Year: 2025
Type: conference-paper
Abstract: Named Entity Recognition (NER) is a key task in natural language processing. It aims to find entity boundaries and predefined entity types in text. For nested NER tasks, it is difficult to achieve both high accuracy and efficiency. Span-based NER methods are common deep learning approaches. They are effective for handling nested entities, but they face semantic fragmentation issues with long spans. Convolutional Neural Networks (CNN) can capture local context, but their fixed receptive fields and weights limit their ability to handle long-range dependencies. To solve this problem, we propose a new Adaptive Context-aware Attention Module (ACAM). This module adjusts sampling points and uses a multi-head separation strategy to extract local and long-range features from different directions. By focusing on important areas in span matrices, it improves entity representations and reduces semantic loss. We combined the ACAM with a contrastive learning-based NER model and tested it on the ACE 2004 and ACE 2005 datasets. The results show that it performs better than baseline methods, achieving F1 scores of 89.2% and 90.0% respectively. Ablation studies show that both adaptive sampling and multi-head separation strategies help capture complex patterns and improve span features.
Authors: Rui Zhang, Chenfan Ling, Qiaochuan Chen, Yuexing Han
Institution: Shanghai University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Awards: G1744688910, G4993317998, G8992341184
OpenAlex ID: W4413845197