fewshot_NER:SpanProto: A Two-stage Span-based Prototypical Network for Few-shot Named Entity Recogni

Introduction

我们提出了一个开创性的基于跨度的原型网络(SpanProto),它通过一个两阶段的方法来解决少量的NER问题,包括跨度提取和提及分类。在跨度提取阶段,我们将顺序标签转化为全局边界矩阵,使模型能够专注于明确的边界信息。对于提及分类,我们利用原型学习来捕捉每个标记的跨度的语义表示,并使模型更好地适应小说类实体。

模型分为两个部分,通过表填充方式解决mention识别问题,使用原型网络解决mention分类问题。

Model

fewshot_NER:SpanProto: A Two-stage Span-based Prototypical Network for Few-shot Named Entity Recogni_第1张图片
第一部分:mention识别是通过pointer network解决的。损失函数时二分类交叉熵损失函数
fewshot_NER:SpanProto: A Two-stage Span-based Prototypical Network for Few-shot Named Entity Recogni_第2张图片

第二部分:原型学习实现关系分类。在每一个episode中,通过平均化相同实体类型的span的表示得到对应class的原型。损失函数是分类损失函数

当识别出的flase positive类型,则将the false positive can be viewed as a special entity mention, which has no type to be assigned in Ttrain, but could be an entity in other
episode data
. In other words, the real type of this
false positive is unknown. Thus, a natural idea is
that we can keep it away from all current prototypes
in the semantic space. S

fewshot_NER:SpanProto: A Two-stage Span-based Prototypical Network for Few-shot Named Entity Recogni_第3张图片

你可能感兴趣的:(EMNLP,人工智能)