Location Matters: Accelerating Historical Cipher Transcription with Detection-Based Models

Abstract

Transcribing historical ciphers is the first step toward decryption and analysis. Machine learning models proposed for this task often neither consume nor produce symbol locations. Such location-discarding approaches do not integrate naturally with transcription and analysis tools that rely on a visual feedback loop linking image regions to transcribed symbols. We argue for location-aware, detection-based deep learning models that preserve location information in both supervision and output, supporting an end-to-end visual workflow with tools such as CTTS (CrypTool Transcription and Solver). To this end, we present two detection-based models with distinct architectures, evaluate them quantitatively across diverse cipher collections, and illustrate the workflow through a case study. The results show that this approach is practical and sample-efficient: it performs well with limited training data and remains effective in a challenging yet typical bootstrap scenario for new cipher collections. It supports human-in-the-loop correction, significantly reduces manual work, and helps produce accurate transcriptions and training data.