An unsupervised text representation learning approach through self-instructed-tuning of a pre-trained encoder-decoder language model can effectively augment corpus representation for zero-shot dense retrieval.
