Welcome to this data labelling guide for historical corpora annotation. This document provides the information and context necessary for your work as an annotator. It is available to you throughout your assignment and you should refer to it if you have any questions.
We want to produce annotations to train a model to detect different types of content in historical documents. To train the model, the annotations must be as consistent as possible.
Our goal is to create an automated system for extracting visual elements using machine learning technology. For this, we need high-quality data, annotated and manually corrected by human hands.
We are currently working on a study of the dissemination of texts by studying images. In this project, we are only interested in the visual elements and we will completely disregard text.
The page of a historical document is composed of text, illustrations, ornaments and other elements added after the creation of the object, such as glosses, comments, stamps, drawings and traces of wear.
You are an annotator and your goal is to annotate the visual elements of the page. To do this, you will draw straight rectangles around the elements that have a label. The 5 labelling tags are the following:
The corpus annotated with these guidelines, and the detection model trained on it, are both openly available:
Annotated pages of historical documents, labelled with the 5 classes defined in this guide.
Multi-class object detection model trained on the dataset above.
Each class is described in detail, with rules and edge cases, in the class definition section.