README.md

Neural-based Propaganda Detection

This repository implements a neural network-based classifier to detect propaganda in textual data. The model is trained using GloVe word embeddings and a Multi-Layer Perceptron (MLP) architecture. It predicts whether a given sentence belongs to the class of "propaganda" or "non-propaganda".

---

Features

• Pre-trained GloVe embeddings for semantic word representations.
• Robust neural architecture using PyTorch.
• Pickle-based model serialization for easy inference.
• Simple and efficient sentence vectorization using GloVe.
• CLI-based inference for predicting individual sentences.

---

Technologies and Tools

1. Python Libraries

• PyTorch: Deep learning framework used for building and training the MLP classifier.
• Gensim: For loading and utilizing pre-trained GloVe word embeddings.
• NLTK: Tokenizer for breaking sentences into words.
• NumPy: Efficient numerical computations for vector operations.
• Scikit-learn: For evaluating the model (precision, recall, F1-score).
• Pandas: Data manipulation during preprocessing.

2. Pre-trained Embeddings

• GloVe (Global Vectors for Word Representation):
  • Version: glove.6B.300d.txt.
  • Provides 300-dimensional dense vector representations for English words.

3. Dataset

• Textual data provided in train.tsv.
  • Contains sentences, article titles, and their corresponding labels (propaganda or non-propaganda).

4. File Serialization

• Pickle:
  • Saves the trained model weights, input parameters, and OOV vector for efficient inference.

---

Setup Instructions

1. Clone the Repository

git clone https://github.com/ash-sha/propaganda-detection.git
cd propaganda-detection

2. Install Dependencies

Ensure Python 3.8+ is installed. Then install required libraries:

pip install -r requirements.txt

3. Download Pre-trained GloVe Embeddings

• Download glove.6B.300d.txt from the GloVe website.
• Place the file in the appropriate directory (e.g., ./glove.6B.300d.txt).

4. Training

(Optional) To retrain the model from scratch:

train.ipynb

5. Inference

• Ensure the model file (propoganda.pickle) is present in the working directory.
• Run the inference script:

test.ipynb

• Enter your sentence when prompted, and get the prediction:

Enter the query: The government is spreading fake news to mislead the public.
Predicted label: propaganda

---

Repository Structure

propaganda-detection/
│
├── glove.6B.300d.txt   # Pre-trained GloVe embeddings
├── train.tsv   # Dataset used for training
├── propoganda.pickle   # Serialized trained model
├── train.ipynb                # Script for training the model
├── test.ipynb            # Script for running inference
├── requirements.txt        # Python dependencies
├── README.md               # Project documentation

---

Model Workflow

1. Data Preprocessing:

   • Load sentences and labels from the dataset.
   • Shuffle and split data into training, validation, and testing sets.
   • Normalize and vectorize sentences using GloVe embeddings.

2. Model Architecture:

   • Multi-Layer Perceptron (MLP) with one hidden layer.
   • Dropout regularization to prevent overfitting.

3. Training:

   • Optimize using the Adam optimizer and CrossEntropy loss.
   • Evaluate on the validation set after each epoch.
   • Save the model with the highest validation F1-score.

4. Inference:

   • Load the trained model (propoganda.pickle).
   • Use GloVe embeddings to vectorize the input query.
   • Predict the class label for the input sentence.

---

Sample Results

• Input Query: The government is spreading fake news to mislead the public.
• Predicted Label: propaganda

---

Future Work

• Add support for multi-class propaganda classification (e.g., detecting specific propaganda techniques).
• Improve sentence vectorization by integrating contextual embeddings (e.g., BERT, RoBERTa).
• Implement a web or mobile interface for real-time predictions.

---

Contributing

Contributions are welcome! Please fork the repository and submit a pull request.

---

License

This project is licensed under the MIT License. See LICENSE for details.