The repository contains the source code of the ACL 2023 long paper "From the One, Judge of the Whole: Typed Entailment Graph Construction with Predicate Generation". Arxiv
- Download the evaluation repository and the dataset for later usage:
git clone https://github.com/mjhosseini/entgraph_eval.git
cd TP-EGG/
wget https://dl.dropboxusercontent.com/s/j7sgqhp8a27qgcf/gfiles.zip
unzip gfiles.zip
rm gfiles.zip
mv gfiles/ ..
ls ../gfiles/ent/
ls ../entgraph_eval/
- Run the training script to get the predicate generator
$G$ based on T5-large, or directly download the checkpoint:
python t5_train.py --t5_size large --lr 1e-3
stat t5_tuned_large_0.001_reannofix.pth
- Run the predicate generator to generate novel predicate sets:
python t5_predgen.py --beam 50 --keep 5000 --filter_level 0 --t5_size large --model_path t5_tuned_large_0.001_reannofix.pth
- Run the training script to get the edge selictor
$M$ based on BERT, or directly download the checkpoint:
mkdir sent_matchers
python -u sent_matcher_ball.py --lr 1e-5 --dnet_lr 5e-4 --n_epoch 300 --pos_repeat 5 --alltest 1 --dist_last exp --d_middim 4 --embmod_dim 16
stat sent_matchers/ball2_alltest_bertbase_0.9_1e-05_0.0005_5_exp_4_em16_best.pth.tar
- Download the edge weight calculator
$W$ checkpoint re-implemented based on EGT2:
mkdir deberta_tars
mv deberta0.8_12_1e-05_1_reannofix_best.pth.tar deberta_tars/
stat deberta_tars/deberta0.8_12_1e-05_1_reannofix_best.pth.tar
- Run the local graph generation script to get the output EG, and you can change the parameters of Line 32-43 in
sent_matcher_modifier_ball_for_t5predgen.py:
python sent_matcher_modifier_ball_for_t5predgen.py
ls ../gfiles/typedEntGrDir_sent_matcher_t5gen_b50k5000fil02nei_large_tunedlarge0.001reannofix_ball2m_1e-5_5e-4_5_exp_4_em16_f2e7_mNone_ty2/
And we upload the final graphs for future research. After downloading corresponding files, get the graphs by:
cat TPEGG_LH_split.z* > graph.zip
unzip graph.zip
- Copy the relaxed evaluation script to the evaluation repo, and get the evaluation results:
cp eval_with_sent.py ../entgraph_eval/evaluation/
cd ../entgraph_eval/evaluation/
python eval.py --gpath typedEntGrDir_sent_matcher_t5gen_b50k5000fil02nei_large_tunedlarge0.001reannofix_ball2m_1e-5_5e-4_5_exp_4_em16_f2e7_mNone_ty2 --test_reanno --sim_suffix _sim.txt --method TPEGG_LH_reannofix --CCG 1 --typed 1 --supervised 0 --oneFeat 1 --useSims 0 --featIdx 0 --exactType --backupAvg --write
cd ../../TP-EGG
python eval_curvefill.py