🤖 Sarcasm Detection Using Tokenization and Neural Networks

This project builds a sarcasm detection system using a neural network trained on real news headlines. It processes text by turning words into numbers and feeds them into a model that improves itself through learning.

---

📊 Dataset

• Source: Kaggle - News Headlines Dataset for Sarcasm Detection
• Format: JSON (each line = one JSON object)
• Fields:
  • headline: the text of the news headline
  • is_sarcastic: label (1 for sarcastic, 0 for not sarcastic)

---

🧪 How It Works (Step-by-Step)

1. Load and Prepare Data

• Headlines and sarcasm labels are loaded from a JSON file.
• First 20,000 examples are used for training, and the rest for testing.

2. Tokenization: Text → Numbers

• A Keras Tokenizer is created with a vocabulary size limit of 10,000.
• Words in the training data are mapped to unique integers (tokens).
• Words not in the vocabulary are replaced with <OOV> (out-of-vocabulary).

3. Padding Sequences

• Headline sequences are converted to the same length (100) using post-padding and post-truncation.
• This makes the data uniform in size, which is required for input to the neural network.

4. Convert to Numpy Arrays

• Both padded sequences and labels are converted into NumPy arrays.
• This step is necessary because TensorFlow models require numerical array input.

5. Build Neural Network Model

The model is created using tf.keras.Sequential() with the following layers:

• Embedding: Converts each word index into a dense 16-dimensional vector.
• GlobalAveragePooling1D: Reduces the sequence into a single feature vector.
• Dense (24, relu): Learns hidden patterns.
• Dense (1, sigmoid): Outputs a probability score between 0 and 1 for sarcasm.

6. Train the Model

• The model is compiled using binary cross-entropy loss and the Adam optimizer.
• It is trained for 30 epochs.
• During training, the model evaluates itself on the test set after each epoch to improve performance.

7. Evaluate and Test

• After training, the model's accuracy is measured on the testing dataset.
• You can input new headlines to see the model predict sarcasm probability.

8. Save the Model and Tokenizer

• The trained model is saved as sarcasm_model.keras.
• The tokenizer is saved using Python’s pickle module as tokenizer.pickle.

---

🧠 Example Prediction

# Input:
test_sentences = [
    "granny starting to fear spiders in the garden might be real haha crazy",
    "game of thrones season finale showing this sunday night"
]

# Output:
Sentence: "granny starting to fear spiders in the garden might be real haha crazy"
Predicted Sarcasm Probability: 98.55%

Sentence: "game of thrones season finale showing this sunday night"
Predicted Sarcasm Probability: 4.87%

Running in action: