Isomorphism 0.2.0 released

Isomorphism 0.2.0 is out, right after the first version. 0.1.0 covered the kinds with check digits. 0.2.0 adds the rest of what a Swiss text contains: people, companies and places. It is still pip install isomorphism, MIT-licensed, Python 3.12 or newer, no dependencies, fully local.

What is in 0.2.0

• Persons. After a title (Hr. Keller, Frau Dr. Müller, M. Jean Dupont, signora Bianchi), as first name plus surname from a list of about 440 common first names, as Muster, Anna, the same surname again later in the text, and a first name on its own.
• Companies by legal form: AG, GmbH, SA, Sàrl, Sagl, KlG, KG, Ltd, Inc, LLC.
• Addresses: street with number and postcode with place, in German, French, Italian and English (Musterstrasse 12, 9000 St. Gallen, Rue de Lausanne 12, 1003 Lausanne, Via Nassa 8, 6900 Lugano).
• Dates of birth, only after context words such as geb., Geburtsdatum, né le, nata il, date of birth. Invoice dates stay untouched.
• Card numbers with a Luhn check and accounts: PostFinance numbers with their mod-10 check and digits after Konto, compte, conto or account.
• Streaming. Restorer and restore_stream restore an answer that arrives in pieces, also when a placeholder such as [AHV_1] is split between two pieces.
• Authorities such as Steueramt St. Gallen or Handelsregisteramt Zürich are never replaced. keep now accepts compiled patterns as well as exact values.
• Presidio as an optional backend for names in running text: anonymize(text, detectors=[presidio_detector(analyzer)]). It is not a dependency and nothing is downloaded by Isomorphism.
• Measurement. The README has a table with expected, found, missed and false alarms for every kind, on invented Swiss texts in four languages. A test keeps it current.

How sure is it

Names, companies and addresses come from patterns, titles and word lists. They are less certain than an AHV number with a valid check digit, and they can report too much: a four-digit number followed by a capitalised word after a comma can look like a postcode and a place. Every hit carries a confidence, so you can filter. In the invented test texts no listed value gets through; the two known false alarms are named in the README. That is a small regression set, not a benchmark on real documents, and no tool finds all personal data.

Install and try

pip install --upgrade isomorphism

from isomorphism import Restorer, anonymize

safe, mapping = anonymize("Hr. Keller, Musterstrasse 12, 9000 St. Gallen, AHV 756.1234.5678.97")
# '[PERSON_1], [ADDRESS_1], AHV [AHV_1]'

r = Restorer(mapping)
for piece in ["Die Nummer [AH", "V_1] von [PER", "SON_1] ist gültig."]:
    print(r.feed(piece), end="")
print(r.flush())

The live demo runs the real package in your browser and shows the streaming step by step. The reference, the FastAPI guide and the Presidio guide are at packages.sweber.dev/isomorphism/docs.