DART-FT: Decomposed, Ambiguity-Aware, Risk-Targeted Fine-Tuning of Transformer Models for Ancient Greek Morphological Tagging

Low-resource languages present challenges for generating accurate morphological tagging models due to data sparsity. Languages which exhibit complex morphological structures (internal word structures) further compound the problem of accurate morphological tagging model creation. Ancient Greek is a low-resource language with complex morphology, leading to frequent morphological ambiguity where a single word form can have multiple grammatical interpretations. Standard fine-tuning techniques, such as training with cross-entropy loss, focus on increasing overall accuracy, but can leave systematic errors concentrated in the most ambiguous tokens. Decomposed, Ambiguity-Aware, Risk-Targeted Fine-Tuning (DART-FT) is a targeted fine-tuning method that identifies a slice of tokens (the high-risk slice) associated with recurrent baseline model errors. The method applies multiple filtering passes to dynamically determine which tokens to target for fine-tuning. The targeted fine-tuning process utilizes oversampling. Experiments were conducted with cross-entropy, class-balanced, and class-balanced combined with focal loss, producing a total of 90 distinct models over 5 random seeds on the UD Ancient Greek PROIEL treebank. The model with the largest increase in accuracy on the high-risk slice, as compared to the corresponding standard cross-entropy fine-tuned baseline model, exhibited a statistically significant increase of 6.4 percentage points. Global macro F1 scores for the models were stable and slightly increased. DART-FT enables targeted repairs to error-prone areas of fine-grained morphological tagging models trained on limited labeled corpora and serves as a resource-efficient alternative when additional pre-training is impractical.

Author

• Weber, Matthew K.

Keyword

• Morphology
• Artificial Intelligence
• Ancient Greek

Date created

• 2026

Type of Work

• Map or Cartographic Material

License

• Attribution-NonCommercial 3.0 United States

Rights statement

• In Copyright

GW Unit

• Computer Science

Degree

• D.Engr.

Advisor

• Vijay Raghavan, Professorial Lecturer of Engineering and Applied Science, Praxis Director
• Amir Etemadi, Adjunct Professor of Engineering and Applied Science, Committee Chair
• Mazen Mheish, Professorial Lecturer of Engineering and Applied Science, Committee Member

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/4x51hk097