Abstract: Few-shot Named Entity Recognition aims to identify entity spans and types in text based on a small amount of labeled
data. Recently, span-based metric learning has achieved promising results, but there are still two issues: first, prototypes may be pulled away from the center of the cluster due to a small number of 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, was proposed for few-shot NER. In the span detection stage, to learn explicit entity boundary information without being affected by label dependency at the token level, a global boundary matrix was used by DFSM to detect candidate spans; In the span classification stage, to increase the number of trainable candidate spans for each entity type, a fuzzy span strategy was proposed to adjust the boundary range of candidate spans. Meanwhile, prototypical contrastive learning was designed to optimize the span-based semantic representation space; Additionally, to eliminate the interference of non-entity noise data, prototypical margin learning was introduced to enlarge the distance between non-entity spans and prototypes. Experimental results on the Few-NERD and CrossNER datasets show that compared to the baseline model TadNER, the average F1 score of DFSM in Few-NERD Inter increases by 8.52%, especially in Inter 10 way 5~10 shot, where the F1 score increases by 10.39%, indicating its superior recognition ability for fine-grained entity types. Compared to the baseline model DecomMeta, the average F1 score of DFSM increases by 3.32% and 1.09% in CrossNER 1-shot and 5-shot, respectively, demonstrating its good generalization ability in cross-domain low-resource scenarios.

Key words: named entity recognition, 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实体跨度和类型。近期，基于跨度的度量学
习取得了不错的效果，但仍然存在以下个问题：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，而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，设计了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 中，DFSM 的平均 F1 值提升了 8.52%，尤其是在 Inter 10 way 5~10-shot 中，F1 值提升
了 10.39%，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
中，DFSM 的平均 F1 值分别提升了 3.32%和 1.09%，这表明它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