TCGA: A Grid-Tagging NER Model Enhanced by Fusing Position and Region Information

The task of Named Entity Recognition (NER) involves the identification and categorization of entities in text through delineating their boundaries and assigning them to predefined categories. In recent years, grid tagging methods have shown a certain superiority in information extraction owing to the adaptable design of their model architecture. So, we adhere to the grid tagging method and propose the unified NER model, namely TCGA. Firstly, the unified NER task is modeled as a two-dimensional grid of words, and attention mechanism is applied to integrate position and region information. Secondly, the introduction of GRU and multi-scale convolutions aims to establish a finer-grained representation of words relationships, capturing internal dependencies and boundary information within phrases. The application of a collaborative predictor is ultimately utilized to effectively reason the intricate relationships between entities. We extensively conduct experiments on three widely recognized benchmark datasets, namely CONLL03, GENIA, and CADEC, where the TCGA improves the SOTA F1-Scores results by approximately 2.79%, 2.61%, and 0.81%, respectively, demonstrating its effectiveness.