Visual Analytics for the Exploratory Analysis and Labeling of Cultural Data

Cultural data can come in various forms and modalities, such as text traditions, artworks, music, crafted objects, or even as intangible heritage such as biographies of people, performing arts, cultural customs and rites. The assignment of metadata to such cultural heritage objects is an important task that people working in galleries, libraries, archives, and museums (GLAM) do on a daily basis. These rich metadata collections are used to categorize, structure, and study collections, but can also be used to apply computational methods. Such computational methods are in the focus of Computational and Digital Humanities projects and research. For the longest time, the digital humanities community has focused on textual corpora, including text mining, and other natural language processing techniques. Although some disciplines of the humanities, such as art history and archaeology have a long history of using visualizations. In recent years, the digital humanities community has started to shift the focus to include other modalities, such as audio-visual data. In turn, methods in machine learning and computer vision have been proposed for the specificities of such corpora. Over the last decade, the visualization community has engaged in several collaborations with the digital humanities, often with a focus on exploratory or comparative analysis of the data at hand. This includes both methods and systems that support classical Close Reading of the material and Distant Reading methods that give an overview of larger collections, as well as methods in between, such as Meso Reading. Furthermore, a wider application of machine learning methods can be observed on cultural heritage collections. But they are rarely applied together with visualizations to allow for further perspectives on the collections in a visual analytics or human-in-the-loop setting. Visual analytics can help in the decision-making process by guiding domain experts through the collection of interest. However, state-of-the-art supervised machine learning methods are often not applicable to the collection of interest due to missing ground truth. One form of ground truth are class labels, e.g., of entities depicted in an image collection, assigned to the individual images. Labeling all objects in a collection is an arduous task when performed manually, because cultural heritage collections contain a wide variety of different objects with plenty of details. A problem that arises with these collections curated in different institutions is that not always a specific standard is followed, so the vocabulary used can drift apart from another, making it difficult to combine the data from these institutions for large-scale analysis. This thesis presents a series of projects that combine machine learning methods with interactive visualizations for the exploratory analysis and labeling of cultural data. First, we define cultural data with regard to heritage and contemporary data, then we look at the state-of-the-art of existing visualization, computer vision, and visual analytics methods and projects focusing on cultural data collections. After this, we present the problems addressed in this thesis and their solutions, starting with a series of visualizations to explore different facets of rap lyrics and rap artists with a focus on text reuse. Next, we engage in a more complex case of text reuse, the collation of medieval vernacular text editions. For this, a human-in-the-loop process is presented that applies word embeddings and interactive visualizations to perform textual alignments on under-resourced languages supported by labeling of the relations between lines and the relations between words. We then switch the focus from textual data to another modality of cultural data by presenting a Virtual Museum that combines interactive visualizations and computer vision in order to explore a collection of artworks. With the lessons learned from the previous projects, we engage in the labeling and analysis of medieval illuminated manuscripts and so combine some of the machine learning methods and visualizations that were used for textual data with computer vision methods. Finally, we give reflections on the interdisciplinary projects and the lessons learned, before we discuss existing challenges when working with cultural heritage data from the computer science perspective to outline potential research directions for machine learning and visual analytics of cultural heritage data