Code for our COLING-2020 paper,

Semantic Role Labeling with Heterogeneous Syntactic Knowledge

Qingrong Xia, Rui Wang, Zhenghua Li, Yue Zhang, and Min Zhang

Python2, Pytorch0.4.1

We conduct experiments on both Chinese and English SRL datasets, i.e, CPB1.0 and CoNLL-2005. Please follow Chinese-SRL for the data prepration of CPB1.0 and lsgn for the data prepration of CoNLL-2005.

For Chinese, we use the PCTB and CDT as the dependency treebanks. For English, we use the PTB and UD as the dependency treebanks. As described in our paper, we employ the biaffine parser to obtain the automatic dependency trees of the SRL data. Since only the UD treebank is publicly available, we give an example as follows: The last column is the dependency edge probabilities. Besides, we upload the CPB1.0 model (without RoBERTa) in url

Preparing all the data, then you can train our model with the command:

bash train.sh GPU_ID

For evaluation, the bash command is:

bash predict.sh GPU_ID