How to train BiLSTM-CNN-CRF with German LER Dataset

The Implementation for Sequence Tagging of BiLSTM-CNN-CRF is used for the training. See the GitHub Repo for more information.

This repository contains a BiLSTM-CRF implementation that used for NLP Sequence Tagging (for example POS-tagging, Chunking, or Named Entity Recognition). The implementation is based on Keras 2.2.0 and can be run with Tensorflow 1.8.0 as backend. It was optimized for Python 3.5 / 3.6. It does not work with Python 2.7.

Create an environment

The best way to use these models is to create an environment in conda with python 3.6:

conda create -n ler python=3.6

To activate an environment:

conda activate ler

Install requirements

pip install -r requirements.txt

Install tensorflow from wheel. I used the following version for Windows and python 3.6.

wget https://raw.githubusercontent.com/fo40225/tensorflow-windows-wheel/master/1.8.0/py36/CPU/sse2/tensorflow-1.8.0-cp36-cp36m-win_amd64.whl
pip install tensorflow-1.8.0-cp36-cp36m-win_amd64.whl

Download dataset

Download the German LER Dataset splits (train, dev, test) from GitHub and save it in the data folder.

wget https://raw.githubusercontent.com/elenanereiss/Legal-Entity-Recognition/master/data/ler_train.conll -P data
wget https://raw.githubusercontent.com/elenanereiss/Legal-Entity-Recognition/master/data/ler_dev.conll -P data
wget https://raw.githubusercontent.com/elenanereiss/Legal-Entity-Recognition/master/data/ler_test.conll  -P data

Train

It is possible to use three models for training:

• BiLSTM-CRF: modelName=blstm-crf;
• BiLSTM-CRF with character embeddings from BiLSTM: modelName=char-blstm-crf;
• BiLSTM-CNN-CRF with character embeddings from CNN: modelName=blstm-cnn-crf.

I want to use char-blstm-crf because that model gives the best results.

python3 train.py char-blstm-crf data/ler_train.conll data/ler_dev.conll data/ler_test.conll

Evaluation

To evaluate the stored model in models/char-blstm-crf.h5, we first need preditions from the test split ler_test.conll. The predictions are written in a file ler_test_pred.conll. After that we can get classification report on entity basis of gold labels and predictions.

python3 predict.py models/char-blstm-crf.h5 data/ler_test.conll data/ler_test_pred.conll
python3 evaluate.py data/ler_test.conll data/ler_test_pred.conll