Knowledge-aware Attention Network for Protein-Protein Interaction Extraction Zhou, H., Liu Z., Ning S. et al. Accepted by The Journal of Biomedical informatics

An implementation of Knowledge-aware attention networks (KAN) for protein-protein extraction task.

This code has been written using Pytorch 0.4.

We have put word embeddings, entity and relation embeddings learned from TransE[1] to the folder of data.

Go to the model path and run:

❱❱❱ python3 main.py

In this setting, the default hyperparameters are used. Or run in specific settings:

❱❱❱ python3 main.py --trainPath ../data/train.txt --validPath ../data/valid.txt --testPath ../data/test.txt --trainGold ../data/trainGold.txt --testGold ../data/testGold.txt --batchSize 100 --wd 100 --ed 100 --hop 2 --clas 2 --epoch 20 --wePath ../data/wordEmb/bio-word2id100 --w2IDPath ../data/wordEmb/bio-embed100 --eePath ../data/KB/entity2vec.vec --rePath ../data/KB/relation2vec.vec --t2idPath ../data/KB/triple2id.txt --e2idPath ../data/KB/entity2id.txt --paraPath ./parameters/kan --results ./results/ --training True --lr 0.1 --wdecay 0.0 --validsetR 0.15

the option you can choose are:

• --trainPath path of training dataset.
• --validPath path of valid dataset.
• --testPath path of test dataset.
• --trainGold path of triples of training dataset
• --testGold path of triples of test dataset
• --batchSize batch size.
• --wd dimension of word embedding.
• --ed dimension of entity embedding learned from TransE.
• --hop number of hop.
• --clas number of class.
• --epoch number of iterations.
• --wePath path of word embedding file.
• --w2IDPath path of file that contains mapping from word to its number.
• --eePath path of entity embedding file.
• --rePath path of relation embedding file.
• --t2idPath path of file that contains the triples.
• --e2idPath path of file that contains mapping from Entrez Gene ID to number.
• --paraPath path of model parameters.
• --results path where the results write to.
• --training bool value for training or not. A non-empty string means training phase. An empty string means test phase.
• --lr learning rate.
• --wdecay weight decay.
• --validsetR this parameter ([0.0, 1.0]) means how much document in training set is selected to be the valid set.

[1] Bordes, Antoine, et al. Translating embeddings for modeling multi-relational data. Proceedings of NIPS, 2013.