Awesome

Higher-order Coreference Resolution with Coarse-to-fine Inference for the german language

Introduction

This repository contains the german adaption from

• Higher-order Coreference Resolution with Coarse-to-fine Inference
• Kenton Lee, Luheng He, and Luke Zettlemoyer
• In NAACL 2018

Getting Started

• Install python3 requirements: pip3 install -r requirements.txt
• Download pretrained model at https://drive.google.com/file/d/1L-kKxzlC0pPr_tJzRyi9xoTOKSPQXfNb/view?usp=sharing
  • Create a 'logs' folder in the root of the repository and extract the model with the folder 'final' in it.
• Download GloVe embeddings and build custom kernels by running setup_all.sh.
  • https://drive.google.com/file/d/1nN_qc3qHtPecxek0LsYf544ipJpfXEfj/view?usp=sharing
• To train your own models, elmo embeddings need to be created and converted to the right format.
  • This assumes access to TüBa-D/Z dataset.
  • Use https://github.com/ChristianHesels/bilm-tf-e2e-german to create embeddings

Training Instructions

• Experiment configurations are found in experiments.conf
• Choose an experiment that you would like to run, e.g. best
• Training: python3 train.py <experiment>
• Results are stored in the logs directory and can be viewed via TensorBoard.
• Evaluation: python3 evaluate.py <experiment>

Other Quirks

• It does not use GPUs by default. Instead, it looks for the GPU environment variable, which the code treats as shorthand for CUDA_VISIBLE_DEVICES.
• The training runs indefinitely and needs to be terminated manually. The model generally converges at about 400k steps.