A Relational Triple Extraction Method Based on Cascading Pointer Annotation

An improved BF-CASRE model has been proposed for the problem of overlapping relational triples, which is currently not addressed by existing relational extraction models. This model leverages a function for relational modeling, facilitating the mapping of subjects to objects within sentences to enhance the understanding of relationships between entities. It encodes contextual information through a pre-trained Bert model. Next, subject and object taggers are implemented based on the pointer network for a specific relationship. Finally, the fusion self-attention mechanism is utilized to enhance the extraction of semantic features from the text by considering the importance of information at different positions within the text, thus improving overall effectiveness. In the feature fusion stage, a dense layer is added to fuse the CLS vector obtained by the Bert encoder and the subject information obtained by the subject tagger to improve the feature expression ability. Experimental validation of the BF-CASRE model was performed on the Chinese relational extraction dataset DuIE. The results showed that the improved BF-CASRE model effectively improves the performance of relational extraction.