PEEL
A lightweight probing method that decodes ontology structure from frozen retrieval embeddings using an order-embedding objective. It requires only a concept vocabulary and no document corpus, and it can transfer zero-shot to new domains.
Timeline (last 90 days)
PEEL is a lightweight probing method that operates on frozen embeddings, decoding a classification DAG from a concept vocabulary via an order-embedding objective and a minimum spanning arborescence algorithm.
PEEL requires only a concept vocabulary and can run without any document corpus.
On two OLLM benchmarks, the classification DAGs recovered by PEEL exhibit high structural fidelity.
PEEL can transfer zero-shot to new domains and maintain performance without edge supervision in the target domain.
PEEL recovers the complete taxonomy DAG on a directed graph via a minimum spanning arborescence algorithm, without presupposing a fixed depth or number of branches.
PEEL substantially lowers the deployment barrier while achieving structural fidelity close to that of the end-to-end OLLM pipeline, avoiding expensive LLM training and document processing.
PEEL's decoding structure originates from the granularity information inherent in MRL embeddings rather than from confounding factors, which was confirmed by control experiments that shuffle and reverse the dimensions.
All Facts (7)
PEEL is a lightweight probing method that operates on frozen embeddings, decoding a classification DAG from a concept vocabulary via an order-embedding objective and a minimum spanning arborescence algorithm.
50%UnverifiedPEEL requires only a concept vocabulary and can run without any document corpus.
50%UnverifiedOn two OLLM benchmarks, the classification DAGs recovered by PEEL exhibit high structural fidelity.
50%UnverifiedPEEL can transfer zero-shot to new domains and maintain performance without edge supervision in the target domain.
50%UnverifiedPEEL recovers the complete taxonomy DAG on a directed graph via a minimum spanning arborescence algorithm, without presupposing a fixed depth or number of branches.
50%UnverifiedPEEL substantially lowers the deployment barrier while achieving structural fidelity close to that of the end-to-end OLLM pipeline, avoiding expensive LLM training and document processing.
50%UnverifiedPEEL's decoding structure originates from the granularity information inherent in MRL embeddings rather than from confounding factors, which was confirmed by control experiments that shuffle and reverse the dimensions.
50%