Exploring intertextuality across the Homeric poems through language models

Maria Konstantinidou, John Pavlopoulos, Elton BarkerACL 2024

TL;DR

This paper explores the linguistic connections and surprising divergences between Homer's Iliad and Odyssey using language models. By analyzing verse-level "surprise" scores, they identify passages that are more linguistically akin to the other epic than their own, offering new insights into the oral composition and potential intertextuality of these foundational works.

What Problem Does This Paper Address?

The origins and composition of Homer's epics, the Iliad and the Odyssey, have long been debated. While scholars agree on the importance of oral tradition and formulaic language, precisely how these poems relate to each other and to their oral heritage remains a complex question. This paper aims to move beyond simply identifying formulaic language to quantitatively assessing the linguistic "affinity" between the Iliad and the Odyssey at a fine-grained, verse level. They want to understand if and how specific verses or passages might have been influenced by, or resonate more strongly with, the other epic, potentially revealing patterns of interpolation, conscious borrowing, or shared compositional strategies.

Methods

The core of this research relies on character-level statistical language models trained on the Homeric poems. They utilize the HoLM (Homeric Language Models) resource, which provides "cross-score" metrics. This metric quantifies the perplexity (PPL), a measure of how "surprising" a verse is to a language model.

They employ three levels of analysis:

  1. Immediate surroundings: Comparing a verse's surprise to its local context (preceding/following verses) versus its book.
  2. Individual book > Source poem: Comparing a verse's surprise within its own book to its surprise when the model is trained on the entire source poem.
  3. Source poem > Other poem: This is the key innovation. They compare a verse's surprise when modeled by its own epic versus when modeled by the other epic. A positive cross value (PCV) indicates a verse is more surprising to its own poem's model than to the other poem's model, suggesting a stronger affinity with the other epic.

Additionally, they use supervised machine learning algorithms (KNN, Logistic Regression, Random Forests) to classify verses as "source-surprising" or not, providing a complementary approach to the PCV analysis.

Datasets

The primary dataset is the HoLM resource, which contains character-level language models and associated perplexity scores for the Iliad and the Odyssey. The paper uses specific editions of the Iliad (Allen, 1931) and the Odyssey (von der Mühll, 1962). For the machine learning classification, they trained models on the entirety of both poems, excluding certain books for evaluation, resulting in approximately 12,103 verses from the Iliad and 9,861 from the Odyssey for training.

Key Insights and Ideas

  • Linguistic Bridges Between Epics: The analysis reveals that a significant number of verses are linguistically more "at home" in the other Homeric poem than their own, suggesting a deeper intertextual relationship than previously quantifiable.
  • PCVs as Indicators of Influence: Positive cross-score (PCV) verses are highlighted as potential indicators of influence, interpolation, or shared compositional practices, pointing to specific passages that might warrant closer literary scrutiny.
  • Book-Level Hotspots: Certain books, particularly Book 24 of the Iliad and Books 11, 15, 22, and 24 of the Odyssey, show a higher concentration of these "source-surprising" verses, suggesting these books might be particularly rich in inter-epic connections.
  • The "Homeric Question" Revisited: While not aiming to definitively answer the authorship question, the method provides a quantitative lens to explore how linguistic patterns might reflect different compositional hands or influences.
  • Orality and Language Models: The study reinforces the idea that oral poetry functions like a language, and that language modeling can effectively capture nuances of traditional poetic practices.

Key Results

  • The Iliad contains a higher number of PCVs (511, or 375 after removing duplicates) compared to the Odyssey (272, or 235 after duplicates), even accounting for its greater length.
  • Machine learning classifiers largely corroborate the PCV findings: 79% of Iliad PCVs were flagged as "Odyssean" by classifiers, and 71% of Odyssey PCVs were flagged as "Iliadic."
  • The ML classifiers also identified a substantial number of additional "source-surprising" verses not flagged by PCVs alone, suggesting a broader scope for identifying inter-epic connections.
  • Logistic Regression (LR) performed best among the tested ML algorithms for classifying verses, achieving high precision, recall, and F1 scores.

Limitations and Future Work

  • Defining "Surprise": The reliance on perplexity as a measure of "surprise" is a proxy for linguistic affinity and may not fully capture all nuances of literary influence or intentionality.
  • Book-Level Granularity: While book-level analysis is informative, the authors note that the low number of PCVs in some books makes reliable statistical analysis challenging.
  • Future Work: The authors suggest focusing on shorter verse clusters, exploring thematic and lexical features associated with each poem, and examining the Homeric poems alongside other ancient authors and genres (like Hesiod and lyric poetry) to further contextualize their findings. They also propose investigating specific types of content, such as catalogs, direct speech, and proper names.

Summary updated

BibTeX
@inproceedings{konstantinidou2024exploring,
  author = {Maria Konstantinidou and John Pavlopoulos and Elton Barker},
  title = {{Exploring intertextuality across the Homeric poems through language models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure