NLP Assignment – Twitter Sentiment Analysis

Author: Neha Narkhede PRN: 202301040146

---

Overview

This project performs binary sentiment analysis on the Sentiment140 dataset, which contains 1.6 million tweets labeled as positive or negative. The goal is to preprocess raw tweet text using NLP techniques and train machine learning classifiers to predict tweet sentiment.

---

Dataset

File	Description
training.1600000.processed.noemoticon.csv	Sentiment140 dataset with 1.6M tweets

Columns used:

• target – Sentiment label (0 = Negative, 4 → remapped to 1 = Positive)
• text – Raw tweet content

---

Project Structure

NLP_Assignment/
├── dataset/
│   └── archive (6)/
│       └── training.1600000.processed.noemoticon.csv
├── Code_nlp.ipynb     # Main Jupyter Notebook
├── requirements.txt   # Python dependencies
└── README.md

---

Pipeline

1. Data Loading

• Loaded the CSV with latin-1 encoding
• Retained only target and text columns
• Dropped rows with missing values (dropna)
• Remapped label 4 → 1 for binary classification

2. Dataset Reduction (moved before preprocessing for efficiency)

• Sampled 50,000 tweets (random state = 42) before running the NLP pipeline
• This avoids preprocessing all 1.6M rows, significantly reducing runtime

3. NLP Preprocessing

Step	Description
Tokenization	Split text into tokens using NLTK word_tokenize (lowercased)
Stopword Removal	Removed English stopwords using NLTK
Stemming	Applied Porter Stemmer to reduce words to their root form
Lemmatization	Applied WordNet Lemmatizer on non-stemmed tokens (independent step)
Clean Text	Combined pipeline: lowercasing → URL removal → mention removal → non-alpha removal → stopword removal → short token filter (len > 2) → stemming + lemmatization

4. Train-Test Split

• 80% train / 20% test (random state = 42)

5. Feature Extraction

Vectorizer	Config
CountVectorizer	max_features=5000, ngram_range=(1,2)
TF-IDF Vectorizer	max_features=5000, ngram_range=(1,2)

6. Model Training

Model	Feature Input
Logistic Regression (max_iter=200)	TF-IDF
Multinomial Naive Bayes	CountVectorizer

---

Results

Model	Accuracy	Precision (avg)	Recall (avg)	F1 (avg)
Logistic Regression	75.17%	0.75	0.75	0.75
Naive Bayes	74.70%	0.75	0.75	0.75

Key Observations

1. Logistic Regression slightly outperforms Naive Bayes (~0.5% higher accuracy).
2. Logistic Regression captures feature relationships, while Naive Bayes assumes word independence.
3. Both models show balanced precision/recall across classes — no class bias.
4. ~75% accuracy is reasonable given the noisy nature of Twitter data (slang, abbreviations, informal language).
5. NLP preprocessing (stopword removal, stemming, lemmatization) significantly improved model performance.

---

Visualizations

• Class Distribution – Bar chart of positive vs. negative tweet counts (with colors)
• Confusion Matrix – Heatmap with true/predicted axis labels for Logistic Regression
• Model Accuracy Comparison – Bar chart with value labels and y-axis limits
• Model Performance Comparison – Line chart of Precision, Recall, F1 for both models
• Top 20 Words – Bar chart of the most frequent words in cleaned tweets (with tight_layout)

---

Improvements Made

Area	Change
Data loading	Added dropna() to remove rows with missing text or labels
Sampling order	Moved sampling to before preprocessing (saves ~30x processing time)
NLTK downloads	Consolidated all downloads into a single cell at the start
Lemmatization	Fixed standalone lemmatized column to use no_stopwords (not stemmed tokens)
Clean text	Added len(w) > 2 filter to remove very short noise tokens
Confusion matrix	Added xticklabels, yticklabels, axis titles, and cmap='Blues'
Accuracy bar chart	Added y-axis limits (0.5–1.0) and numeric value labels on bars
All charts	Added tight_layout() to prevent label cutoff

---

Requirements

Install all dependencies at once:

pip install -r requirements.txt

Or manually:

pip install pandas numpy matplotlib seaborn scikit-learn nltk

Download NLTK resources (run once inside the notebook):

import nltk
nltk.download('punkt')
nltk.download('punkt_tab')
nltk.download('stopwords')
nltk.download('wordnet')

---

How to Run

1. Open Code_nlp.ipynb in Google Colab or Jupyter Notebook.
2. Upload the dataset CSV when prompted (Colab) or place it in the working directory.
3. Run all cells sequentially.