Handling text features

[bjschoenfeld]

Text columns are currently treated as categorical columns. When these columns get one-hot encoded, computing metafeatures becomes intractable (e.g. PCA and knn metafeatures). There are several easy solutions: disallow text columns by throwing an error, ignore text columns, or don't compute expensive metafeatures which depend of categorical columns.

What metafeatures exist for datasets which contain text? Do we implement those here?