Classifying NESTED entities on Arabic data using StagedNER and Biaffine classifier inspired by Named Entity Recognition as Dependency Parsing paper

Data: Wojood corpus

• consists of 21 entity types
• Majority of data comes in MSA while the rest is in the Levant dialects (Palestinian and Lebanese)

Method:

• Staged NER trains two instances of the same transformer
• One transformer is trained on classifying BIO spans in the first stage and the second transformer is trained on classifying entity types in the second stage
• For NESTED ANER, we employ a biaffine classifer to classify BIO spans of NESTED entities while the second stage remains as is
• to reproduce stagedNER checkout our paper here

Demo: (coming soon)

Code in the repositroy provides implementation only for biaffine classifier and post-processing constraints described here