Universal Information Extraction as Unified Semantic Matching

The challenge of information extraction (IE) lies in the diversity label schemas and heterogeneity structures. Traditional methods require task-specific model design rely heavily on expensive supervision, making them difficult to generalize new schemas. In this paper, we decouple IE into two basic abilities, structuring conceptualizing, which are shared by different tasks Based paradigm, propose universally various with Unified Semantic Matching (USM) framework, introduces three unified token linking operations abilities conceptualizing. way, USM can jointly encode schema input text, uniformly extract substructures parallel, controllably decode target structures demand. Empirical evaluation 4 shows that proposed method achieves state-of-the-art performance under supervised experiments strong generalization ability zero/few-shot transfer settings.