与知识相结合的提示学习研究综述

Summary

This Chinese-language survey from the National University of Defense Technology examines how prompt learning and knowledge can work together. It covers two directions: using prompts to extract knowledge, and using knowledge to build better prompts. Prompt learning recasts a downstream task in the format a pre-trained language model already handles, such as masked-token filling or generation. This avoids the cost and objective mismatch of full fine-tuning.

The authors first set out the core vocabulary, mostly following Liu et al.:

• template (prompt) engineering,
• answer engineering,
• multi-prompt strategies.

They then organise the literature into a taxonomy covering entity, relation and event extraction. A further category covers methods that inject external knowledge into prompt construction and label mapping. For each family of methods they note typical weaknesses, for example:

• error propagation when prompts are decomposed,
• the expense of hand-crafted templates,
• how to choose and order demonstrations.

The survey reports no new experiments. Its value is a structured map of an emerging area, especially for low-resource and few-shot extraction, together with a discussion of open limitations and future trends.

Key points

• Venue and pipeline: Published in 中文信息学报 (Journal of Chinese Information Processing), vol. 37, no. 7, 2023, pp. 1–12. It frames prompt learning as three steps: template construction, answer search, and label-word mapping. Answers can be tokens, spans or sentences.
• Entity extraction is grouped into three families:
  • prompt decomposition: PLET and TemplateNER, where TemplateNER scores enumerated spans of up to 8 tokens with BART;
  • prefix prompts: plug-tagger and LightNER, which freeze the backbone and train inserted vectors;
  • demonstration-based methods: Lee et al. and EntLM.
• Demonstrations (Lee et al.): Entity-type-oriented demonstrations beat instance-oriented ones, which are retrieved with SBERT or BERTScore. Among the three ways of selecting examples (random, most frequent, search), the search strategy worked best.
• Relation extraction methods covered include:
  • typed entity markers (Zhou & Chen);
  • manual templates: AdaPrompt and RelationPrompt, the latter generating synthetic data for zero-shot relation extraction;
  • continuous templates with continuous label mapping: DART.
• Event extraction and knowledge constraints:
  • Event extraction largely follows the ACE2005 conventions. The surveyed methods include L2A for argument extraction, generation-based pipeline approaches, and prompt fusion (Lin et al.).
  • Knowledge-constrained prompting includes PTR and work by Hu, Chen and Ye et al. These methods use knowledge to shrink the search space.
  • Selective knowledge injection relies on a visible matrix to limit noise and heterogeneity.