Basic version of FCT model for relation extraction. The package has two executable files: RE_FCT_fixed corresponds to the log-linear model, RE_FCT is the log-quadratic (log-binear) model The data directory
Usage: ./RE_FCT[_fixed] trainfile devfile resfile baseline_embfile num-iter learning-rate
Use the following command to reproduce the results I reported: ./RE_FCT ../data/SemEval.train.fea.sst ../data/SemEval.test.fea.sst predict.fea.fullnerpair.onlyne.txt ../data/vectors.nyt2011.cbow.semeval.filtered 30 0.005 &> training.log &
Sorry that currently early-stopping should be done manually :)
I did not actually tune the learning rate much. You can try a grid search and hopefully better results can be achieved
When running on SemEval, it is better to close the sub-models with sst-pair features, since when WordNet super sense tags are used instead of NE tags, the model will have much more number of entity type pairs and will lead to overfitting.
The word embeddings can be found at http://yumo.asiteof.me/yumo/data/vectors.nyt2011.cbow.bin