Emotion Detection from Tweets

Overview

This repository contains the implementation of a logistic regression classifier and a multi-layer neural network to detect emotions from Tweets. This homework is part of the NLP course, focusing on building these models from scratch without using built-in functions from libraries like sklearn for the machine learning part.

The script will train the logistic regression and neural network models on the training data and predict emotions on the test data. The predictions will be saved in test_lr.csv and test_nn.csv for logistic regression and neural network models, respectively.

Dataset

The dataset consists of 1200 labeled Tweets for training and 800 Tweets for testing. The Tweets are annotated with one of six emotions: joy, love, sadness, anger, fear, surprise. The test dataset does not include labels.

Implementation Details

Logistic Regression

• The LR() function in main.py learns a logistic regression classifier using cross-validation and outputs predictions in test_lr.csv.
• Loss Function: Cross-Entropy with L1 regularization.
• Feature Representation: TF-IDF values for each word.

Multi-layer Neural Network

• The NN() function in main.py learns a multi-layer neural network classifier using cross-validation and outputs predictions in test_nn.csv.
• Architecture: One hidden layer with 50 nodes. Input layer based on TF-IDF value dimensions and output layer for one-hot encoded emotions.
• Activation Functions: ReLU for the hidden layer and softmax for the output layer.
• Loss Function: Cross-Entropy combined with L1 or L2 regularization.

Both functions ensure that the ID-to-text-to-emotion mapping is preserved as required.

Model Details and Hyperparameters

Hyperparameter Tuning

• Logistic Regression: Learning rate and regularization were tuned using 5-fold cross-validation and grid search.
• Neural Network: Learning rate, regularization rate and type, number of hidden nodes, and batch size were tuned. Observations on overfitting were noted based on training and validation loss comparisons.
• Detailed learning curves were used to decide the best parameters and to observe the effects of different hyperparameter settings on model performance and overfitting.

Contact

For any queries regarding the code or the homework, please create an issue in this repository.

License

This project is licensed under the MIT License - see the LICENSE.md file for details.