One of the increasingly important NLP tasks is claim verification (or automated fact-checking), where the goal is to validate a claim’s veracity based on relevant evidence. The process usually consists of retrieving evidence related to the claim and then assessing whether it supports the claim. There have been numerous datasets constructed for the task in various domains, including encyclopedic knowledge, society, and rumors. Especially challenging is the scientific domain, in particular covering biomedical and health-related claims, which can often be uncertain and change with time.
In scope of the Master Thesis, a student would explore some of the challenges related to claim verification and related knowledge-intensive tasks like question answering. The thesis would consist of posing research questions, doing literature search, implementing and conducting experiments, and evaluating the results. Some starting ideas that can guide the direction of the thesis include:
- Open-ended claim verification, where the relevant evidence and labels are not assumed to be known and the challenge becomes how to effectively search through large knowledge bases to find appropriate evidence.
- Uncertainty in verification, which includes detecting and assessing claims that change their nature over time, claims with highly conflicting evidence, claims with insufficient evidence, etc.
- LLMs in claim verifcation, which explores how to utilize the few-shot learning capabilities and multi-turn nature of generative LLMs to help the process of verifying facts.
- Explainability of verification, which seeks to provide clear and relevant justifications for a given decision and make the whole process interpretable to humans and human needs.
Ideally, a candidate should have prior knowledge of machine learning and natural language processing, and experience with programming in Python and using NLP frameworks like HuggingFace. If you are interested, please send a CV, transcript of records, and a short motivation e-mail to Juraj Vladika (juraj.vladika@tum.de)
There are no subpages or files.