Introduction and Overview
In recent years, Automated Text Recognition (ATR) has made great developmental strides, especially regarding handwritten sources. ATR is the technique of having documents read by the computer or more specifically by trained models. Deep-learning algorithms form the core of ATR and are also important components in today’s flourishing AI, providing e.g. the basis for Chat-GPT. In other words, these are trained language models that excel at sentence formation and content retrieval.
As time-efficient indexing of large handwritten text corpora advances, ATR tools become all the more crucial for research on archival documents. This tutorial serves as an introduction to the methods of applying ATR to different types of corpora, ranging from smaller to more extensive research projects, and therefore complements the existing transcription exercises on Ad fontes. Please note that the user‘s ability to read and comprehend sources remains a fundamental requirement in this process. On Ad fontes, you also find an introduction to paleography as well as excersises for scripts in different languages and from various ages.
This module will
- teach you what to consider when working with ATR
- teach you key metrics and terms in working with ATR
- introduce you to the modern ATR workflow on a conceptual level
This module will not
- give detailed instructions on specific platforms or providers as they change quickly. For this information, please refer to the individual providers.
The first chapter provides a general introduction to ATR: how does ATR work and what does it contribute to historical research?
The second chapter outlines various preliminary factors that should be considered when using ATR: we establish four dimensions of ATR (heterogenity of hands, amount of text, research question, and method). These dimensions are then used to examine what constitutes a good transcription (depending on specific goals). Our main aim is to show when the use of ATR is time-efficient and which factors play a role in this.
The third chapter uses four examples to explain how to apply the preliminary considerations to specific research scenarios.
The fourth chapter consists of an interactive tool which shows the four outlined dimensions of ATR and all possible combinations. Here, you can find some general recommendations on how to proceed with your own sources, once you have ascertained the four established dimensions (part 2) for your own project.
The fifth and last chapter provides more detailed information on how to select, test and train ATR models, the foundation of text recognition in your own documents. Here, we answer the question: how do you find a good text recognition model for your documents?