metadata

language: bn
tags:
  - bengali-ner
  - bengali
  - bangla
  - NER
license: MIT
datasets:
  - wikiann
  - xtreme

Multi-lingual BERT Bengali Name Entity Recognition

mBERT-Bengali-NER is a transformer-based Bengali NER model build with bert-base-multilingual-uncased model and Wikiann Datasets.

How to Use

pip install bntransformer

from bntransformer import BanglaNER

# passing "sagorsarker/mbert-bengali-ner" in bntransformer
bnner = BanglaNER("sagorsarker/mbert-bengali-ner")
sentence = "আমি জাহিদ হাসান এবং আমি ঢাকায় বাস করি ।"
output = bnner.ner_tag(sentence)
print(output)

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("sagorsarker/mbert-bengali-ner")
model = AutoModelForTokenClassification.from_pretrained("sagorsarker/mbert-bengali-ner")

nlp = pipeline("ner", model=model, tokenizer=tokenizer)
example = "আমি জাহিদ এবং আমি ঢাকায় বাস করি।"

ner_results = nlp(example)
print(ner_results)

Label and ID Mapping

Label ID	Label
0	O
1	B-PER
2	I-PER
3	B-ORG
4	I-ORG
5	B-LOC
6	I-LOC

Training Details

mBERT-Bengali-NER trained with Wikiann datasets
mBERT-Bengali-NER trained with transformers-token-classification script
mBERT-Bengali-NER total trained 5 epochs.
Trained in Kaggle GPU

Evaluation Results

Model	F1	Precision	Recall	Accuracy	Loss
mBert-Bengali-NER	0.97105	0.96769	0.97443	0.97682	0.12511