BERT NER Organization Name Extraction

Overview

Fine-tune of the bert-base-uncased model, trained to identify and classify named entities within affiliation strings, focusing on organizations and locations.

Training Data

Training data comprised approximately 500,000 programatically annotated items, where named entities in affiliation strings were tagged with their respective types (organizations, locations), and all other text is marked as extraneous. Example annotation format:

O: Internal
O: Medicine
O: Complex
B-ORG: College
I-ORG: of
I-ORG: Medical
I-ORG: Sciences
B-LOC: New
I-LOC: Delhi
B-LOC: India

The training data was derived from OpenAlex affiliation strings and their ROR ID assignments. Tagging was done using the corresponding name and location metadata from the assigned ROR record. Location names were further supplemented with aliases derived from the Unicode Common Locale Data Repository (CLDR).

Training Details

Dataset Size: ~500,000 items
Number of Epochs: 3
Optimizer: AdamW
Training Environment: Google Colab T-4 High Ram instance
Training Duration: Approximately 8 hours

Usage

See https://github.com/ror-community/affiliation-matching-experimental/tree/main/ner_tests/inference for example usage.