An NLP based anonymizer for archival documents metadata.
This repository contains code written for the redaction of sensitive strings from the entries of the Swiss Federal Archive (Schweizerisches Bundesarchiv - BAR).
For purposes of illustration synthetic example data (in German) is included.
The synthetic example data contains the following set of (manually applied) labels in the format of html markup:
• <n> natural person: usually a first or family name
• <j> legal entity: companies, associations; not explicitly including product names
• <d> birth date: full day, month, year
• <a> social security number, military registration number or other identifier, such as the foreigners register (Ordipro) maintained by the Federal Department of Foreign Affairs.
The task is to recognize and blacken all strings of text that correspond to the redactable labels <n>, <j>, <d> and <a> in unseen data.
Example:
Input: « Strafverfahren gegen Muster, Max, 14.10.1967 wegen Betrug »
Output: « Strafverfahren gegen ███, ███, ███ wegen Betrug »
The system combines three different types of classifiers into a feature-based ensemble system: Conditional Random Fields (CRFs), Support Vector Machines (SVMs) and Multilayer Perceptrons (MLP) based on their implementation in Scikit-learn.
Optional relabeling according to a defined set of rules is applied within a post-processing step to reach a final prediction.
We distinguish between token-based and context-based types of features.
Token-based features:
• The token itself (lower-cased)
• The lemma (base form) of the token
• The part-of-speech (pos) tag of the token
• The shape of the token (e.g., Xxxx oder dd.dd.dd)
• If the token is alphabetic/numeric/a stopword/a punctuation character
• Character 2-grams of the token
• If the token looks like a date according to regular expressions
• If the token looks like an identifier (label <a>) according to regular expressions
• If the token is found in the dictionary of personnames
• If the token is found in the dictionary of swiss place names
Context-based features:
Features within a context window of n (default: 4) preceding and subsequent tokens is taken into account.
• The (lower-cased) context tokens
• The part-of-speech (pos-tags) of the context tokens
• If the context tokens are punctuation characters
• The shape of the context tokens
We use tokens, lemmas, pos-tags, token shape, information regarding token nature (alphabetic/numeric/a stopword/a punctuation character) as provided by SpaCys tokenization (adapted de_core_news_lg pipeline).
Installation of the necessary environment:
conda create --name tcc-bar-anonym
conda activate tcc-bar-anonym
conda install python=3.10
conda install ipykernel
conda install joblib
conda install pandas
conda install -c conda-forge scikit-learn
conda install spacy
python -m spacy download de_core_news_lg
conda install babel
conda install seaborn
conda install python-crfsuite
pip install sklearn-crfsuiteUse the configuration file (config.ini) to set up the following paths and parameters:
The code can be used from the command line or by importing it as a package (see included jupyter notebook cross_validate.ipynb).
Call from within the code directory.
python3 train_models.py --in_path ../data_syn/annotated_data_example.tsv --models_dir ../models_syn_new --config_file ../config_syn.inipython3 predict_labels.py --in_path ../data_syn/annotated_data_example.tsv --out_path ../data_syn/output_anon.tsv --models_dir ../models_syn_new --config_file ../config_syn.iniIf you have questions or comments, please do not hesitate to contact us under the following address:
