This repo contains Theano implementations of the models described in the following paper:
The model gives about F1 = 78.00 on the dev & test set in the following hyperparameter setting:
- word embedding=SENNA, unit=GRU, hidden dimension=128, L2 regularization=0.0005, layers=4
- CoNLL-2004 Shared Task (http://www.lsi.upc.edu/~srlconll/st04/st04.html)
- CoNLL-2005 Shared Task (http://www.cs.upc.edu/~srlconll/)
- Word Embedding: SENNA (http://ronan.collobert.com/senna/)
python -m srl.cons_srl.main -mode train --train_data path/to/data/conll2005.train.txt --dev_data path/to/data/conll2005.dev.txt --test_data path/to/data/conll2005.test.txt --init_emb path/to/data/senna.txt --unit gru --layer 1 --hidden 32 --reg 0.0001
An DT B-NP (S* - (A0* (A0*
A.P. NNP I-NP * - * *
Green NNP I-NP * - * *
official NN I-NP * - *A0) *A0)
declined VBD B-VP * decline (V*V) *
to TO I-VP * - (A1* *
comment VB I-VP * comment * (V*V)
on IN B-PP * - * *
the DT B-NP * - * (A1*
filing NN I-NP * - *A1) *A1)
. . O *S) - * *
Ms. NNP (S1(S(NP* * - (A0*
Haag NNP *) (LOC*) - *)
plays VBZ (VP* * play (V*)
Elianti NNP (NP*)) * - (A1*)
. . *)) * - *