Uncovering Tibet’s Oldest Religion through AI-enhanced Handwritten Text Recognition

This poster presents our AI-enhanced workflow to digitise a vast corpus of manuscripts, hand-written in an extremely difficult form of Tibetan. These offer a unique glimpse into archaic non-Buddhist rituals and narratives, enabling us to reconstruct Tibetan Pagan religion for the first time. The palm-leaf page layout differs from any previously trained models and the wide range of abbreviations, a mixture of scripts and languages, and varying image quality make the task extremely challenging. To tackle these challenges, we developed a unique collaborative Handwritten Text Recognition (HTR) workflow making optimal use of new technologies while bringing together a diverse range of expertise from multiple stakeholders. Central to this workflow is the custom-developed PechaTools that streamlines the process of transcribing from scratch and correcting computer-generated output, with multiple proofreading stages to ensure accuracy. Within six months, the above-sketched workflow enabled us to successfully train Layout Analysis (LA) and HTR models using Transkribus. We furthermore show how we improve results by providing detailed feedback to the correction teams (including customised glyph-lookup dictionaries), optimising parameters as well as fine-tuning. We also discuss the trade-off of getting more-readable results by using dictionaries and language models that yield more normalised text, which can be useful for a more general audience, but is highly detrimental to retrieving essential philological details. We also discuss postprocessing leading to the next steps of creating both diplomatic as well as normalised transcriptions to make these unique text collections maximally useful for researchers in a wide range of disciplines.