Toward Efficient Unsupervised Web Data Extraction: From Unsupervised to Self-Trained Wrappers