Abstract:

Few-shot Named Entity Recognition （few-shot NER） aims to identify entity spans and their types in text based on limited labeled data. Although span-based metric learning has achieved promising results in recent years， two challenges remain： first， prototypes may be pulled away from cluster centers due to sparse candidate spans； second， some non-entity spans may be produced by span detectors that are irrelevant to the categories. To address these issues， a decomposed model integrating fuzzy span， namely DFSM （Decomposed Fuzzy Span Model）， was proposed for few-shot NER. In the span detection stage， a global boundary matrix was used to detect candidate spans， enabling the learning of explicit entity boundary information without dependency on labels at token level. In the span classification stage， a fuzzy span strategy was proposed to adjust the boundary ranges of candidate spans， thereby increasing the number of trainable candidate spans for each entity type. Meanwhile， a prototypical contrastive learning was designed to optimize the span-based semantic representation space. Besides， prototypical boundary learning was introduced to enlarge the distance between non-entity spans and prototypes， eliminating interference from non-entity noisy data. Experimental results on Few-NERD and CrossNER datasets show that： compared to the baseline model TadNER， DFSM achieves an average F1-score gain of 8.52 percentage points under the Few-NERD Inter setting， with a notable 10.39 percentage points improvement in the Inter 10-way 5 - 10-shot scenario， highlighting its enhanced capability for fine-grained entity recognition； compared to the baseline model DecomMeta， DFSM achieves F1-score improvements of 3.32 and 1.09 percentage points in CrossNER 1-shot and CrossNER 5-shot setting， respectively， demonstrating the good generalization ability of DFSM in cross-domain low-resource scenarios.

Key words: Named Entity Recognition (NER), few-shot learning, prototypical network, global boundary matrix, fuzzy span

摘要：

小样本命名实体识别（few-shot NER）旨在基于少量标记数据识别文本中的实体跨度和类型。近年来，基于跨度的度量学习虽然取得了不错的效果，但仍然存在2个问题：一是少量的候选跨度可能导致原型偏离群组的中心；二是与类别无关的跨度检测器可能会产生一些非实体跨度。为了解决以上问题，提出一种用于few-shot NER的融合模糊跨度的分解式模型DFSM（Decomposed Fuzzy Span Model）。在跨度检测阶段，为学习明确的实体边界信息且不受标记级别的标签依赖影响，DFSM采用全局边界矩阵检测候选跨度；而在跨度分类阶段，为增加可训练的每种实体类型的候选跨度数量，提出一种模糊跨度策略，以调整候选跨度的边界范围。同时，设计一种原型对比学习以优化基于跨度的语义表示空间。此外，为消除非实体噪声数据的干扰，引入原型边界学习以扩大非实体跨度与原型的距离。在Few-NERD和CrossNER数据集上的实验结果显示：与基线模型TadNER相比，在Few-NERD Inter设置中，DFSM的平均F1值提升了8.52个百分点，尤其是在Inter 10 way 5~10-shot设置中，DFSM的平均F1值提升了10.39个百分点，这表明DFSM对于细粒度实体类型具有更强的识别能力；与基线模型DecomMeta相比，在CrossNER 1-shot和5-shot设置中，DFSM的平均F1值分别提升了3.32和1.09个百分点，这表明DFSM在跨领域低资源场景下具有良好的泛化能力。

关键词: 命名实体识别, 小样本学习, 原型网络, 全局边界矩阵, 模糊跨度