Airline Tweet Sentiment Analysis

Author: Nicolette Mtisi
Tools: Python · NLTK · scikit-learn · pandas · Matplotlib · Seaborn

Overview

This NLP project sorts airline customer tweets into positive, neutral or negative sentiment. Airlines get thousands of these messages every day, and an automatic classifier helps customer-experience teams spot complaints and trends without reading every tweet by hand.

Dataset: Twitter US Airline Sentiment (Kaggle), 14,640 tweets about six U.S. airlines. It is included here as Tweets.csv.

Approach

1. Text cleaning: lowercasing and removing URLs, @mentions, #hashtags and punctuation.
2. NLP preprocessing with NLTK: tokenization, stopword removal and lemmatization.
3. Features: TF-IDF vectorization, keeping the top 5,000 terms.
4. Model: Logistic Regression trained on an 80/20 train/test split.
5. Evaluation: accuracy, a per-class classification report and a confusion matrix.

Results

Accuracy: 80.0% on the held-out test set (2,928 tweets).

Sentiment	Precision	Recall	F1	Test Tweets
Negative	0.82	0.94	0.88	1,889
Neutral	0.67	0.49	0.57	580
Positive	0.82	0.62	0.71	459

Sentiment Distribution	Confusion Matrix

Key takeaways

• Negative tweets make up about 63% of the data, and the model detects them very reliably (94% recall).
• Neutral is the hardest class. These tweets are often questions or plain statements that share wording with the other two classes.
• Example: "I love the friendly service on this airline!" is classified as positive.

How to Run

git clone https://github.com/nic-stack/Twitter-Sentiment-Analysis.git
cd Twitter-Sentiment-Analysis
pip install -r requirements.txt
jupyter notebook sentiment_analysis.ipynb

The notebook downloads the NLTK resources it needs (punkt, stopwords, wordnet) on the first run.

Files

File	Description
sentiment_analysis.ipynb	Full pipeline: preprocessing, training, evaluation and charts
Tweets.csv	Twitter US Airline Sentiment dataset
sentiment_distribution.png	Class distribution chart
confusion_matrix.png	Test-set confusion matrix
requirements.txt	Python dependencies

Next Steps

• Handle the class imbalance with class weights or resampling to improve neutral and positive recall.
• Compare against transformer models. A follow-up project fine-tuned DistilBERT on this task (model · live demo).

Links

• 📝 Medium write-up
• 🌐 Portfolio · LinkedIn