RPT

Can AI help automate human-easy but computer-hard data preparation tasks that burden scientists, practitioners, and crowd workers? We answer this question by presenting RPT, a denoising autoencoder for tuple-to-X models (" X " could be tuple, token, label, JSON, so on). RPT is pre-trained tuple-to-tuple model corrupting the input tuple then learning to reconstruct original tuple. It adopts Transformer-based neural translation architecture consists of bidirectional encoder (similar BERT) left-to-right autoregressive decoder GPT), leading generalization both BERT GPT. The can already support several common such as cleaning, auto-completion schema matching. Better still, fine-tuned on wide range tasks, value normalization, transformation, annotation, etc. To complement we also discuss appealing techniques collaborative training few-shot entity resolution, NLP question-answering information extraction. In addition, identify series research opportunities advance field preparation.