Unified Information Extraction Method for Power Operation Inspection Based on ERNIE

Unified Information Extraction (UIE) has emerged as a critical task in Natural Language Processing. By employing a unified modeling approach, it enhances both the accuracy and generalizability of information extraction while addressing the issues of resource wastage and insufficient knowledge sharing in single-task model. However, existing UIE methods struggle to adapt to the power sector due to the unique characteristics of its textual data, including ambiguous entity boundaries and complex semantics. To this end, a UIE framework tailored for the power industry has been designed, introducing the Prompt-based ERNIE and Pointer Network for Unified Information Extraction (PENP-UIE) model. PENP-UIE effectively tackles low-resource challenges through a prompt-based methodology, employing the ERNIE 3.0 pre-trained language model in the encoding layer to accommodate the specific features of power texts, while the decoding layer utilizes pointer networks to address entity nesting issues. To facilitate effective domain-specific finetuning, a dataset encompassing equipment defect maintenance has been constructed, which includes eight core entity types relevant to the power sector, thereby aligning closely with real-world applications. Comparative experiments across nine named entity recognition tasks, five relation extraction tasks, two event extraction tasks, and five UIE methods demonstrate that PENP-UIE consistently outperforms existing models. This dataset has been publicly released, significantly contributing to the structuralization of textual data in the power domain.