Daniel Vollmers successfully completed his doctorate on the topic of "Neural Entity Linking for Question Answering over Knowledge Graphs" under the supervision of Professor Dr. Axel-Cyrille Ngonga Ngomo. Congratulations on this achievement!
Summary of the work:
One of the central applications of the Internet continues to be the targeted search for correct information and the exploration of new or previously unknown topics. In this context, search engines and related applications that retrieve relevant documents, such as web pages, in response to user queries have long played a central role in research. However, the information needs of users are often more complex and require the integration of information from multiple sources, for example when determining the most favourable offer on a car or answering factual questions such as "Which museums are located in Paderborn?".
This thesis deals with the research area of Question Answering. Recent advances in the development of large language models (LLMs), which are trained on large text corpora, have fundamentally changed this field of research. Although LLMs represent the state of the art, they are prone to hallucination, that is, they produce answers that are not always factually correct or based on reliable knowledge. This limitation is particularly problematic given the prevalence of disinformation on the Internet. To overcome this problem, enriching LLM prompts with information from trustworthy knowledge sources, such as text corpora, databases or knowledge graphs (KGs), offers a promising approach. In this Retrieval Augmented Generation (RAG) paradigm, information is first retrieved from knowledge sources and provided to the model as additional input to improve factual accuracy. In the case of knowledge graphs, this task is referred to as Question Answering over Knowledge Graphs (KGQA). To answer questions based on the information of a knowledge graph, semantic parsing techniques can be applied, which translate a natural language question into an executable SPARQL query. In this context, the thesis focusses on the research and application of neural entity linking methods for the research topic Question Answering over Knowledge Graphs (KGQA). Methods for contextual enhancement, missing keyword extraction and robust retrieval are developed to improve the identification of entities, relations and types (ERL). Furthermore, we analyse how ERL influences the overall performance of KGQA systems.
