This model is a Named Entity Recognition (NER) model built using PyTorch and fine-tuned on the CoNLL-2003 dataset. The model is designed to identify and classify named entities in text into categories such as persons (PER), organizations (ORG), locations (LOC), and miscellaneous entities (MISC).

Intended Uses & Limitations

Intended Uses:

Text Analysis: This model can be used for extracting named entities from unstructured text data, which is useful in various NLP tasks such as information retrieval, content categorization, and automated summarization.
NER Task: Specifically designed for NER tasks in English.

Limitations:

Language Dependency: The model is trained on English data and may not perform well on texts in other languages.
Domain Specificity: Performance may degrade on text from domains significantly different from the training data.
Error Propagation: Incorrect predictions may propagate to downstream tasks, affecting overall performance.

How to Use

To use this model, you can load it using the Hugging Face Transformers library. Below is an example of how to perform inference using the model:

python

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("Ashaduzzaman/bert-finetuned-ner")
model = AutoModelForTokenClassification.from_pretrained("Ashaduzzaman/bert-finetuned-ner")

# Create a pipeline for NER
ner_pipeline = pipeline("ner", model=model, tokenizer=tokenizer)

# Example inference
text = "Hugging Face Inc. is based in New York City."
entities = ner_pipeline(text)

print(entities)

Troubleshooting

If the model isn't performing as expected, consider checking the following:

Ensure that the input text is in English, as the model was trained on English data.
Adjust the model's confidence threshold for entity detection to filter out less confident predictions.

Limitations and Bias

Bias in Data: The model is trained on the CoNLL-2003 dataset, which may contain biases related to the sources of the text. The model might underperform on entities not well represented in the training data.
Overfitting: The model may overfit to the specific entities present in the CoNLL-2003 dataset, affecting its generalization to new entities or text styles.

Training Data

The model was trained on the CoNLL-2003 dataset, a widely used benchmark dataset for NER tasks. The dataset contains annotated text from news articles, with labels for persons, organizations, locations, and miscellaneous entities.

Training Procedure

The model was fine-tuned using the pre-trained BERT model (bert-base-cased) with a token classification head for NER. The training process involved:

Optimizer: AdamW optimizer with betas=(0.9, 0.999) and epsilon=1e-08
Learning Rate: A linear learning rate scheduler was employed starting from 2e-05
Batch Size: 8 for both training and evaluation
Epochs: The model was trained for 3 epochs
Evaluation: Model performance was evaluated on a validation set with metrics like F1-score, precision, recall, and accuracy.

Training Hyperparameters

Learning Rate: 2e-05
Batch Size (train/eval): 8/8
Seed: 42
Optimizer: Adam with betas=(0.9, 0.999) and epsilon=1e-08
LR Scheduler Type: Linear
Number of Epochs: 3

Evaluation Results

This model was evaluated on the CoNLL-2003 test set, with performance measured using standard NER metrics:

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
0.076	1.0	1756	0.0657	0.9076	0.9337	0.9204	0.9819
0.0359	2.0	3512	0.0693	0.9265	0.9418	0.9341	0.9847
0.0222	3.0	5268	0.0599	0.9347	0.9512	0.9429	0.9864

Framework Versions

Transformers: 4.42.4
PyTorch: 2.3.1+cu121
Datasets: 2.21.0
Tokenizers: 0.19.1 !

ashaduzzaman/bert-finetuned-ner

bert-finetuned-ner

Model Description