DCA
DCA (Dynamic Context Augmentation) provides global entity linking models featuring:
-
Efficiency: Comparing to global entity linking models, DCA only requires one pass through all mentions, yielding better efficiency in inference.
-
Portability: DCA can introduce topical coherence into local linking models without reshaping their original designs or structures.
Remarkablely, our DCA models (trained by supervised learning or reinforcement learning) achieved:
- 94.64% in-KB acc. on AIDA-CONLL testset (AIDA-B).
- 94.57% F1 score on MSBNC dataset and 90.14% F1 score on ACE2004 dataset.
Details about DCA can be accessed at: https://arxiv.org/abs/1909.02117.
This implementation refers to the project structure of mulrel-nel.
Written and maintained by Sheng Lin ([email protected]) and Xiyuan Yang ([email protected]).
Overall Workflow
Data
Download data from here and unzip to the main folder (i.e. your-path/DCA).
The above data archive mainly contains the following resource files:
-
Dataset: One in-domain dataset (AIDA-CoNLL) and Five cross-domain datasets (MSNBC / AQUAINT / ACE2004 / CWEB / WIKI). And these datasets share the same data format.
-
Mention Type: Adopted to compute type similarity between mention-entity pairs. We predict types for each mention in datasets using a typing system called NFETC model trained by the AIDA dataset.
-
Wikipedia inLinks: Surface names of inlinks for a Wikipedia page (entity) are used to construct dynamic context in our model learning process.
-
Entity Description: Wikipedia page contents (entity description) are used by one of our base model -- Berkeley-CNN
Installation
Requirements: Python 3.5 or 3.6, Pytorch 0.3, CUDA 7.5 or 8
Important Parameters
mode: train or eval mode.
method: training method, Supervised Learning (SL) or Reinforcement Learning (RL)
order: three decision orders -- offset / size / random. Please refer to our paper for their concrete definition.
n_cands_before_rank: the number of candidates.
tok_top_n4inlink: the number of inlinks for a Wikipedia page (entity) would be considered as candidates for the dynamic context.
tok_top_n4ent: the number of inlinks for a Wikipedia page (entity) would be added into the dynamic context.
isDynamic: 2-hop DCA / 1-hop DCA / without DCA. Corresponding to the experiments of Table 4 in our paper.
dca_method: soft+hard attention / soft attention / average sum. Corresponding to the experiments of Table 5 in our paper.
Running
cd DCA/
export PYTHONPATH=$PYTHONPATH:../
Supervised Learning: python main.py --mode train --order offset --model_path model --method SL
Reinforcement Learning: python main.py --mode train --order offset --model_path model --method RL
Citation
If you find the implementation useful, please cite the following paper: Learning Dynamic Context Augmentation for Global Entity Linking.
@inproceedings{yang2019learning,
title={Learning Dynamic Context Augmentation for Global Entity Linking},
author={Yang, Xiyuan and Gu, Xiaotao and Lin, Sheng and Tang, Siliang and Zhuang, Yueting and Wu, Fei and Chen, Zhigang and Hu, Guoping and Ren, Xiang},
booktitle = {Proceedings of EMNLP-IJCNLP},
year={2019}
}