Featured image of post Modularized Interaction Network for Named Entity Recognition

Modularized Interaction Network for Named Entity Recognition

现有NER模型缺点:

  • 基于序列标注的NER模型(The sequence labeling-based NERmodels)

    • 长实体识别不佳,只关注词级信息
  • 基于分段的NER模型(The segment-based NERmodels)

    • 处理分段,而非单个词,不能捕获分段中的词级依赖关系

边界检测(boundary detection)和类型预测(type pre-diction)可以相互配合,两个子任务可共享信息,相互加强。

提出模块化交互网络模型MIN(Modularized Interaction Network)

利用段级信息和词级依赖关系,结合一种交互机制,支持边界检测和类型预测之间的信息共享。

1

a recurrentneural network encoder-decoder framework with apointer network is used to detect entity segmentsfor segment information.

2