From Text to Vector Space: Modeling Gendered Semantics in Ottoman Turkish

Abstract

This article examines how gendered meaning was constructed in late Ottoman Turkish periodicals by applying word-embedding models to a research-grade corpus that we digitized and processed through Optical Character Recognition/Handwritten Text Recognition. Ottoman Turkish presents distinctive challenges for computational analysis: it was written in a now-obsolete script, incorporates extensive Arabic and Persian vocabulary, and has agglutinative morphology. Very little of this material has been digitized, making Ottoman Turkish a severely under-resourced language for computational study. In light of these constraints, we approach corpus preparation not as a neutral prelude but as a constitutive stage of the research itself, since text recognition and preprocessing shape the semantic relations recoverable through modeling. Using Word2Vec, we analyzed gendered semantics through similarity measures, clustering, and projection along a constructed gender axis. Across parameter settings, the results converge on a consistent asymmetry: femininity is explicitly marked within reformist and domestic registers, while masculinity appears as an unmarked and diffuse norm spanning civic, institutional, and familial vocabularies. We argue that this structure arises not from grammatical encoding but from cultural discourse, and that word embeddings provide a heuristic means of rendering such asymmetries analytically visible. More broadly, the study demonstrates how computational methods can serve interpretive historical-linguistic inquiry in non-Western and digitally under-resourced contexts, where corpus construction and modeling decisions are inseparable from scholarly interpretation.