Name Entity Recognition (NER) on FiNER Dataset

Dataset

• A sequence is a single sentence
• Every word is given an NER label
• NER tags: Person, Organization, Location, non NER tag
• 6 labels
  • {'O': 0, 'PER_B': 1, 'PER_I': 2, 'LOC_B': 3, 'LOC_I': 4, 'ORG_B': 5, 'ORG_I': 6}
  • O is not an NER tag
  • Multi token entities are split into separate tokens
  • Ex: "United States of America is a country." -> [LOC_B] [LOC_I] [LOC_I] [LOC_I] [O] [O] [O].
  • so every NER tag has a ..._B (beginning) and ..._I (intermediate)

Results

• NER seen as a classification task
  • NLL Loss or CE loss
• Underlying problem is determining which tokens in a multi token entity is the beginning and is the last intermediate (requires understanding of context and long spatial awareness) -> BERT appropriate (Transformer and allows fine tuning for classifcation task like NER)
• Fine Tuned Hugging Face BERT-base model in 5 epochs and achieved 0.96 F1 score on test set
• Documentation and understanding can be found here