Dual-Encoder-Based Zero-Shot Entity Linking

Quick Starts in 3 minutes

git clone https://github.com/izuna385/Zero-Shot-Entity-Linking.git
cd Zero-Shot-Entity-Linking
python -m spacy download en_core_web_sm

# ~ Multiprocessing Sentence Boundary Detection takes about 2 hours under 8 core CPUs.
sh preprocessing.sh
python3 ./src/train.py -num_epochs 1

For further speednizing to check entire script, run the following command.

python3 ./src/train.py -num_epochs 1 -debug True

also, multi-gpu is supported.

CUDA_VISIBLE_DEVICES=0,1 python3 ./src/train.py -num_epochs 1 -cuda_devices 0,1

Descriptions

• This experiments aim to confirm whether fine-tuning pretraind BERT (more specifically, encoders for mention and entity) is effective even to the unknown domains.

  • Following [Logeswaran et al., '19], entities are not shared between train-dev and train-test.

  • If you are interested in what this repository does, see the original paper, or unofficial slides.

    • Original paper

    • Unofficial slides

Requirements

• torch,allennlp,transformers, and faiss are required. See also requirements.txt.

• ~3 GB CPU and ~1.1GB GPU are necessary for running script.

How to run experiments

1. Preprocessing

• Run sh preprocessing.sh at this directory.

  • The Datasets are derived from [Logeswaran et al., '19].

2. Training and Evaluate Bi-Encoder Model

• python3 ./src/train.py

  • This script trains encoder for mention and entity.
  

3. Logging Each Experiment

• See ./src/experiment_logdir/.

  • Log directory is named after when the experiment starts.

TODO

• Preprocess with more strict sentence boundary.

LICENSE

• MIT