https://github.com/msthamizh/toxicity-detection-in-tweets

Developed a machine learning model to classify tweets as toxic or non-toxic using NLP techniques like tokenization, lemmatization, and TF-IDF vectorization. Evaluated multiple classifiers and analyzed performance using classification reports, confusion matrices, and ROC-AUC curves.
https://github.com/msthamizh/toxicity-detection-in-tweets

machine-learning natural-language-processing nltk python

Last synced: about 1 month ago
JSON representation

Developed a machine learning model to classify tweets as toxic or non-toxic using NLP techniques like tokenization, lemmatization, and TF-IDF vectorization. Evaluated multiple classifiers and analyzed performance using classification reports, confusion matrices, and ROC-AUC curves.

• Host: GitHub
• URL: https://github.com/msthamizh/toxicity-detection-in-tweets
• Owner: MSThamizh
• Created: 2024-11-21T15:01:09.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2024-12-20T12:42:43.000Z (4 months ago)
• Last Synced: 2025-01-26T04:11:20.939Z (3 months ago)
• Topics: machine-learning, natural-language-processing, nltk, python
• Language: Jupyter Notebook
• Homepage:
• Size: 3.05 MB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# **Toxicity Detection in Tweets**

This project focuses on detecting toxic tweets using advanced Natural Language Processing (NLP) techniques. The goal is to classify tweets as either toxic or non-toxic, based on the content, to help prevent harmful online communication.

## **Problem Statement**

The task is to build a model that classifies tweets into two categories: toxic and non-toxic. Toxicity is defined as language that is harmful, abusive, or hateful. The model is trained using a dataset of labeled tweets and evaluated using classification metrics.

## Workflow

1. **Data Collection**: The dataset consists of labeled tweets, where each tweet is classified as either toxic or non-toxic.

2. **Data Preprocessing**: Text data is cleaned and preprocessed by:

   - Removing special characters, URLs, and stopwords.

   - Tokenizing text and lemmatizing words.

3. **Feature Extraction**: 

   - Used **TF-IDF vectorization** to convert the text data into numerical features for the machine learning models.

4. **Model Training**: 

   - Trained multiple machine learning classifiers like **Logistic Regression**, **Random Forest**, and **SVM**.

5. **Model Evaluation**: 

   - Evaluated models using metrics such as **Accuracy**, **Precision**, **Recall**, **F1-Score**, and **ROC-AUC**.

6. **Prediction**: 

   - The model predicts whether a given tweet is toxic or non-toxic.

## Features

- **Text Preprocessing**: 

  - Removal of special characters, URLs, and non-alphabetic characters.

  - Tokenization and lemmatization for better feature extraction.

  

- **TF-IDF Vectorization**: 

  - Convert text data into numerical features using Term Frequency-Inverse Document Frequency (TF-IDF).

  

- **Model Training & Evaluation**:

  - Multiple models are trained, evaluated, and compared based on classification metrics.

- **User Interface**: 

  - User can input a tweet and the model will classify it as toxic or non-toxic.

  

## Technologies Used

- **Python**: Primary language for model development.

- **NLP Libraries**: 

  - **NLTK** for text preprocessing (tokenization, lemmatization).

  - **Scikit-learn** for machine learning algorithms and evaluation.

- **TF-IDF**: For converting text into numerical features.

## **Results**

### **Classification Models**

| Model                  | Test Accuracy | Precision (Toxic) | Recall (Toxic) | F1-Score (Toxic) | Key Insights                              |

|-------------------------|---------------|--------------------|----------------|------------------|-------------------------------------------|

| Decision Tree           | 87%           | 80%               | 91%            | 85%             | High interpretability; prone to overfitting. |

| Random Forest           | 91%           | 88%               | 90%            | 89%             | Robust to overfitting; high accuracy.       |

| Multinomial Naive Bayes | 89%           | 87%               | 87%            | 87%             | Performs well with text data.              |

| K-Nearest Neighbors     | 81%           | 71%               | 91%            | 80%             | Struggles with imbalanced classes.         |

### **Visualizations**

- **ROC Curves**:

  Each model's ROC curve shows its capability to distinguish between toxic and non-toxic tweets.

- **Confusion Matrices**:

  Evaluate the model's ability to avoid false positives and false negatives.

## References

- **Python**: [https://docs.python.org/3/](https://docs.python.org/3/)

- **NLTK**: [https://www.nltk.org/](https://www.nltk.org/)

- **Scikit-learn Documentation**: [https://scikit-learn.org/stable/](https://scikit-learn.org/stable/)