Learning to rank method combining multi-head self-attention with conditional generative adversarial nets.