Named Entity Recognition for the Legal Domain

Dependencies

Get a copy of the repository:

$ git clone [email protected]:openlegaldata/legal-ner.git

Before getting started you have to install the Python dependencies, which will also install the required language models.

$ cd legal-ner
$ pipenv --python 3.7
$ pipenv install

To be able to run python scripts the project root needs to be added to the python module search path.

$ export PYTHONPATH=$PWD

All python statements below should be run in the shell provided by pipenv.

$ pipenv shell

Extract Entities

To extract entities for the Open Legal Data Platform (OLDP) run:

$ python legal_ner/oldp/annotate.py -k=your_api_key -p

To get more information about the usage run:

$ python legal_ner/oldp/annotate.py --help

You can also extract and locally visualize entities for a single case using:

$ python legal_ner/oldp/visualize.py -k=your-your_api_key-key -i=case_id -p=joined

Training your Model

Obtaining Data

You can download cases from the OLDP website using:

$ python legal_ner/utils/oldp_scraper.py -o=data -k=your_api_key -c=case_id_1,case_id2,...

The data has to be annotated in the following format:

{
  "text": "Denn das FG hat --wie oben dargelegt-- bindend festgestellt, dass die Klägerin das Motorrad gerade nicht zur Ausfuhr, sondern zur Nutzung in den USA erworben hat.",
  "entities": [[9, 11, "ORG"], [145, 148, "LOC"]],
}

Each line contains one json object. Store the labeled sentences in data/annotations.txt and split them into a training and testing dataset with:

$ python legal_ner/utils/split_data.py --data=data/annotations.txt --train=data/train.txt --test=data/test.txt

Training

Currently only the NER module can be trained. The following command loads the training and test datasets from data/ and saves the trained model to models/legal-de.

$ python legal_ner/training/train_ner.py -t=data/train.txt -l=data/test.txt --epochs=4 -o=models/legal-de -v

Evaluation

You can evaluate the performance on a given model (e.g. models/legal-de) by providing an evaluation dataset (e.g. data/test.txt) and running:

$ python eval.py -l=data/test.txt -m=models/legal-de

Pretrained Language Models

This repository hosts its own pretrained language models, specific for the German legal domain:

• German Legal-Domain Language Model

Usage:

import spacy
nlp = spacy.load(path_to_model)