TL;DR
This paper introduces a novel set of 23 stylometric features for classifying ancient Greek literary texts into prose and verse, and further distinguishing verse into epic and drama. Achieving over 97% accuracy without relying on complex syntactic parsing, this work opens doors for computational literary analysis of pre-modern, non-English traditions.
What Problem Does This Paper Address?
Genre classification is a fundamental task in natural language processing, with extensive research focused on modern English texts. However, applying these techniques to ancient languages, particularly classical Greek, has been largely unexplored. This gap is significant because classical Greek literature forms the bedrock of Western literary tradition, and understanding its genre distinctions computationally could unlock new avenues for literary scholarship. Existing challenges include the lack of robust syntactic parsers for ancient Greek, making feature extraction difficult. This paper aims to bridge this gap by developing a feature set and methodology that bypasses the need for syntactic parsing, enabling accurate genre classification for this rich corpus.
Methods
The core of this research lies in a custom-built "stylometric feature set" designed specifically for ancient Greek. This set comprises 23 features, categorized into pronouns and non-content adjectives, conjunctions and particles, subordinate clauses, and miscellaneous features. Crucially, these features are designed to be computed without requiring syntactic parsing, a significant advantage given the nascent state of ancient Greek parsing tools. The features are primarily based on word frequencies and sentence-level statistics.
For classification, the authors employed a random forest classifier, a standard machine learning algorithm known for its robustness. They utilized a stratified 5-fold cross-validation approach, repeated 400 times with different random seeds, to ensure the reliability and generalizability of their results. Feature rankings were determined by average Gini importance.
Datasets
The study uses a corpus of 751 ancient Greek text files, primarily sourced from the Perseus Digital Library and processed by the Tesserae Project. This corpus encompasses a substantial portion of surviving classical Greek literature, spanning from the 8th century BCE to the 6th century CE.
For the initial prose vs. verse classification, the corpus was hand-annotated into 610 prose files and 141 verse files. For the subsequent epic vs. drama classification, the verse texts were further categorized into 82 epic files and 45 drama files, with 14 files containing other poem genres set aside.
Key Insights and Ideas
- Parsing-Free Stylometry for Ancient Languages: The most significant contribution is demonstrating that accurate genre classification can be achieved for ancient Greek without relying on syntactic parsing. This is a major breakthrough for computationally analyzing historical texts where parsing tools are underdeveloped.
- Language-Specific Features Matter: While some features are common in English stylometry (like pronouns), the inclusion of language-specific elements, such as Greek particles, enhances classification accuracy. This highlights the importance of tailoring features to the linguistic nuances of the target language.
- Pronouns as Prose Indicators: The high ranking of pronouns and pronominal adjectives in prose classification suggests that their frequent use in sustained discussions within prose genres helps avoid noun repetition. This offers a linguistic explanation for a statistical observation.
- Conjunctions and Sentence Length: The strong correlation between conjunctions and prose, along with the generally longer sentences in prose, points to structural differences that are indicative of genre.
- Drama's Distinctive Style: For epic vs. drama, shorter sentence lengths and a higher frequency of interrogative sentences were found to be key discriminators for drama, likely reflecting the dynamic nature of dramatic dialogue and character interactions.
Key Results
The study reports exceptionally high classification accuracies:
- Prose vs. Verse: Over 97% accuracy and F1 score.
- Epic vs. Drama: Comparable performance to prose vs. verse, with an overall accuracy of 99.8%.
The top features for prose vs. verse classification were identified as "αὐτός" (a pronoun/adjective), conjunctions, demonstrative pronouns, reflexive pronouns, and "μέν" (a particle). For epic vs. drama, the top features were mean sentence length, demonstrative pronouns, interrogative sentences, and variance of sentence length.
Limitations and Future Work
The authors acknowledge that some misclassifications, particularly for the epic vs. drama task, occurred with highly fragmentary or short texts. They also note that the feature set is restricted to the Attic dialect to mitigate issues with dialectical variation and homonyms.
Future work suggested includes extending the experiments to encompass a wider range of traditional prose genres (historiography, philosophy, oratory) and applying the feature set to other questions in classical literary criticism. The authors also express hope that their heuristic approach will inspire similar work on other premodern traditions where NLP research is still in its early stages.
Summary updated