Lecture by
William Mattingly
Yale University
The application of AI has undergone a fundamental shift in the last decade. Instead of custom-designed and trained neural networks, the current approach often involves leveraging larger, generalized models and adapting them to specific tasks through methods such as zero-shot, few-shot, or finetuning. However, a single model is insufficient to address the diverse challenges in cultural heritage. This lecture will focus on designing and constructing AI pipelines that integrate various components to tackle these issues. We will start with a physical text and explore its conversion into a digital representation. This digitized form then enters the AI pipeline. The initial stage involves visual components, including object detection and either Handwriting Text Recognition (HTR) or Optical Character Recognition (OCR). Following the visual processing, the pipeline transitions to text-based components. We will discuss Named Entity Recognition (NER) and Information Extraction (IE). We will then examine techniques for identifying quotations and textual reuse by comparing the text against other sources. To enhance access and discoverability, we will explore AI techniques like semantic search and retrieval-augmented generation. Finally, the lecture will cover how Large Language Models (LLMs) and knowledge graphs can be utilized to link our data with Linked Open Data (LOD).