An overlapping entity relation extraction method for electricity marketing audit information

In the field of electricity marketing inspection, large amounts of unstructured text data (such as inspection reports, work orders, and supporting documentation) present significant challenges for traditional information extraction methods. Traditional pipeline-based entity relation extraction methods and joint extraction methods are difficult to solve complex problems such as overlapping relations and nested entities. To tackle these challenges, this paper proposes an entity-relationship extraction technique for electricity marketing based on a cascaded pointer network. First, in order to improve the accuracy of entity recognition, a domain-specific PowerSpanBERT model was trained by incorporating both domain text and the specialized lexicon, and adversarial training with gradient accumulation was introduced to enhance the model’s accuracy in extracting electricity marketing knowledge and predicting unfamiliar information. In the joint recognition layer for relationships and tail entities, a multi-head attention mechanism was introduced to strengthen the feature rep-resentation of the output vectors from the encoding layer, enabling joint extraction of entities and relationships. Finally, experimental results show that the pro-posed joint extraction method significantly improves knowledge extraction accuracy, achieving an F1 score of $86.04 \%$, making it a powerful tool to support advanced applications such as intelligent information retrieval and decisionmaking in marketing inspections.