Sentiment Classifiers on Twitter Dataset

Overview

The goal is to develop a sentiment classifier in Python using four different approaches on a given English-language Twitter dataset. Results and comparisons are documented in distinct reports for each implementation. The four approaches are:

1. Logistic Regression with TF-IDF
2. Neural Network in PyTorch with Word2Vec embeddings
3. Fine-tuning pretrained BERT (without Hugging Face's Trainer)
4. Fine-tuning pretrained DistilBERT (without Hugging Face's Trainer)

Detailed description

Dataset

The dataset contains three columns:

• ID: Unique text identifier
• Text: Tweet content
• Label: Sentiment (0 = negative, 1 = positive)

Training, validation, and test sets are pre-split.

Preprocessing

Preprocessing includes four functions: my_lower, my_stopword, my_unpunct, and my_lemmatize. For each classifier, we experiment to determine which preprocessing methods improve performance. All implementations share the same preprocessing options.

Vectorization and Embeddings

Approach 1: TF-IDF vectorization. The sparse TF-IDF matrix feeds directly into logistic regression.

Approach 2: Uses pretrained GloVe embeddings (glove.6B/glove.twitter.27B) converted to Word2Vec format. Tweet embeddings are generated by averaging word vectors.

Approaches 3-4: BERT and DistilBERT use their own pretrained contextual embeddings.

Experiments

Includes manual experiments and automated hyperparameter tuning using the Optuna framework. Key aspects explored:

• Hyperparameter optimization
• Embedding dimensions
• Regularization techniques
• Optimization algorithms
• Architectural variations

All experimentation aims to prevent underfitting/overfit and maximize validation performance.

Evaluation

Metrics include accuracy, precision, recall, and F1-score. Visualizations include:

• Learning curves
• ROC curves
• Confusion matrices

Notes

• Results for each model on the test set were obtained via a private Kaggle competition as part of the course for which these assignments were developed. Consequently, we cannot reproduce those test‑set results locally.
• Reports for BERT and DistilBERT models are combined in a single document.

License

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