RHIA-EOP

A pytorch implementation of the paper: "Distantly Supervised Relation Extraction via Recursive Hierarchy-Interactive Attention and Entity-Order Perception".

Requirements

The model is implemented using Pytorch. The versions of packages used are shown below.

• python==3.7.9
• pytorch==1.7.1
• cuda == 10.1
• numpy==1.19.2
• tqdm==4.55.1
• scikit_learn==0.23.2

Data

Unzip the data ./raw_data/data.zipunder ./data/

unzip ./raw_data/data.zip -d ./data/

Use pretrained model

The pretrained model is already saved at ./result/checkpoint/. To directly evaluate on it, run the following command:

python evaluate.py --model RHIA-EOP --processed_data_dir ./_processed_data/RHIA-EOP/ --save_dir result/checkpoint/ --ent_order eop --hier_rel_net rhia --pone --ptwo --pall

P-R curve, the curve will be saved at ./result/checkpoint/:

python show_pr.py RHIA-EOP

Train the model

python main.py --model RHIA-EOP --processed_data_dir ./_processed_data/RHIA-EOP/ --save_dir ./result/RHIA-EOP/ --ent_order eop --hier_rel_net rhia

Evaluate the model

python evaluate.py --model RHIA-EOP --processed_data_dir ./_processed_data/RHIA-EOP/ --save_dir ./result/RHIA-EOP/ --ent_order eop --hier_rel_net rhia --pone --ptwo --pall

Experiment results

P-R curve

AUC & Max_F1

AUC	Max_F1
0.56	0.546

P@N (%)

P@100	P@200	P@300	P@500	P@1000	P@2000	Mean
95.0	94.0	89.7	85.2	71.7	53.2	81.5

P@N (Use the setting of Li et al.)

• pone

P@100	P@200	P@300	Mean
96.0	92.5	86.7	91.7

• ptwo

P@100	P@200	P@300	Mean
98.0	95.5	92.3	95.3

• pall

P@100	P@200	P@300	Mean
98.0	96.5	93.3	95.9

Hits@K (macro)

< 100	10	15	20
RHIA-EOP	66.7	83.3	94.4

< 200	10	15	20
RHIA-EOP	72.7	86.4	95.5

Acknowledgements

We thank a lot for the following repos:

• thunlp/HNRE
• tmliang/SeG