Twitter-Sentiment140-Analysis

This project uses an LSTM-based neural network to perform binary classification of the Sentiment140 dataset. The model predicts whether a given tweet expresses positive or negative sentiment.

Features

• Data Preprocessing: Includes tokenization, cleaning, and lemmatization of tweets to create structured input for the model.
• Custom Embedding Layer: Utilizes pre-trained GloVe embeddings for word representation, ensuring robust vectorization of textual data.
• Model Design: A two-layer bidirectional LSTM with dropout for regularization, trained using the Adam optimizer.
• Performance: Trained on a subset of the Sentiment140 dataset (~1.6M tweets), achieving effective sentiment classification.

Example Predictions

- Sentences closer to **1.0** indicate positive sentiment. - Sentences closer to **0.0** indicate negative sentiment.

Project Highlights

1. Data Preprocessing:
   • Tokenization using NLTK.
   • Lemmatization with POS tagging.
   • Noise removal (URLs, mentions, punctuation, and stopwords).
2. Visualization: Generates word clouds of frequent words for exploratory data analysis.
3. Model:
   • Embedding layer built from GloVe.
   • Bidirectional LSTMs for context-aware text analysis.
   • Dropout for overfitting prevention.
4. Custom Predictions: Easily classify custom text strings using the trained model.

Results

The model can predict custom sentences with confidence, such as:

print(predict_custom_string("I love chocolate"))  # Output: ~0.93 (positive)
print(predict_custom_string("I hate chocolate"))  # Output: ~0.09 (negative)

Dataset

The Sentiment140 dataset is used for training and testing. It contains 1.6 million tweets labeled as positive (4) or negative (0).

For more details, refer to the notebook.