Schema-aware Reference as Prompt Improves Data-Efficient Knowledge Graph Construction

• DOI: 10.48550/arxiv.2210.10709
• Publication Date: 2022-01-01
• Authors: Yunzhi Yao; Shengyu Mao; Xiang Chen; Ningyu Zhang; Shumin Deng; Huajun Chen
• Keywords: Knowledge graph; Schema (genetic algorithms); Leverage (statistics)

With the development of pre-trained language models, many prompt-based approaches to data-efficient knowledge graph construction have been proposed and achieved impressive performance. However, existing learning methods for are still susceptible several potential limitations: (i) semantic gap between natural output structured with pre-defined schema, which means model cannot fully exploit constrained templates; (ii) representation locally individual instances limits performance given insufficient features, unable unleash analogical capability models. Motivated by these observations, we propose a retrieval-augmented approach, retrieves schema-aware Reference As Prompt (RAP), construction. It can dynamically leverage schema inherited from human-annotated weak-supervised data as prompt each sample, is model-agnostic be plugged into widespread approaches. Experimental results demonstrate that previous integrated RAP achieve gains in low-resource settings on five datasets relational triple extraction event Code available https://github.com/zjunlp/RAP.