Year: 2026
Type: dissertation
Abstract: Declaration of the Use of AI v List of Figures 4.1 KG Entity Alignment example in the Star Wars domain.Equivalent entities often appear under different names. . . . . . . . . . . . . . . .40 4.2 Cumulative percentage of gold-standard pairs retrieved at each rank position in SentenceBERT similarity rankings for the Memory Alpha-ST Expanded dataset. . . . . . . . . . . . . . . . . . . . . . . . . . . .42 4.3 Pipeline's dataflow.The brackets indicate which property was used. . .45 5.1 Overview of the proposed sparse entity alignment pipeline.For each anchor entity in the source KG, we: (1) generate textual representations (Dogtags) incorporating node attributes and graph structure, (2) retrieve top-k candidates using embedding similarity, (3) select matches via exact matching, embedding-based reranking, or LLM prompting, and (4) deduplicate to enforce 1:1 alignments.Multiple strategies can be merged using threshold calibration from exact matches. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .56 5.2 Comparison of entity representation strategies (Dogtags): Node-based representations use limited context from attributes and abstracts, while the LLM Summary incorporates rich relational and attribute information from the knowledge graph neighborhood.All four representations serve as input to embedding-based retrieval and matching algorithms. . . . . . . . . . . . . . .
Author: Kardos Péter
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204678376