Adapting CycleGAN architecture for Unpaired Diachronic Text Style Transfer

Diachronic text style transfer aims to transform text from one historical period into the style of another while preserving its meaning.However, the scarcity of parallel corpora across time periods makes supervised approaches impractical.In this work, we propose to adapt the CycleGAN architecture, originally developed for unpaired image-to-image translation, to model linguistic change over time.Our method employs a generator and discriminator, both conditioned on temporal information, and trained using a combination of adversarial and cycle-consistency losses.We propose a time-conditioned generative framework that supports both discrete and continuous temporal representations, enabling the model to interpolate between historical language styles.The model is trained on unaligned historical texts and can transform language from any period to another.This approach offers a data-efficient solution for diachronic language modeling and opens new research directions in historical linguistics, digital humanities, and unsupervised style transfer.