[arXiv]score: 0.11
Knowledge Graph Augmentation Tested Across 15 LLMs for Zero-Shot Multi-Label Classification
June 1, 2026
A zero-shot multi-label topic classification framework with eight variants — including keyword enhancement, self-consistency decoding, and per-article knowledge graph augmentation via subject-predicate-object triples — is evaluated across 15 LLMs and eight datasets. Keyword-enhanced classification is the top-performing base method; KG augmentation effects vary by dataset and model.
cs.CL
HOW THIS AFFECTS YOU
●
builderIf you are building zero-shot document classification pipelines, the keyword-enhancement finding is a low-cost baseline worth testing before adding KG infrastructure.
●
researcherSystematic comparison across 15 LLMs and 8 datasets provides a reasonably broad empirical baseline for KG-augmented zero-shot classification.