Sherlock is a neural network trained on real-world datasets collected from the web. Unlike rule-based approaches based on hard-coded regular expressions and values, Sherlock detects types using word embeddings and distributions of characters.

Because Sherlock is trained on real data, it is robust to messy entries (e.g. blanks, mispellings, malformatted values) and includes over 20 semantic types. See for yourself by changing table values: