Turkish Named Entity Recognition (NER) Model

This model is the fine-tuned model of dbmdz/convbert-base-turkish-cased (ConvBERTurk) using a reviewed version of well known Turkish NER dataset


The ConvBERT architecture is presented in the "ConvBERT: Improving BERT with Span-based Dynamic Convolution" paper.

Fine-tuning parameters:

task = "ner"
model_checkpoint = "dbmdz/convbert-base-turkish-cased"
batch_size = 8 
label_list = ['O', 'B-PER', 'I-PER', 'B-ORG', 'I-ORG', 'B-LOC', 'I-LOC']
max_length = 512 
learning_rate = 2e-5 
num_train_epochs = 3 
weight_decay = 0.01 

How to use:

model = AutoModelForTokenClassification.from_pretrained("akdeniz27/convbert-base-turkish-cased-ner")
tokenizer = AutoTokenizer.from_pretrained("akdeniz27/convbert-base-turkish-cased-ner")
ner = pipeline('ner', model=model, tokenizer=tokenizer, aggregation_strategy="first")
ner("<your text here>")

# Pls refer "https://huggingface.co/transformers/_modules/transformers/pipelines/token_classification.html" for entity grouping with aggregation_strategy parameter.

Reference test results:

  • accuracy: 0.9937648915431506
  • f1: 0.9610945644080416
  • precision: 0.9619899385131359
  • recall: 0.9602008554956295

Select AutoNLP in the “Train” menu to fine-tune this model automatically.

Downloads last month
Hosted inference API
Token Classification
This model can be loaded on the Inference API on-demand.