Demonstration-Based and Attention-Enhanced Grid-Tagging Network for Mention Recognition

Abstract

Concepts empower cognitive intelligence. Extracting flat, nested, and discontinuous name entities and concept mentions from natural language texts is significant for downstream tasks such as concept knowledge graphs. Among the algorithms that uniformly detect these types of name entities and concepts, Li et al. proposed a novel architecture by modeling the unified mention recognition as the classification of word–word relations, named W2NER, achieved state-of-the-art (SOTA) results in 2022. However, there is still room for improvement. This paper presents three improvements based on W2NER. We enhanced the grid-tagging network by demonstration learning and tag attention feature extraction, so our modified model is named DTaE. Firstly, addressing the issue of insufficient semantic information in short texts and the lack of annotated data, and inspired by the demonstration learning from GPT-3, a demonstration is searched during the training phase according to a certain strategy to enhance the input features and improve the model’s ability for few-shot learning. Secondly, to tackle the problem of W2NER’s subpar recognition accuracy problem for discontinuous entities and concepts, a multi-head attention mechanism is employed to capture attention scores for different positions based on grid tagging. Then, the tagging attention features are embedded into the model. Finally, to retain information about the sequence position, rotary position embedding is introduced to ensure robustness. We selected an authoritative Chinese dictionary and adopted a five-person annotation method to annotate multiple types of entities and concepts in the definitions. To validate the effectiveness of our enhanced model, experiments were conducted on the public dataset CADEC and our annotated Chinese dictionary dataset: on the CADEC dataset, with a slight decrease in recall rate, precision is improved by 2.78%, and the comprehensive metric F1 is increased by 0.89%; on the Chinese dictionary dataset, the precision is improved by 2.97%, the recall rate is increased by 2.35%, and the comprehensive metric F1 is improved by 2.66%.