POS-Signature Retrieval Outperforms Semantic Search for In-Context Dependency Parsing

Replacing semantic similarity with a POS-signature policy for demonstration retrieval improves Universal Dependencies parsing accuracy. Using weighted part-of-speech statistics to select length-compatible examples, this method achieved labeled attachment scores of up to 73.2 on open-weight models, outperforming semantic retrieval by up to 2.7 points.