Thesis Topics

My areas of research are in trustworthy NLP, discourse, semantics, pragmatics (particularly figurative language and sentiment analysis) and summarization. Please contact me if you are interested in a BA or MA thesis in those topics. As I am currently oversubscribed, the following holds: For MA students, preference goes to students who have done the research module with me. In other cases, students must have done at least 2 seminars with me or with my doctoral student Jakob Schuster. For BA students, students must have done at least 2 seminars/software project with me or my doctoral student Jakob Schuster. Note that I normally require you to register for your thesis after 4 (BA) or 5 (MA) meetings. Exceptions due to heavy workload for financing your studies or due to ongoing, chronic health concerns must be discussed with me at beginning of supervision. Sudden interruptions due to sudden illnesses or family matters during supervision will be handled in a flexible manner. Current ideas for thesis topics and areas include but are not limited to the following:

1. Influence of Sources on LLM judges, Question Answering or Other Tasks (BA/MA)

   In a recent paper we have shown in a synthetic world in a multiple choice question setup that language models change their answer substantially between several conflicting answers if answers are attributed to various actors (such as newspapers or governments), leading to a source preference hierarchy. In addition, we have shown that simple repetition of low-reliability sources can override this hierarchy, making LLMs vulnerable to being flooded with misinformation (akin to the illusory truth effect or bandwagon bias in humans). Future Students can expand this framework to (i) real-world scenarios (ii) free-form answers (iii) different languages (iv) different tasks (v) the interaction between source preferences and cognitive biases other than illusory truth/bandwagon effects.

2. Language Models for Timeline Summarization (MA).

   Sebastian Martschat and me developed a submodular framework called TILSE for the generation of timeline overviews such as timelines on long-running wars or other events. One question is whether one can integrate language models into that framework, for example by letting the algorithm work on language model generated summaries of individual articles instead of on the whole article which should also make it more efficient.

3. Bias in Language Models: The case of Sinti and Roma (BA/MA).

   It is well known that Language Models reproduce and even amplify bias against minorities. Most research has concentrated on English as well as gender bias but also racial bias against Blacks or religious bias against Jews or Muslims. To the best of my knowledge, there is no work on language model bias against Sinti and Roma. This thesis would establish the first benchmark test suite for bias against Sinti and Roma and evaluate language models against this benchmark.

4. Measuring Linguistic Capabilities of Large Language Models(BA).

   It is unclear how good the linguistic and metalinguistic capabilities of language models are with conflicting results on grammaticality judgements, dependent on whether you measure via prompting or directly via probability measurements. This work would continue prior work on this topic and extend it towards discourse and pragmatic NLP problems.

5. Cryptic Crosswords (BA/MA)

   Whereas standard crosswords are an almost solved problem in NLP (at least for English ) , results for cryptic crosswords that use clues needing character-level and phonological knowledge as well as making use of word play and puns are extremely low. This thesis would investigate the use of expert modules (such as anagram solvers) to improve the state-of-the art and/or the integration of fine-grained clue annotation to help LM solvers. As an alternative, one could investigate the integration of LLM individual cue solution into a search-based solution for a whole crossword grid. If you are interested in algorithms for other word games, instead of cryptic crosswords, this might also be possible.

If you do not find anything you like, but still would like something in the area of trustworthy NLP, summarization, semantics/pragmatics or discourse please contact me.