Thesis Advisement

Topic

Contact 

New initialisation technique for topic models

As a rule, topic models are initialised by random assignments. Unfortunately, this means that the repeatability of the inference is not very stable and the topics look slightly different. The aim of this thesis is to investigate how initialisation using similarities from embedding methods (language models) can help to design the initialisation in such a way that this instability is reduced and the results become more coherent.

Automation of topic model evaluation with language model prompts

The Tea Leaves test is a way to evaluate topic models by asking a human (acting as an oracle) to identify which word from a group of words does not belong to a given topic. The underlying idea is that topic models, such as Latent Dirichlet Allocation (LDA), generate clusters of related words, and a well-performing model should produce word groups that make intuitive sense to humans. Automating this test with language models (LLMs) involves using LLMs to simulate human judgment. By prompting an LLM to determine which word seems out of place in a topic, we can use the model's predictions to assess the coherence of the topics. LLMs, with their vast knowledge and contextual understanding, can effectively mimic human responses, providing a scalable way to evaluate topic model performance without requiring human intervention.

Automated Analysis and Contextualization of Archaeological Artifacts from Digital Collection Catalogs Using Modern AI Methods.

This research would explore how information from digital archaeological collection catalogs can be extracted and automatically processed to gain insights into artifacts, their historical significance, and cultural contexts. Natural Language Processing (NLP) techniques could be applied to analyze descriptive texts, metadata, and contextual information. Additionally, computer vision techniques could be employed to identify patterns in images of artifacts, comparing them to known findings. The goal would be to make digital archaeological collections more accessible while using AI to enable deeper, automated analysis and interpretation of the data, potentially leading to new scholarly discoveries.

Historical research expeditions

The archives of the Leibniz Institut für Länderkunde (IfL) contain, among other things, collections of the photographic material of various research expeditions. The digitalisation of the documents of a specific explorer or a specific group of researchers lends itself as a thesis subject. Subsequently, these should be assigned to the places depicted and the resulting route visualised.

Immersive Reading: Augmented reality for books

This rather experimental topic is dedicated to the extent to which AR technologies can be used to meaningfully supplement physical books with statistics and other distant reading overlays on the respective chapter.

Augmented reality as a tool for collaborative annotations

The annotation of data plays a fundamental role in the digital humanities. When working with very large amounts of data, this work cannot be done by individuals, but only collaboratively. Especially when working with non-textual documents (such as maps or scans of physical objects), AR technologies open up new possibilities for collaborative annotation, which are to be explored in this thesis.

Mental visualisation of literary passages 

Unlike visual media, images of literary works are created in the minds of readers and are therefore strongly dependent on subjective factors. Utilising Stable Diffusion, passages from literature can be visualised in ways that incorporate diverse cultural patterns. With the results, a test person study is to be conducted in order to find out how well each visualisation corresponds with the "image in the mind" of readers from different backgrounds. Due to its scope and complexity, this topic is primarily suitable for M. Sc. students.

Literary Animal Embeddings 

Anthropomorphic animals appear again and again in literature. This topic deals with the question of whether word embeddings can be used to determine whether the characterisation of these figures aligns more closely to humans or to the respective animal, and how this relationship changes throughout history or over the course of a literary work.

Audio Description

Sound effects are an important component in film and television. We are interested in finding out which features or descriptors are suitable to describe sound effects in such a way that similarities between effects of the same category can be shown. In this way, we want to investigate whether sound reuse in sound design and music production can be measured with quantitative methods. One possible research question is, for example, which sound is continued in which genres (e.g. the warp drive noise in the Star Trek franchise).

Position of light sources in paintings

In art studies, a certain symbolic power is attributed to the lighting of the subjects, especially in historical works. According to the Distant Viewing principle, the aim is to identify the position of the light source in relation to the subject in paintings of different epochs using computational methods. On the basis of this data, changes in the artistic use of light over time can be examined and classified using existing art historical findings.

New algorithm for dominant colour analysis

Colour design is an important stylistic device in cinematography. Existing tools for capturing dominant colours usually identify those clusters of similar colour values that occupy the largest area of the image. However, smaller clusters of colour that are nonetheless very prominent to the human eye despite cannot be reliably identified with these methods, but are certainly used as stylistic devices by filmmakers. An extreme example is the character of the girl in the red coat in the film "Schindler's List", whose colour can easily be identified by human viewers as the dominant colour in the otherwise grey scenes. The goal is to develop an algorithm that can better recognise the use of such contrasting colours.

Textual patterns of the Hero’s Journey

In his 1949 work The Hero with a Thousand Faces, Joseph Campbell conceptualised the hero's journey (or monomyth) as a narratological pattern underlying many stories. Since then, various academic papers have dealt with this theory and critically examined it. A thesis on this topic could explore whether a distant reading approach can be used to identify textual patterns that suggest a certain archetype or station in the hero's journey.

Exploring the Taishō Shinshū Daizōkyō with the help of topic modelling

The Taishō Shinshū Daizōkyō (jap. 大正新脩大藏經, Eng. "Taishō-revised Tripiṭaka") is a fundamental text in the Chinese Buddhist canon. The texts refer to different Buddhist schools. We are interested in using topic modelling to explore questions such as these: Can religious patterns be identified in the texts? What are the main topics? Are the topics evenly distributed or is there a clustering of certain topics around certain dates? Pre-processing has already taken place.

Colour analysis of video games

This topic also concerns the field of Game Studies and aims to explore the relationship between the genre of a game and its colour design, as well as the development of this relationship over time.

Detecting cultural patterns in music sampling

WhoSampled is the world's largest music discovery service for sample-based music, covers and remixes. We want to use this data source to explore the question of whether there are cultural patterns in music sampling. Are there significant temporal relationships between music styles and popular culture? Can films with strong music reuse be identified from their soundtrack data and are there any connections to the structure of the film? Or is the phenomenon of music reuse in films dependent on genre or year of release?