We use the cross-sentence n-ary relation extraction dataset provided by Peng et al. It is a biomedical dataset from the PubMed literature database that contains three types of entities: "drug", "gene", and "mutation", including 6987 instances of "drug-gene-mutation" ternary relationships and 6087 instances of "mutation-drug" binary relationships. You can get the original dataset here: https://github.com/freesunshine0316/nary-grn.
Please modify the following parameters in the preprocessing.py file according to the original data storage path.
self.data_path_list = ['origin_data/drug_gene_var/0/data_graph_1', 'origin_data/drug_gene_var/0/data_graph_2',
'origin_data/drug_gene_var/1/data_graph_1', 'origin_data/drug_gene_var/1/data_graph_2',
'origin_data/drug_gene_var/2/data_graph_1', 'origin_data/drug_gene_var/2/data_graph_2',
'origin_data/drug_gene_var/3/data_graph_1', 'origin_data/drug_gene_var/3/data_graph_2',
'origin_data/drug_gene_var/4/data_graph_1', 'origin_data/drug_gene_var/4/data_graph_2']
self.vec_path = 'vec/vocab_morph.wordvec.st'
After the program is executed, the following data will be stored in the data/ folder.
--edge_id2words.json
--edge_word_vec.npy
--edge_word2ids.json
--id2words.json
--word2ids.json
--wordvec.npy
--data(1)
--test_entity_index.npy
--test_label_ids.npy
--test_sentence_ids.npy
--test_y.npy
--train_entity_index.npy
--train_label_ids.npy
--train_sentence_ids.npy
--train_y.npy
Set class_num (2 or 5, class_num indicates the number of relational categories, i.e., two classification and five classification), execute python run.py
training 0 epoch...
1%|▏ | 72/5219 [00:23<28:04, 3.06it/s]
...
training 150 epoch...
###
test set macro accuracy: 0.9133435582822086
test set macro precision: 0.8902972279845123
test set macro recall: 0.9084080450541909
test set macro f1: 0.8977559548819014
Five-fold cross-validation average accuracy
Drug-Mutation
Two-class: 93.78%
Five-class: 93.45%
---------------------------------
Drug-Gene-Mutation
Two-class: 93.22%
Five-class: 92.88%