Deberta for Named Entity Recognition
I used a Pretrained Deberta-v3-base and finetuned it on Few-NERD, A NER dataset that contains over 180k examples and over 4.6 million tokens.
The Token labels are Person, Organisation, Location, Building, Event, Product, Art & Misc.
How to use the model
from transformers import pipeline
def print_ner(sentences):
"""Cleaning and printing NER results
"""
for sentence in sentences:
last_entity_type = sentence[0]['entity']
last_index = sentence[0]['index']
word = sentence[0]['word']
for i, token in enumerate(sentence):
if (i > 0):
if (token['entity'] == last_entity_type) and (token['index'] == last_index + 1):
word = word + '' + token['word']
else:
word = word.replace('▁', ' ')
print(f"{word[1:]} {last_entity_type}")
word = token['word']
last_entity_type = token['entity']
last_index = token['index']
if i == len(sentence) - 1:
word = word.replace('▁', ' ')
print(f"{word[1:]} {last_entity_type}")
pipe = pipeline(model='RashidNLP/NER-Deberta')
sentence = pipe(["Elon Musk will be at SpaceX's Starbase facility in Boca Chica for the orbital launch of starship next month"])
print_ner(sentence)