ym88659208ym87991671
Multi-language data anonymization with Microsoft Presidio | Документация для разработчиков

Multi-language data anonymization with Microsoft Presidio

Обновлено 4 марта 2024

Open In Colab

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

# Download model
!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 Sofía")  # "My name is Sofía" in Spanish
    'Me llamo Sofía'

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

anonymizer.anonymize("Yo soy Sofía")  # "I am Sofía" in Spanish
    'Kari Lopez soy Mary Walker'

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 Sofía", language="es")
) # "My name is Sofía" in Spanish
print(anonymizer.anonymize("Yo soy Sofía", language="es")) # "I am Sofía" in Spanish
    Me llamo Christopher Smith
Yo soy Joseph Jenkins

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 Shawna Bennett

Usage with other frameworks

Language detection

One of the drawbacks of the presented approach is that we have to pass the language of the input text directly. However, there is a remedy for that - language detection libraries.

We recommend using one of the following frameworks:

  • fasttext (recommended)
  • langdetect

From our experience fasttext performs a bit better, but you should verify it on your use case.

# Install necessary packages
%pip install --upgrade --quiet fasttext langdetect

langdetect

import langdetect
from langchain.schema import runnable


def detect_language(text: str) -> dict:
language = langdetect.detect(text)
print(language)
return {"text": text, "language": language}


chain = runnable.RunnableLambda(detect_language) | (
lambda x: anonymizer.anonymize(x["text"], language=x["language"])
)
chain.invoke("Me llamo Sofía")
    es
    'Me llamo Michael Perez III'
chain.invoke("My name is John Doe")
    en
    'My name is Ronald Bennett'

fasttext

You need to download the fasttext model first from https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.ftz

import fasttext

model = fasttext.load_model("lid.176.ftz")


def detect_language(text: str) -> dict:
language = model.predict(text)[0][0].replace("__label__", "")
print(language)
return {"text": text, "language": language}


chain = runnable.RunnableLambda(detect_language) | (
lambda x: anonymizer.anonymize(x["text"], language=x["language"])
)
    Warning : `load_model` does not return WordVectorModel or SupervisedModel any more, but a `FastText` object which is very similar.
chain.invoke("Yo soy Sofía")
    es
    'Yo soy Angela Werner'
chain.invoke("My name is John Doe")
    en
    'My name is Carlos Newton'

This way you only need to initialize the model with the engines corresponding to the relevant languages, but using the tool is fully automated.

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 się 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 się Wiktoria", language="pl")
) # "My name is Wiktoria" in Polish
    Nazywam się 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 się Wiktoria", language="pl")
) # "My name is Wiktoria" in Polish
    Nazywam się 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. Płock 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. Płock 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 Sofía', language='es')}"
)
    Model: es_core_news_sm. Result: Me llamo Sofía
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.

ПАО Сбербанк использует cookie для персонализации сервисов и удобства пользователей.
Вы можете запретить сохранение cookie в настройках своего браузера.