Skip to main content

Mutli-language data anonymization with Microsoft Presidio

Open In Collab

Use case​

Multi-language support in data pseudonymization is essential due to differences in language structures and cultural contexts. Different languages may have varying formats for personal identifiers. For example, the structure of names, locations and dates can differ greatly between languages and regions. Furthermore, non-alphanumeric characters, accents, and the direction of writing can impact pseudonymization processes. Without multi-language support, data could remain identifiable or be misinterpreted, compromising data privacy and accuracy. Hence, it enables effective and precise pseudonymization suited for global operations.

Overview​

PII detection in Microsoft Presidio relies on several components - in addition to the usual pattern matching (e.g. using regex), the analyser uses a model for Named Entity Recognition (NER) to extract entities such as:

  • PERSON
  • LOCATION
  • DATE_TIME
  • NRP
  • ORGANIZATION

[Source]

To handle NER in specific languages, we utilize unique models from the spaCy library, recognized for its extensive selection covering multiple languages and sizes. However, it's not restrictive, allowing for integration of alternative frameworks such as Stanza or transformers when necessary.

Quickstart​

# Install necessary packages
# ! pip install langchain langchain-experimental openai presidio-analyzer presidio-anonymizer spacy Faker
# ! python -m spacy download en_core_web_lg
from langchain_experimental.data_anonymizer import PresidioReversibleAnonymizer

anonymizer = PresidioReversibleAnonymizer(
analyzed_fields=["PERSON"],
)

By default, PresidioAnonymizer and PresidioReversibleAnonymizer use a model trained on English texts, so they handle other languages moderately well.

For example, here the model did not detect the person:

anonymizer.anonymize("Me llamo SofĂ­a")  # "My name is SofĂ­a" in Spanish
    'Me llamo SofĂ­a'

They may also take words from another language as actual entities. Here, both the word 'Yo' ('I' in Spanish) and SofĂ­a have been classified as PERSON:

anonymizer.anonymize("Yo soy SofĂ­a")  # "I am SofĂ­a" in Spanish
    'Bridget Kirk soy Sally Knight'

If you want to anonymise texts from other languages, you need to download other models and add them to the anonymiser configuration:

# Download the models for the languages you want to use
# ! python -m spacy download en_core_web_md
# ! python -m spacy download es_core_news_md
nlp_config = {
"nlp_engine_name": "spacy",
"models": [
{"lang_code": "en", "model_name": "en_core_web_md"},
{"lang_code": "es", "model_name": "es_core_news_md"},
],
}

We have therefore added a Spanish language model. Note also that we have downloaded an alternative model for English as well - in this case we have replaced the large model en_core_web_lg (560MB) with its smaller version en_core_web_md (40MB) - the size is therefore reduced by 14 times! If you care about the speed of anonymisation, it is worth considering it.

All models for the different languages can be found in the spaCy documentation.

Now pass the configuration as the languages_config parameter to Anonymiser. As you can see, both previous examples work flawlessly:

anonymizer = PresidioReversibleAnonymizer(
analyzed_fields=["PERSON"],
languages_config=nlp_config,
)

print(
anonymizer.anonymize("Me llamo SofĂ­a", language="es")
) # "My name is SofĂ­a" in Spanish
print(anonymizer.anonymize("Yo soy SofĂ­a", language="es")) # "I am SofĂ­a" in Spanish
    Me llamo Michelle Smith
Yo soy Rachel Wright

By default, the language indicated first in the configuration will be used when anonymising text (in this case English):

print(anonymizer.anonymize("My name is John"))
    My name is Ronnie Ayala

Advanced usage​

Custom labels in NER model​

It may be that the spaCy model has different class names than those supported by the Microsoft Presidio by default. Take Polish, for example:

# ! python -m spacy download pl_core_news_md

import spacy

nlp = spacy.load("pl_core_news_md")
doc = nlp("Nazywam siÄ™ Wiktoria") # "My name is Wiktoria" in Polish

for ent in doc.ents:
print(
f"Text: {ent.text}, Start: {ent.start_char}, End: {ent.end_char}, Label: {ent.label_}"
)
    Text: Wiktoria, Start: 12, End: 20, Label: persName

The name Victoria was classified as persName, which does not correspond to the default class names PERSON/PER implemented in Microsoft Presidio (look for CHECK_LABEL_GROUPS in SpacyRecognizer implementation).

