A Joint Model for Hierarchical Nested Information Extraction

Abstract

During the long-term power construction process, the power dispatching department has saved many notification texts related to adjustment of grid operation mode. There is an urgent need to study named entity recognition techniques to automatically recognize the power equipment and operation mode, in order to support automatic verification of grid operation mode. By analyzing the characteristics of notification texts, a classification method of hierarchical nested named entities is proposed for the first time in power domain. The entities are divided into two layers with nested relationships, and the corpus of grid operation mode is constructed. We further propose a joint model based on character-word feature fusion and attention mechanism. The model is based on the parameter sharing approach for joint recognition of hierarchical nested entities in the corpus and further introduces an attention mechanism to optimize the feature interaction between hierarchical nested entities. In addition, we splice embeddings of characters and words as feature input to obtain richer semantic features. Experimental results show that our model achieves state-of-the-art results. Eventually, the recognition results can be stored as a standardized verification information chain, providing effective data support for automatic verification of the grid operation mode and ensuring safe and stable operation of the grid.

Keywords

• Hierarchical nested named entity
• joint model
• attention mechanism
• feature fusion
• named entity recognition