XREF

This repository contains code and links to data for our AKBC 2020 paper:

XREF: Entity Linking for Chinese News Comments with Supplementary Article Reference
Xinyu Hua, Lei Li, Lifeng Hua, and Lu Wang. AKBC 2020

Data

link

Environment

python 3.7
pytorch 1.5
pytorch-lightning 0.8

Training

The following script will train XRef model for 10 epochs with batch size 32. It will dump checkpoints every 1 epoch, with the top 5 ones (base on validation loss) saved. (we assume the data is stored under ./data/.)

python train.py --domain=[ent,product] \
    --ckpt-dir=model/demo/ \
    --batch-size=32 \
    --model=xref \
    --tensorboard-dir=demo \
    --save-interval=1 \
    --save-topk=5 \
    --max-train-epochs=10

Inference

The following script loads the latest checkpoint from --ckpt-dir and run inference over the test set. The results will be saved to --output-path.

python infer.py --domain=[ent,product] \
    --output-path=demo.jsonl \
    --ckpt-dir=model/demo/ \

License