We present a novel training paradigm for neural abstractive summarization. Instead of using MLE training alone, we introduce a contrastive learning component, which encourages the abstractive models to estimate the probability of system-generated summaries more accurately.