Half-time seminar: Micaella Bruton, PhD student at the Department of Linguistics, Stockholm University

Abstract

World archives contain thousands of historical documents written in cipher –diplomatic correspondence, religious texts, personal letters, and other manuscripts – many of which remain unread. Decrypting these texts is challenging not only cryptographically but also linguistically: orthographic variation across centuries, distorted symbol distributions, noisy transcriptions, and the scarcity of aligned plaintext–ciphertext data all complicate analysis.

In this talk, I present the first half of my doctoral work on applying neural methods to historical cipher decryption. I introduce ChronoFidelius, a Python library for generating historically grounded synthetic ciphertext, and HistCiph, a multilingual dataset of plaintext–ciphertext pairs spanning ten languages and eight centuries. Building on these resources, I present initial experiments with neural models for ciphertext–plaintext alignment and key extraction, deceipherment, and direct decryption, highlighting both their potential and current limitations.

On Micaella Bruton