Awesome

<br/><br/>

Introduction

This repo contains all existing pretrained models that are fine-tuned for the Named Entity Recognition (NER) task. These models trained on a mixed NER dataset collected from ARMAN, PEYMA, and WikiANN that covered ten types of entities:

• Date (DAT)
• Event (EVE)
• Facility (FAC)
• Location (LOC)
• Money (MON)
• Organization (ORG)
• Percent (PCT)
• Person (PER)
• Product (PRO)
• Time (TIM)

Dataset Information

Download You can download the dataset from here

Evaluation

The following tables summarize the scores obtained by pretrained models overall and per each class.

Bert

Roberta

Distilbert

Albert

How To Use

You use this model with Transformers pipeline for NER.

Installing requirements

pip install sentencepiece
pip install transformers

How to predict using pipeline

from transformers import AutoTokenizer
from transformers import AutoModelForTokenClassification  # for pytorch
from transformers import TFAutoModelForTokenClassification  # for tensorflow
from transformers import pipeline

# model_name_or_path = "HooshvareLab/bert-fa-zwnj-base-ner"  # Roberta
# model_name_or_path = "HooshvareLab/roberta-fa-zwnj-base-ner"  # Roberta
model_name_or_path = "HooshvareLab/distilbert-fa-zwnj-base-ner"  # Distilbert
# model_name_or_path = "HooshvareLab/albert-fa-zwnj-base-v2-ner"  # Albert

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)

model = AutoModelForTokenClassification.from_pretrained(model_name_or_path)  # Pytorch
# model = TFAutoModelForTokenClassification.from_pretrained(model_name_or_path)  # Tensorflow

nlp = pipeline("ner", model=model, tokenizer=tokenizer)
example = "در سال ۲۰۱۳ درگذشت و آندرتیکر و کین برای او مراسم یادبود گرفتند."

ner_results = nlp(example)
print(ner_results)

Models

Hugging Face Model Hub

• Bert
• Roberta
• Distilbert
• Albert

Training

All models were trained on a single NVIDIA P100 GPU with following parameters.

Arguments

"task_name": "ner"
"model_name_or_path": model_name_or_path
"train_file": "/content/ner/train.csv"
"validation_file": "/content/ner/valid.csv"
"test_file": "/content/ner/test.csv"
"output_dir": output_dir
"cache_dir": "/content/cache"
"per_device_train_batch_size": 16
"per_device_eval_batch_size": 16
"use_fast_tokenizer": True
"num_train_epochs": 5.0
"do_train": True
"do_eval": True
"do_predict": True
"learning_rate": 2e-5
"evaluation_strategy": "steps"
"logging_steps": 1000
"save_steps": 1000
"save_total_limit": 2
"overwrite_output_dir": True
"fp16": True
"preprocessing_num_workers": 4

Cite

Please cite this repository in publications as the following:

@misc{ParsNER,
  author = {Hooshvare Team},
  title = {Pre-Trained NER models for Persian},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/hooshvare/parsner}},
}

Questions?

Post a Github issue on the ParsNER Issues repo.