This repository is the code for "Fine-grained Entity Typing via Label Reasoning" EMNLP2021
We evaluate the performance on two datasets: OntoNotes and Ultra-Fine from Choi. Considering the size of the data, we only put the Ultra-Fine in the folder data/
| Data | Train | Dev | Test |
|---|---|---|---|
| Ultra-Fine | 1998 | 1998 | 1998 |
| OntoNotes | 251039 | 2202 | 8963 |
| OntoNotes augmentation | 793487 | 2202 | 8963 |
Configure the environment in python=3.7.7：
pip install -r requirement.txtthen download the Glove Vector and put it in the data/, here we use glove.6b.300d.txt
If you want to run the model, download the mode and put it into model/:
bash view_typing.bash- You can train a LRN w/o IR for example:
python -m entity_typing.train seq_typing \
-model-path model/ultra_fine/seq_typing_entity_marker_macro_att1_att2 \
-label-emb-size 100 \
-bert-type entity_marker \
-device 2 \
-batch 32 \
-patience 10 \
-fine-tune-transformer \
-transformer bert-base-uncased \
-lr 1e-3 \
-overwrite-model-path \
-lr-diff \
-edit 0.05 \
-teaching-forcing-rate 0 \
-seq-type sequence \
-shuffle-num 0 \
-loss-type match \
-decoder-type lstm2 \
-evaluation macro \
$*- You can train a LRN for example:
python -m entity_typing.train mem_typing \
-model-path model/ultra_fine/kv_typing_entity_marker_macro_att1_att2 \
-label-emb-size 100 \
-data-folder-path data/open_type_original_lemma_synthetic \
-value-file probe_experiment/data/all_types.txt \
-memory-emb data/glove.6B.300d.txt \
-memory-emb-size 300 \
-bert-type entity_marker_kv \
-device 3 \
-batch 32 \
-patience 10 \
-fine-tune-transformer \
-transformer bert-base-uncased \
-lr 1e-3 \
-overwrite-model-path \
-lr-diff \
-edit 0.05 \
-teaching-forcing-rate 0 \
-seq-type sequence \
-shuffle-num 0 \
-loss-type match \
-decoder-type lstm2 \
-evaluation macro \
-loss-lambda 0 \
-loss-lambda2 1 \
-similar-thred 0.2 \
-sparce-rate 0.9 \
$*If you use other Glove Vector, edit the dimension of -memory-emb-size
If you want to train another model, you can refer to the ./train_*.bash file.
If you want use other attributes, you can add it to data and edit the process_line function in entity_typing/data_utils.py