Detecting Bias in News Articles

Detecting bias in news articles is a challenging task due to the nuanced and subtle ways bias can be expressed at the sentence level. This repository evaluates the performance of several transformer-based models, including BERT, RoBERTa, DistilBERT, and DeBERTa, as well as a traditional machine learning model, Support Vector Machine (SVM), for the task of sentence-level bias detection.

Models Evaluated

The following models were evaluated for bias detection:

• BERT (Bidirectional Encoder Representations from Transformers)
• RoBERTa (A Robustly Optimized BERT Pretraining Approach)
• DistilBERT (A Smaller, Faster, Cheaper, and Lighter version of BERT)
• DeBERTa (Decoding-enhanced BERT with disentangled attention)
• SVM (Support Vector Machine) - A traditional machine learning approach

Dataset and Preprocessing

The models were tested on a dataset specifically designed for sentence-level bias detection. Preprocessing techniques, such as text cleaning and oversampling, were applied to the data to improve the models' performance.

Key Results

• DistilBERT achieved the best overall performance, balancing computational efficiency with high precision and recall.
• Transformer models generally performed well, highlighting their effectiveness for bias detection in text.
• Surprisingly, the SVM model outperformed certain transformer models, demonstrating that traditional machine learning approaches can still be competitive, especially when combined with the right preprocessing techniques.

Comparative Analysis

This repository provides a comparative analysis of the different models, allowing for a deeper understanding of how transformer models and traditional machine learning approaches differ in their ability to detect bias in news articles.

Getting Started

To get started with the code, clone this repository:

git clone https://github.com/sadaf-rad/bias-detection.git
cd bias-detection