1.
Sukanya Saha
– Department of Computer Science and Engineering, University of Calcutta, West Bengal, India.
2.
Santanu Modak
– Department of Information Technology, Bengal College of Engineering and Technology, West Bengal, India.
3.
Bablu Pramanik
– Department of Computer Science and Engineering, NSHM Knowledge Campus, Durgapur Group of Institution, West Bengal, India.
Abstract
Organisations are embracing sentiment analysis to measure the opinion of people and make informed decisions based on information. In spite of that, real-time sentiment analysis systems tend to propagate prejudices towards protected attributes such as race, gender, and age. The paper proposes a new model of real-time bias detection and reduction in sentiment analysis. We apply a Convolutional Neural Network (CNN) with optimised reweighting and adaptive reweighting mechanisms alongside a fairness-aware loss function using the CI-MNIST data as a proxy for sentiment data with protected attributes. The system is constantly checking fairness indicators and provides alerts whenever bias goes beyond a set of predefined limits. Experiments also show
that our methodology has high classification accuracy (81.5%) and shows a significant positive impact on measures of fairness, including demographic parity and equal opportunity. The framework is low-latency and modular and can be integrated into existing sentiment analysis pipelines, which will enhance the creation of ethical and fair artificial intelligence (AI) systems.
Keywords Bias mitigation, Convolutional neural networks, Fairness in AI, Real-time systems, Sentiment analysis.
