Neural Mention Detection

Introduction

This repository contains code introduced in the following paper:

Neural Mention Detection
Juntao Yu, Bernd Bohnet and Massimo Poesio
In Proceedings of the 12th Language Resources and Evaluation Conference (LREC), 2020

Setup Environments

• The code is written in Python 2, the compatibility to Python 3 is not guaranteed.
• Before starting, you need to install all the required packages listed in the requirment.txt using pip install -r requirements.txt.
• After that run setup.sh to compile the Tensorflow custom kernels.
• If you want to use Lee MD you need to uncomment the first few lines of setup.sh to download the GloVe embeddings that required by the system.

To use a pre-trained model

• Pre-trained models can be download from this link. We provide two pre-trained models trained with our best model (Biaffine MD):

  • One trained on the CoNLL 2012 shared task data in which singletons and non-referring expressions are not annotated.
  • The other trained on the CRAC 2018 shared task data that has both single mentions and the non-referring expressions annotated.

• Choose the model you want to use and put the model.max.ckpt.* files under the logs/biaffinmd folder.

• Modifiy the test_path accordingly:

  • the test_path is the path to .jsonlines file, each line of the .jsonlines file must in the following format:

  {
  "clusters": [[[0,0],[5,5]],[[2,3],[7,8]], 
  "doc_key": "nw",
  "sentences": [["John", "has", "a", "car", "."], ["He", "washed", "the", "car", "yesteday","."],["Really","?","it", "was", "raining","yesteday","!"]],
  
  }
  

  • If you only have mentions annotated, but not the coreference clusters, then you can simply give every mention a cluster, the reason we use clusters instead of mentions is to allow the same data also be used by our coreference resolution system.

• The model has two output mode high-recall and high-f1 which can be configured in the experiments.conf

• Then use python evaluate.py config_name to start your evaluation

To train your own model

• If you plan to use LEE MD or Biaffine MD, you need to run the python cache_elmo.py train.jsonlines dev.jsonlines to store the ELMo embeddings in the disk, this will speed up your training a lot.
• If you plan to use LEE MD you will additionally need to create the character vocabulary by using python get_char_vocab.py train.jsonlines dev.jsonlines
• If you plan to use Bert MD we would suggest you train on the CPU instead unless you have a very small data set, see BERT page for more information about the GPU memory issue.
• Finally you can start training by using python train.py config_name

Training speed

Both LEE MD and Biaffine MD can be trained in just a few (4-6) hours on a GTX 1080Ti GPU. The Bert MD however takes about 1 week to finish on 48 CPU cores.