Entity Relation Joint Extraction Method Based on Insertion Transformers

Existing multi-module multi-step and multi- module single-step methods for entity relation joint extraction suffer from issues such as cascading errors and redundant mistakes. In contrast, the single-module single-step modeling approach effectively alleviates these limitations. However, the single-module single-step method still faces challenges when dealing with complex relation extraction tasks, such as excessive negative samples and long decoding times. To address these issues, this paper proposes an entity relation joint extraction method based on Insertion Transformers, which adopts the single-module single-step approach and integrates the newly proposed tagging strategy. This method iteratively identifies and inserts tags in the text, and then effectively reduces decoding time and the count of negative samples by leveraging attention mechanisms combined with contextual information, while also resolving the problem of entity overlap. Compared to the state-of-the-art models on two public datasets, this method achieves high F1 scores of 93.2% and 91.5%, respectively, demonstrating its efficiency in resolving entity overlap issues.