A Discontinuous Entity Recognition Model Based on Global Feature Interaction Mechanism

So far, discontinuous Named Entity Recognition (NER) has become a focus of attention for many researchers and has spawned numerous methodologies. Among them, the grid tagging model has particularly stood out due to its flexibility and adaptability in discontinuous NER tasks. The core idea of the model is to convert text into a two-dimensional grid format and capture entity segments as well as their relationships within the text by tagging elements in the grid. However, while grid tagging models typically emphasize learning tag embeddings to represent entity segment information and relationships between word pairs within the same contextual space, they often neglect the extraction of global features. We believe that designing two distinct feature generators to capture two different types of information, namely entity segment tag features and relationship tag features, during the learning process, is beneficial for improving the model’s performance. In this work, we propose novel tag feature generators, specifically designing two different generators–a relationship tag feature generator and an entity segment tag feature generator. These generators are designed to assist each other in more effectively generating feature information during the representation learning process. Experimental results demonstrate that our model achieves significant performance improvements on multiple discontinuous NER datasets, exhibiting higher accuracy and efficiency compared to other state-of-the-art (SOTA) methods. Furthermore, we conducted detailed analyses and discussions on the different components of the model, verifying the effectiveness of each component and its contribution to the overall model performance.