🤖 BERT for Sentiment Analysis — Airline Tweet Complaint Detection

Fine-tuning bert-base-uncased on airline tweets to classify whether a tweet is a complaint or non-complaint — achieving strong performance over a TF-IDF + Naive Bayes baseline.

---

📌 Project Overview

This project implements a full NLP pipeline for binary sentiment classification on airline tweets, comparing:

Approach	Description
Baseline	TF-IDF Vectorizer + Multinomial Naive Bayes
Fine-tuned	bert-base-uncased via Hugging Face Transformers + PyTorch

The dataset contains 3400 tweets (1700 complaints, 1700 non-complaints) from airline handles, split into train/validation/test sets.

---

🗂️ Repository Structure

Bert-for-sentiment-analysis/
├── BERT-for-sentiment-analysis.ipynb   # Main notebook (end-to-end pipeline)
├── requirements.txt                     # Python dependencies
├── .gitignore                           # Git ignore rules
└── README.md                            # This file

---

🧠 Model Architecture

Input Tweet
    ↓
[BERT Tokenizer] — bert-base-uncased, max_len=64
    ↓
[BertModel] — 768-dim CLS token embedding
    ↓
[Linear(768 → 50)] → ReLU
    ↓
[Linear(50 → 2)]  → Logits
    ↓
Complaint / Non-Complaint

---

📊 Pipeline Steps

1. Data Loading & Preprocessing

• Load complaint1700.csv and noncomplaint1700.csv
• Assign binary labels: complaint=0, non-complaint=1
• 80/20 train-validation split with random_state=2020
• Text cleaning: lowercasing, removing mentions, punctuation normalization, stopword removal

2. Baseline — TF-IDF + Naive Bayes

• TfidfVectorizer with ngram_range=(1,3), binary=True
• MultinomialNB with hyperparameter tuning via 5-fold stratified cross-validation
• Evaluated using AUC-ROC curve and accuracy

3. BERT Fine-tuning

• Tokenization via BertTokenizer (bert-base-uncased, max_length=64)
• PyTorch DataLoader with batch size 32
• Custom BertClassifier with a 2-layer classification head
• Optimizer: AdamW (lr=5e-5, eps=1e-8)
• Scheduler: Linear warmup with Hugging Face get_linear_schedule_with_warmup
• Trained for 2 epochs on train set, evaluated on validation set
• Final model trained on the full dataset (train + val) for test predictions

4. Evaluation

• ROC-AUC curve plotted for both baseline and BERT
• Accuracy and AUC reported on validation set
• Test predictions generated with confidence threshold 0.9

---

⚙️ Setup & Usage

Prerequisites

• Python 3.8+
• CUDA-capable GPU (optional but recommended)

Installation

git clone https://github.com/DevyaniD19/Bert-for-sentiment-analysis.git
cd Bert-for-sentiment-analysis
pip install -r requirements.txt

Running the Notebook

jupyter notebook BERT-for-sentiment-analysis.ipynb

Google Colab: This notebook was developed on Colab. Simply upload it and run — all dependencies and data are downloaded automatically.

---

📦 Dependencies

See requirements.txt for the full list. Key packages:

• torch — deep learning framework
• transformers — Hugging Face BERT model & tokenizer
• scikit-learn — TF-IDF, Naive Bayes, evaluation metrics
• pandas, numpy — data manipulation
• matplotlib — ROC curve visualization
• nltk — stopwords for text preprocessing

---

📈 Results

Model	Validation AUC	Validation Accuracy
TF-IDF + Naive Bayes	~0.87	~81%
Fine-tuned BERT	~0.94+	~88%+

Results may vary slightly due to random seeds and hardware.

---

👩‍💻 Author

Devyani Deore 📧 devyanid@umich.edu 🔗 github.com/DevyaniD19

---

📄 License

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