Thema der Dissertation:
Continual Relation Extraction for Knowledge Graph Completion Semantic Continuity in Large Language Models Thema der Disputation:
Semantic Continuity in Dynamic Knowledge Graph Generation
Continual Relation Extraction for Knowledge Graph Completion Semantic Continuity in Large Language Models Thema der Disputation:
Semantic Continuity in Dynamic Knowledge Graph Generation
Abstract: Semantic continuity has been studied in various machine learning settings, including classification, segmentation, trajectory planning, and conversational agentic systems. Broadly, semantic discontinuity, namely, the loss of semantic information, can be understood as a loss of semantic consistency, for example through interclass misclassification or intraclass discontinuity. With the increasing use of large language models (LLMs) as parametric memory for structured knowledge acquisition, an important question arises: When does a change in an LLM’s prediction constitute information loss, and when does generating a semantically related but different relation preserve the semantic continuity of previously learned knowledge?
This talk presents semantic continuity in the context of dynamic knowledge graph generation, focusing on continual learning with LLMs and the relationship between catastrophic forgetting and knowledge graph completeness when continual relation extraction is utilized. It discusses how previously acquired knowledge about relations can be preserved while new relations are continuously learned and how semantically valid novel relations can be distinguished from hallucinations and semantic drift. Finally, the talk introduces semantic continuity as a framework for understanding knowledge preservation and evolution in dynamic knowledge graph generation.
This talk presents semantic continuity in the context of dynamic knowledge graph generation, focusing on continual learning with LLMs and the relationship between catastrophic forgetting and knowledge graph completeness when continual relation extraction is utilized. It discusses how previously acquired knowledge about relations can be preserved while new relations are continuously learned and how semantically valid novel relations can be distinguished from hallucinations and semantic drift. Finally, the talk introduces semantic continuity as a framework for understanding knowledge preservation and evolution in dynamic knowledge graph generation.
Zeit & Ort
02.10.2026 | 13:30
Seminarraum 005
(Fachbereich Mathematik und Informatik, Takustr, 9, 14195 Berlin)
&
WebEx