NER Model using Roberta
This markdown presents a Robustly Optimized BERT Pretraining Approach (RoBERTa) model trained on a combination of two diverse datasets for two languages: English and Persian. The English dataset used is CoNLL 2003, while the Persian dataset is PEYMA-ARMAN-Mixed, a fusion of the "PEYAM" and "ARMAN" datasets, both popular for Named Entity Recognition (NER) tasks.
The model training pipeline involves the following steps:
Data Preparation: Cleaning, aligning, and mixing data from the two datasets. Data Loading: Loading the prepared data for subsequent processing. Tokenization: Utilizing tokenization to prepare the text data for model input. Token Splitting: Handling token splitting (e.g., "jack" becomes "_ja _ck") and using "-100" for optimization and ignoring certain tokens. Model Reconstruction: Adapting the RoBERTa model for token classification in NER tasks. Model Training: Training the reconstructed model on the combined dataset and evaluating its performance. The model's performance, as shown in the table below, demonstrates promising results:
| Epoch | Training Loss | Validation Loss | F1 | Recall | Precision | Accuracy |
|---|---|---|---|---|---|---|
| 1 | 0.072600 | 0.038918 | 89.5% | 0.906680 | 0.883703 | 0.987799 |
| 2 | 0.027600 | 0.030184 | 92.3% | 0.933840 | 0.915573 | 0.991334 |
| 3 | 0.013500 | 0.030962 | 94% | 0.946840 | 0.933740 | 0.992702 |
| 4 | 0.006600 | 0.029897 | 94.8% | 0.955207 | 0.941990 | 0.993574 |
The model achieves an impressive F1-score of almost 95%.
To use the model, the following Python code snippet can be employed:
from transformers import AutoConfig, AutoTokenizer, AutoModel
config = AutoConfig.from_pretrained("AliFartout/Roberta-fa-en-ner")
tokenizer = AutoTokenizer.from_pretrained("AliFartout/Roberta-fa-en-ner")
model = AutoModel.from_pretrained("AliFartout/Roberta-fa-en-ner")
By following this approach, you can seamlessly access and incorporate the trained multilingual NER model into various Natural Language Processing tasks.