Abstract:
Relation extraction is a key task in text data mining, where its technical goal is to mine a triples consisting of entities and semantic relations between them. The problem of overlapping relations is the current difficulty in extraction relation, where hierarchical cascade tagging has a remarkable performance in solving this problem. But there are missing features in existing methods of this strategy because only BERT (Bidirectional Encoder Representations from Transformers) is used as the feature input for the subject tagging module on the one hand, and no relevant feature mining is done for the identified subjects on the other hand. In response, this paper proposes a dynamic hierarchical cascade tagging model for overlapping relation extraction. Firstly, a dynamic character-word fusion feature learning model (RWG) is constructed based on RoBERTa-wwm-ext (A Robustly Optimized BERT Pre-training Approach-whole word masking-extended), WoBERT (Word BERT) and gated mechanism. This effectively avoids the problems of missing features in the subject tagging module and inability to compute in parallel; secondly, the introduction of a dynamic weight local self-attention (LSA) autonomously learns subject-level semantic information; finally, based on an effective fusion of global and subject local features of the input text, the RWG-LSA model is implemented for the extraction of entity pairs and relations in the text. The experiments on the SKE Chinese dataset show that the model is significantly effective in extracting overlapping relations, with the F1 value reaching 82.44%.