A Robust Optimization Framework for Historical Word-Sense Disambiguation in Digital Philology

Abstract

In historical word-sense disambiguation, interpretation of a language with meanings, spelling norms, and texts that might differ from those used by current language models is required. Incomplete documents, unknown dates, and insufficient dictionaries could render an unambiguous interpretation with one top score doubtful. This article discusses the use of a robust optimization method for the ranking of historical senses based on contextual, lexical, and historical evidence. Each score of evidence is accompanied by a finite adversarial deviation, and a robustness budget regulates to what extent the scores could suffer deterioration at the same time. The optimal interpretation is defined by maximizing the minimum value of its score within the given uncertainty set. The example case is concerned with three possible interpretations of the word “art.” In the nominal scoring procedure, the interpretation that is deemed most appropriate is the aesthetic one with the score 0.8650, followed by the practical skill interpretation with the score 0.8200. Under the robustness budget of one, the winning candidate would be the practical skill with the score 0.7700 and the aesthetic interpretation with the score 0.6650. However, both interpretations would still be eligible for consideration due to overlapping evidence intervals. The numeric approach shows a sensitivity to uncertainty of evidence but not the accuracy evaluated against a historical corpus. This methodology makes it possible to provide inspectable interpretation by documenting the evidence, assumptions, and perturbations involved in each ranking.