SafeEdgeNER

Named Entity Recognition for Security and Defense

📌 Description

SafeEdgeNER is an academic Named Entity Recognition (NER) project applied to the security and defense domain. The goal is to automatically identify relevant entities such as military platforms, weapons, locations, temporal references, and official documents in large amounts of unstructured textual data.

The system uses advanced Deep Learning models, including BERT and RoBERTa, to ensure high accuracy in entity extraction.

🚀 Technologies Used

• Python
• Hugging Face Transformers
• PyTorch
• Scikit-learn
• Pandas & NumPy
• Matplotlib & Seaborn (for result analysis)

📂 Project Structure

SafeEdgeNER/               
│── documents/                
│── notebooks/             # Jupyter notebooks
│   │── EDA.ipynb               
│   │── NER_BERT.ipynb
│   │── NER_CRF.ipynb
│   │── NER_RoBERTa_test1.ipynb
│   │── NER_RoBERTa_test2.ipynb
│   │── NER_RoBERTa_test3.ipynb
│   │── NER_RoBERTa_test4.ipynb
│   │── NER_RoBERTa_test5.ipynb
│   │── NER_RoBERTa_test6.ipynb
│   │── NER_RoBERTa_test7.ipynb
│── README.md             
│── LICENSE                # Project license

📊 Dataset

The project uses a dataset adapted from the re3d corpus, which includes documents related to conflicts in Syria and Iraq. The dataset has been preprocessed and converted into the CoNLL 2003 IOB2 format for proper entity annotation.

The main recognized entities are:

• Location (e.g., cities, geographic coordinates)
• MilitaryPlatform (e.g., ships, tanks, aircraft)
• Weapon (e.g., AK-47, missiles)
• Temporal (e.g., dates, times)
• DocumentReference (e.g., official document identifiers)
• Person (names of individuals)
• Organisation (e.g., governments, military groups)
• Money (currencies, amounts of money)

🔧 Usage

you can run the notebooks on Google Colab by uploading them to a Colab workspace and running the cells.

1️⃣ Exploratory Data Analysis (EDA)

To analyze the dataset, use:

jupyter notebook notebooks/EDA.ipynb

2️⃣ Train Models

Various models were tested, including BERT, CRF, and RoBERTa. To run a specific training notebook, use:

jupyter notebook notebooks/NER_BERT.ipynb   # For BERT model training
jupyter notebook notebooks/NER_CRF.ipynb    # For CRF model training
jupyter notebook notebooks/NER_RoBERTa_test5.ipynb  # For RoBERTa model training 

Additional RoBERTa training attempts are available in the other numbered notebooks.

📈 Results

Tests on models showed that RoBERTa achieved the best performance in entity recognition, with an average F1-score of 0.88 and an accuracy of 89%.

Optimizations

Several optimization techniques were applied to improve performance of our RoBERTa model, including:
✔️ Data Augmentation (Synonym Augmentation, Contextual Embedding Augmentation)
✔️ Class Weights to handle class imbalance
✔️ Focal Loss to reduce the impact of more frequent classes
✔️ Increased number of epochs to enhance training

👥 Contributors

The project was developed by:

• Micol Zazzarini
• Andrea Fiorani
• Antonio Antonini

📜 License

This project is distributed under the MIT license. See the LICENSE file for more details.

---