Tweet Classification for Disaster Information

Natural Language Processing Project – Naive Bayes & Logistic Regression

This project was developed for academic purposes (NLP midterm) and can be used as a learning resource.

Project Description

Twitter is often a primary source of real-time information during disasters. However, not all tweets containing words such as “ablaze” or “fire” actually refer to real disaster events.

Examples:

• "The concert was ablaze!" → Not a disaster
• "The entire forest is ablaze." → Disaster

This project aims to build Naive Bayes and Logistic Regression models to classify whether a tweet truly contains disaster-related information.

---

Dataset

• Source: Kaggle – Real or Not? Disaster Tweets

• Total data: 10,876 English tweets

  • 7,613 training data (used in the analysis)
  • 3,263 test data (not used because labels are unavailable)

Labels:

• 1 → Tweet references a real disaster
• 0 → Tweet does not reference a disaster

---

Methodology

1. Exploratory Data Analysis (EDA) – tweet distribution, text length, frequent words
2. Preprocessing – lowercase, removing URLs, numbers, punctuation; stopword removal; lemmatization; tokenization
3. Vectorization – TF-IDF with ngram_range=(1,2) and max_features=10,000
4. Feature Selection – Chi-Square (SelectKBest, top 3,000 features)
5. Modeling – Naive Bayes and Logistic Regression with GridSearchCV (5-fold cross-validation)
6. Evaluation – confusion matrix, precision, recall, F1-score, accuracy
7. Interpretability – error analysis, POS tagging, Named Entity Recognition (NER)

---

Results

Model	Precision	Recall	F1-Score	Accuracy
Logistic Regression	0.84	0.83	0.83	83.7%
Naive Bayes	0.86	0.85	0.86	85.9%

• Naive Bayes outperforms Logistic Regression on all evaluation metrics.
• Both models perform well for short-text classification tasks such as tweets.

---

Insights

• Preprocessing is crucial to reduce noise commonly found in social media text.
• Naive Bayes performs strongly on short, sparse text due to its independence assumptions.
• Logistic Regression is more stable on balanced feature distributions.
• POS tagging and NER help identify recurring locations, times, and organizations mentioned in disaster-related tweets.

---

How to Run

1. Clone the repository

   git clone https://github.com/username/tweet-disaster-classification.git
   cd tweet-disaster-classification