Improvements to the Explanations of Ithaca’s Chronological Attributions of Ancient Greek Inscriptions

Abstract
Transformer-based models, such as Ithaca, have proven effective in dating ancient Greek inscriptions. However, their chronological predictions remain challenging for scholars to interpret and trust. This thesis enhances Ithaca’s chronological-attribution component by integrating andsystematically comparingpost-hocexplainabilitymethodsappliedtoboth character-level and word-level embeddings. We introduce a unified saliency-processing pipeline that normalizes and fuses explanations into clear, token-wise heatmaps. While preserving Ithaca’s original dating accuracy, our refined explanations consistently spot light historically relevant names and terms. The accompanying code and methodological frameworkenhancetheinterpretability of Ithaca’s chronological attributions, helping historians validate and understand the model’s decisions.

Recommended citation: Schoinaki, M., 2025. Improvements to the Explanations of Ithaca’s Chronological Attributions of Ancient Greek Inscriptions, BSc Thesis, Athens University of Economics and Business, School of Information Sciences and Technology, Department of Informatics, Athens, Greece. Supervisors: Ion Androutsopoulos, John Pavlopoulos, Yannis Assael.
Download Paper