原型网络直接应用于小样本命名实体识别(few-shot named entity recognition, FEW-NER)时存在以下问题: 非实体之间不具有较强的语义关系, 对实体和非实体都采用相同的方式构造原型将会造成非实体原型不能准确表示非实体的语义特征; 仅使用平均实体向量表示作为原型的计算方式将难以捕捉语义特征相差较大的同类实体. 针对上述问题, 提出基于细粒度原型网络的小样本命名实体识别(FEW-NER based on fine-grained prototypical networks, FNFP)方法, 有助于提高小样本命名实体识别的标注效果. 首先, 为不同的查询集样本构造不同的非实体原型, 捕捉句子中关键的非实体语义特征, 得到更为细粒度的原型, 提升模型对非实体的识别效果; 然后, 设计一个不一致性度量模块以衡量同类实体之间的不一致性, 对实体与非实体采用不同的度量函数, 从而减小同类样本之间的特征表示, 提升原型的特征表示能力; 最后, 引入维特比解码器捕捉标签转换关系, 优化最终的标注序列. 实验结果表明, 采用基于细粒度原型网络的小样本命名实体识别方法, 在大规模小样本命名实体识别数据集FEW-NERD上, 较基线方法获得提升; 同时在跨领域数据集上验证所提方法在不同领域场景下的泛化能力.
When prototypical networks are directly applied to few-shot named entity recognition (FEW-NER), there are the following problems: Non-entities do not have strong semantic relationships with each other, and using the same way to construct the prototype for both entities and non-entities will make non-entity prototypes fail to accurately represent the semantic characteristics of non-entities; using only the average entity vector as the computing method of the prototype will make it difficult to capture similar entities with different semantic features. To address these problems, this study proposes a FEW-NER based on fine-grained prototypical networks (FNFP) to improve the annotation effect of FEW-NER. Firstly, different non-entity prototypes are constructed for different query sets to capture the key semantic features of non-entities in sentences and obtain finer-grained prototypes to improve the recognition effect of non-entities. Then, an inconsistent metric module is designed to measure the inconsistency between similar entities, and different metric functions are applied to entities and non-entities, so as to reduce the feature representation between similar samples and improve the feature representation of the prototype. Finally, a Viterbi decoder is introduced to capture the label transformation relationship and optimize the final annotation sequence. The experimental results show that the performance of the proposed method is improved compared with that of the large-scale FEW-NER dataset, namely FEW-NERD; and the generalization ability of this method in different domain scenarios is verified on the cross-domain dataset.