Uncovering, Understanding, and Mitigating Social Biases in Language Models