey — or an artifact of the late Book division acting on a self-contained narrative block, or of the style of the first-person wanderings narrative.

(3) Method validity. Test character-level language-model perplexity, including its frequency-corrected variants, as an authorship signal for ancient Greek in attribution scenarios beyond the two-epic contrast and the already demonstrated Homer-versus-Hesiod verification: attribution among multiple poets working within a single tradition (for example Homeric, Hesiodic, and hymnic hexameter considered jointly), and checking the scores against known or philologically debated attributions.

(4) Human baseline. Re-run the expert classification experiment with Ancient Greek language and literature specialists (rather than Homer specialists) under a design that separates linguistic from content-based recognition, so that the machine-versus-human comparison measures linguistic classification as claimed.

These four components jointly constitute the single objective above: a hymn-by-hymn characterization of linguistic affinity and grouping (1); a criteria-based ruling on whether the Apologoi's distinctness is attributable to authorship, Book division, or narrative register (2); a verdict on whether character-level perplexity carries authorship information in the stated attribution scenarios, including any case in which it fails (3); and a rigorous, fully specified experimental design separating linguistic from content-based recognition (4).

References

  1. Maria Fasoi, John Pavlopoulos, and Maria Konstantinidou. Computational Authorship Analysis of Homeric Language. In Proceedings of the Digital Humanities Workshop (DHW '21), Kyiv, Ukraine, December 23, 2021, pages 78-88. ACM. DOI: 10.1145/3526242.3526256. https://doi.org/10.1145/3526242.3526256
  2. John Pavlopoulos and Maria Konstantinidou. Computational authorship analysis of the Homeric poems. International Journal of Digital Humanities, 5(1):45-64, 2022. DOI: 10.1007/s42803-022-00046-7. https://doi.org/10.1007/s42803-022-00046-7
  3. John Pavlopoulos, Ryan Sandell, Maria Konstantinidou, and Chiara Bozzone. HoLM: Analyzing the Linguistic Unexpectedness in Homeric Poetry. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 8166-8172, 2024. DOI: 10.63317/297itczqip3w. https://aclanthology.org/2024.lrec-main.715/
  4. Maria Konstantinidou, John Pavlopoulos, and Elton Barker. Exploring intertextuality across the Homeric poems through language models. In Proceedings of the 1st Workshop on Machine Learning for Ancient Languages (ML4AL 2024), pages 260-268, 2024. DOI: 10.18653/v1/2024.ml4al-1.25. https://aclanthology.org/2024.ml4al-1.25/
  5. Natale Musso. Iliad, Odyssey, and statistics. The Journal of Classics Teaching, 27(53):16-21, 2025. DOI: 10.1017/S205863102510024X. https://doi.org/10.1017/S205863102510024X
  6. Joel P. Christensen. State of Play: Computing Homer from Wissenschaft to AI. American Journal of Philology, 146(4):577-589, 2025. DOI: 10.1353/ajp.2025.a980346. https://doi.org/10.1353/ajp.2025.a980346
  7. Colin Swaelens, Maxime Deforche, Ilse De Vos, and Els Lefever. Verse, Variation, and Meaning: Annotating Similarity in Greek Marginal Texts. Journal of Open Humanities Data, 12, 2026. DOI: 10.5334/johd.587. https://doi.org/10.5334/johd.587
  8. Efthimios Gianitsos, Thomas Bolt, Pramit Chaudhuri, and Joseph P. Dexter. Stylometric Classification of Ancient Greek Literary Texts by Genre. In Proceedings of the 3rd Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2019), pages 52-60, 2019. DOI: 10.18653/v1/w19-2507. https://aclanthology.org/W19-2507/