You can find out more about custom labels in spaCy models (including your own, trained ones) in this thread.

That's why our sentence will not be anonymized:

nlp_config = {
"nlp_engine_name": "spacy",
"models": [
{"lang_code": "en", "model_name": "en_core_web_md"},
{"lang_code": "es", "model_name": "es_core_news_md"},
{"lang_code": "pl", "model_name": "pl_core_news_md"},
],
}

anonymizer = PresidioReversibleAnonymizer(
analyzed_fields=["PERSON", "LOCATION", "DATE_TIME"],
languages_config=nlp_config,
)

print(
anonymizer.anonymize("Nazywam siÄ™ Wiktoria", language="pl")
) # "My name is Wiktoria" in Polish
    Nazywam siÄ™ Wiktoria

To address this, create your own SpacyRecognizer with your own class mapping and add it to the anonymizer:

from presidio_analyzer.predefined_recognizers import SpacyRecognizer

polish_check_label_groups = [
({"LOCATION"}, {"placeName", "geogName"}),
({"PERSON"}, {"persName"}),
({"DATE_TIME"}, {"date", "time"}),
]

spacy_recognizer = SpacyRecognizer(
supported_language="pl",
check_label_groups=polish_check_label_groups,
)

anonymizer.add_recognizer(spacy_recognizer)

Now everything works smoothly:

print(
anonymizer.anonymize("Nazywam siÄ™ Wiktoria", language="pl")
) # "My name is Wiktoria" in Polish
    Nazywam siÄ™ Morgan Walters

Let's try on more complex example:

print(
anonymizer.anonymize(
"Nazywam się Wiktoria. Płock to moje miasto rodzinne. Urodziłam się dnia 6 kwietnia 2001 roku",
language="pl",
)
) # "My name is Wiktoria. PĹ‚ock is my home town. I was born on 6 April 2001" in Polish
    Nazywam się Ernest Liu. New Taylorburgh to moje miasto rodzinne. Urodziłam się 1987-01-19

As you can see, thanks to class mapping, the anonymiser can cope with different types of entities.

Custom language-specific operators​

In the example above, the sentence has been anonymised correctly, but the fake data does not fit the Polish language at all. Custom operators can therefore be added, which will resolve the issue:

from faker import Faker
from presidio_anonymizer.entities import OperatorConfig

fake = Faker(locale="pl_PL") # Setting faker to provide Polish data

new_operators = {
"PERSON": OperatorConfig("custom", {"lambda": lambda _: fake.first_name_female()}),
"LOCATION": OperatorConfig("custom", {"lambda": lambda _: fake.city()}),
}

anonymizer.add_operators(new_operators)
print(
anonymizer.anonymize(
"Nazywam się Wiktoria. Płock to moje miasto rodzinne. Urodziłam się dnia 6 kwietnia 2001 roku",
language="pl",
)
) # "My name is Wiktoria. PĹ‚ock is my home town. I was born on 6 April 2001" in Polish
    Nazywam się Marianna. Szczecin to moje miasto rodzinne. Urodziłam się 1976-11-16

Limitations​

Remember - results are as good as your recognizers and as your NER models!

Look at the example below - we downloaded the small model for Spanish (12MB) and it no longer performs as well as the medium version (40MB):

# ! python -m spacy download es_core_news_sm

for model in ["es_core_news_sm", "es_core_news_md"]:
nlp_config = {
"nlp_engine_name": "spacy",
"models": [
{"lang_code": "es", "model_name": model},
],
}

anonymizer = PresidioReversibleAnonymizer(
analyzed_fields=["PERSON"],
languages_config=nlp_config,
)

print(
f"Model: {model}. Result: {anonymizer.anonymize('Me llamo SofĂ­a', language='es')}"
)
    Model: es_core_news_sm. Result: Me llamo SofĂ­a
Model: es_core_news_md. Result: Me llamo Lawrence Davis

In many cases, even the larger models from spaCy will not be sufficient - there are already other, more complex and better methods of detecting named entities, based on transformers. You can read more about this here.

Future works​

  • automatic language detection - instead of passing the language as a parameter in anonymizer.anonymize, we could detect the language/s beforehand and then use the corresponding NER model.