MetaKGR

Source codes and datasets for EMNLP 2019 paper Adapting Meta Knowledge Graph Information for Multi-Hop Reasoning over Few-Shot Relations

Requirements

• python3 (tested on 3.6.6)
• pytorch (tested on 0.4.1)

Installation

python3 -m pip install -r requirements.txt

Data Preparation

Unpack the data files

unzip data.zip

and there will be two datasets under folder data.

# dataset FB15K-237
data/FB15K-237

# dataset NELL-995
data/NE

Pretrain Knowledge Graph Embedding

./experiment-emb.sh configs/<dataset>-<model>.sh --train <gpu-ID>

dataset is the name of datasets and model is the name of knowledge graph embedding model. In our experiments, dataset could be fb15k-237 and NE, model could be conve and distmult. <gpu-ID> is a non-negative integer number representing the GPU index.

Meta Learning

# take FB15K-237 for example
./experiment-rs.sh configs/fb15k-237-rs.sh --train <gpu-ID> --few_shot

Fast Adaptation

# take FB15K-237 for example
./experiment-rs.sh configs/fb15k-237-rs.sh --train <gpu-ID> --adaptation --checkpoint_path model/FB15K-237-point.rs.conve-xavier-n/a-200-200-3-0.001-0.3-0.1-0.5-400-0.02/checkpoint-<Epoch>.tar

<Epoch> is a non-negative integer number representing the training epoch for meta-learning. We can assign <Epoch> to be 15 on FB15K-237.

Test

# take FB15K-237 for example
./experiment-rs.sh configs/fb15k-237-rs.sh --inference <gpu-ID> --few_shot --checkpoint_path model/FB15K-237-point.rs.conve-xavier-n/a-200-200-3-0.001-0.3-0.1-0.5-400-0.02/checkpoint-<Epoch_Adapt>-[relation].tar

<Epoch_Adapt> is a non-negative integer number representing the training epoch for fast adaptation. We can assign <Epoch_Adapt> to be 39 on FB15K-237.

Cite

If you use the code, please cite this paper:

Xin Lv, Yuxian Gu, Xu Han, Lei Hou, Juanzi Li, Zhiyuan Liu. Adapting Meta Knowledge Graph Information for Multi-Hop Reasoning over Few-Shot Relations. The Conference on Empirical Methods in Natural Language Processing (EMNLP 2019).