SkinDeepRESEARCHSteve Seguin

Exploratory results

Find names to redact

Return marked text spans instead of rewriting a document.

Back to decisions without text

Before sharing a document, identify the names that may need masking. This first test covers names in court documents; it does not establish complete anonymization.

MethodName tokens foundOther tokens incorrectly marked
Title rule646 / 1,2296
Trained token classifier1088 / 1,229339

The trained classifier found more names, but also marked more unrelated words. It missed 141 name tokens.

What was tested?

50 real TAB court documents, limited to the first 1,200 tokens each. Human annotations supply the references. A separate 100 documents train the classifier and 25 select its threshold. No privacy or speed guarantee is established.

Data, method and every outcome