Liri Fang's Dissertation Defense
PhD Candidate Liri Fang will present her dissertation defense, "Modeling of Graph and Text Data: Applications in Citation Networks, Taxonomy Expansion, Entity Resolution and Claim Verification." Liri's dissertation committee includes Associate Professor Vetle I. Torvik (Chair), Professor Jingrui He, Associate Professor Halil Kilicoglu, and Professor Bertram Ludäscher.
Abstract
Modern AI systems are increasingly built on pretrained language models and large language models, yet many high-stakes settings require more reliability than sequence-only modeling can provide. This dissertation argues that explicit structure is a necessary foundation for improving reliability and interpretability in graph–text modeling. Across the thesis, structure appears as citation graphs, taxonomy hierarchies, external structured knowledge for entity resolution, and evidence networks for claim verification. The recurring challenge is therefore not only to understand text, but also to construct, validate, expand, and reason over structure in ways that make downstream decisions more trustworthy.
To support this claim, the dissertation develops and evaluates structure-aware methods through four complementary studies. First, it constructs uCite, a unified PubMed-scale citation graph integrated from nine public sources, and shows that reliable graph construction requires explicit provenance, multi-source
comparison, and reliability filtering rather than naive source union. The thesis then studies taxonomy expansion, where new concepts must be inserted into existing hierarchies by combining textual semantics with structural constraints and geometry-aware modeling. It next investigates knowledge-augmented entity resolution, where external structured knowledge helps consolidate noisy, heterogeneous records only when retrieved evidence is selected and injected in a controlled way. Finally, it studies scientific claim verification through learned evidence graphs that connect claims to task-relevant sentences, showing that sparse, task-specific structure can improve both verification accuracy and evidence interpretability.
Taken together, these studies contribute a unified view of graph–text modeling in which structure is useful when it is explicit, selective, and task-relevant. The dissertation, therefore, contributes not only four application results but also a common perspective on how language models should interact with structure across data construction, constrained prediction, integration, and reasoning.
Questions? Contact Liri Fang.