ADCL-MNER：基于自适应难度课程学习的多模态命名实体识别方法
ADCL-MNER: A Multimodal Named Entity Recognition Method Based on Adaptive Difficulty Curriculum Learning
胡慧云 1肖波1
作者信息
- 1. 北京邮电大学人工智能学院，北京 100876
- 折叠
摘要
多模态命名实体识别（MNER）通过融合文本与图像信息提升实体识别准确性，在社交媒体分析等场景中具有重要应用价值。然而，现有 MNER 方法面临跨模态信息融合不充分、无关视觉干扰以及静态训练策略忽视样本难度差异的问题，传统课程学习（CL）则存在难度评估主观化和调度策略固定化的局限。为此，提出一种基于自适应难度课程学习（ADCL）的 MNER 方法。该方法构建 "自动难度评估器 - 自适应课程调度器 - 数据采样器" 的闭环训练系统：通过损失值量化样本难度，动态跟踪模型学习状态调整训练课程，基于难度分数实现差异化采样。在 Twitter-2017 和 Twitter-2015 数据集上的实验表明，ADCL 方法的整体 F1 分数分别达到 93.58% 和 84.45%，相比基线模型提升 5.87~8.07 个百分点，在困难样本识别上性能优势显著。该方法为多模态命名实体识别提供了高效的训练策略解决方案。
Abstract
Multi-modal Named Entity Recognition (MNER) enhances entity recognition accuracy by integrating text and image information, and holds significant application value in scenarios such as social media analysis. However, existing MNER methods face challenges such as insufficient cross-modal information fusion, irrelevant visual interference, and static training strategies that neglect sample difficulty differences. Traditional Curriculum Learning (CL) suffers from limitations of subjective difficulty assessment and fixed scheduling strategies. To address these issues, we propose an MNER method based on Adaptive Difficulty Curriculum Learning (ADCL). This method constructs a closed-loop training system consisting of an "automatic difficulty evaluator - adaptive curriculum scheduler - data sampler": it quantifies sample difficulty through loss values, dynamically tracks the model\'s learning state to adjust the training curriculum, and implements differentiated sampling based on difficulty scores. Experiments on the Twitter-2017 and Twitter-2015 datasets show that the overall F1 score of the ADCL method reaches 93.58% and 84.45%, respectively, representing an improvement of 5.87~8.07 percentage points compared to the baseline model, with significant performance advantages in recognizing difficult samples. This method provides an efficient training strategy solution for multi-modal named entity recognition.关键词
多模态命名实体识别/自适应课程学习/自动难度评估/差异化采样Key words
Multimodal Named Entity Recognition/Adaptive Curriculum Learning/Automatic Difficulty Assessment/Differential Sampling引用本文复制引用
胡慧云,肖波.ADCL-MNER：基于自适应难度课程学习的多模态命名实体识别方法[EB/OL].(2026-01-26)[2026-01-28].http://www.paper.edu.cn/releasepaper/content/202601-56.学科分类
计算技术、计算机技术
评论