Label-Centric Knowledge Retrieval for Zero-Shot Text Classification

Key points

• Flagship large language models (LLMs) offer rich general knowledge and strong reasoning capabilities, but their high inference cost and latency limit their use for large-scale data labeling.1
• We study zero-shot text classification with semantically meaningful labels and task descriptions, without task-specific labeled examples or classifier feedback.1
• Online, hybrid semantic and lexical retrieval first identifies the top-$K$ candidate labels and then selects up to $M$ input-relevant boundary rules connecting candidate-label pairs.1
• Boundary ablations further show that retrieving a small set of input-relevant rules preserves performance close to using all boundaries with substantially less context, supporting efficient reuse of flagship-generated label knowledge for lightweight classification.1

Sources (1)

• [1]Label-Centric Knowledge Retrieval for Zero-Shot Text Classification
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 05:00 AM

  We propose SLKR (Semantic Label Knowledge Retrieval), a label-centric retrieval-augmented generation framework that separates offline knowledge construction from lightweight online inference.

  Flagship large language models (LLMs) offer rich general knowledge and strong reasoning capabilities, but their high inference cost and latency limit their use for large-scale data labeling.

Extractive summary: sentences quoted from the sources.