Skip to content

Latest commit

 

History

History
116 lines (82 loc) · 7.16 KB

readme.md

File metadata and controls

116 lines (82 loc) · 7.16 KB
Raw

For Using Syntactic Supervision for Cross-lingual SRL

Hi, this describes our implementation for our EMNLP-21 paper: "On the Benefit of Syntactic Supervision for Cross-lingual Transfer in Semantic Role Labeling".

Please refer to the paper for more details: [paper] [bib]

Repo

When we were carrying out our experiments for this work, we used the repo at this commit here. In later versions of this repo, there may be slight changes (for example, default hyper-parameter change or hyper-parameter name change).

Clone this repo:

git clone https://github.com/zzsfornlp/zmsp/ src

Environment

Prepare the environment using conda:

conda create -n p20 python=3.8
conda activate p20
conda install pytorch=1.5.0 -c pytorch
conda install numpy scipy cython pybind11 pandas pip
pip install transformers==3.1.0

Please refer to conda_p20.yml for our specific conda environment and you can also specify conda environment using this.

Data Preparation

There are three extra steps (extra data) to prepare before the final preparation:

  • For PATB_INT: follow UD_Arabic-NYUAD to prepare PATB data.
  • For EWT_FOLDER: use this script prepP.sh to prepare EWT PropBank dataset.
  • For CONLL12D: Please first use srl_span/prep.sh to prepare OntoNotes data for further use here.
  • For CONLL09, PTB3, CTB6, they should point to the corresponding LDC's dataset folders.
  • In addition, udapy is also needed, please install it following here.
  • In addition, stanza is also needed, please install it following here. And further download stanza models for [en,zh,ar].

Then, please use prep.sh to prepare data, refer to it for more details.

For example, run it with:

export PYTHONPATH=/your/path/to/src
export PATB_INT=/your/path/to/PATB/integrated
export EWT_FOLDER=/your/path/to/processed_EWT 
export CONLL12D=/your/path/to/data/pb/conll12d 
export CONLL09=/your/path/to/CONLL09 
export PTB3=/your/path/to/treebank_3 
export CTB6=/your/path/to/ctb_v6 
bash prep.sh

After all of these (probably takes several hours), you can get the data files prepared in data/ud/cl{0,1,2,3}/*, corresponding to the data for different experiments. (cl0=EWT/UPB, cl1=EWT/FiPB, cl2=OntoNotes, cl3=CONLL09)

Running

Please use go.py for running:

# create a new dir
mkdir run; cd run;
# ln data
ln -s /your/path/to/prepared//data/ud/ ./ud/
# copy go.py & examples.py
cp /your/path/to/src//msp2/docs/srl_cl/{examples.py,go.py} .
# use the example confs to train/test, for example, here we run the `cl0_base_xlmr` instance with GPU=0.
python3 -m pdb go.py settings:clsrl0 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl0_base_xlmr
  • Notice that in the above example running of cl0_base_xlmr, the program will run in the background in a new sub-folder called run_cl0_base_xlmr_0. By default, it will not print to the stdout, but we can check run_cl0_base_xlmr_0/_log_* for the loggings.
  • Notice that we need different settings:?? for running different experiments(datasets), corresponding to the data folder names, they are: clsrl0,clsrl1,clsrl2,clsrl3. And for clsrl2 and clsrl3, we need an additional one cl2_lang:?? or cl3_lang:?? to specify the target language.

Here are more runnings and explanations for the confs with the example file:

# `cl0_base_xlmr`: 
# Running the UPB zero-resource experiment (cl0) with the baseline model trained on English-SRL and directly tested on all languages.
python3 -m pdb go.py settings:clsrl0 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl0_base_xlmr

# `cl0_syn_xlmr`: 
# Running the UPB zero-resource experiment (cl0) with the baseline model trained on English-SRL and as well as syntactic trees on all languages.
python3 -m pdb go.py settings:clsrl0 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl0_syn_xlmr

# `cl1_both`: 
# Running the EWT/FiPB experiment (cl1) with English + 1k Finnish-SRL.
python3 -m pdb go.py settings:clsrl1 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl1_both

# `cl1_syn`: 
# Running the EWT/FiPB experiment (cl1) with English + 1k Finnish-SRL, as well as syntactic trees.
python3 -m pdb go.py settings:clsrl1 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl1_syn

# `cl2_zh_both`: 
# Running the OntoNotes experiment (cl2) with English + 1k Chinese SRL.
python3 -m pdb go.py settings:clsrl2 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl2_zh_both cl2_lang:zh

# `cl2_zh_syn`: 
# Running the OntoNotes experiment (cl2) with English + 1k Chinese SRL, as well as syntactic trees.
python3 -m pdb go.py settings:clsrl2 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl2_zh_syn cl2_lang:zh

# `cl3_es_both`: 
# Running the CoNLL09 experiment (cl3) with English + 1k Spanish SRL.
python3 -m pdb go.py settings:clsrl3 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl3_es_both cl3_lang:es

# `cl3_es_syn`: 
# Running the CoNLL09 experiment (cl3) with English + 1k Spanish SRL, as well as syntactic trees.
python3 -m pdb go.py settings:clsrl3 tune_table_file:examples.py shuffle:0 gpus:0 tune_name:cl3_es_syn cl3_lang:es

Alternatively, if not using the configuration files, we can also concat all the configuration strings by using train_extras:... or test_extras:..., for example, still running the cl0_base_xlmr example. In this case, the running dir will directly be the current one and the loggings will be printed to stderr:

# directly run one
python3 go.py settings:clsrl0 gpus:0 'train_extras:bert_model:xlm-roberta-base reg0.mod_trg:Menc.bert reg0.reg_method:update reg0.l2_reg:1'

Please refer to the running and example files for more details, here are some brief explanations:

  • We have a msp2.tasks.zmtl2.drive.data_center.DataCenter which contains various groups of train/dev/test datasets.
  • Each group, train*/dev*/test*, will be configurated by a msp2.tasks.zmtl2.core.data.ZDatasetConf item, whose group_files indicate the input file paths of this dataset group.
  • For example, when specifying train1.group_files:en.json,zh.json, we assign these two data files to this group.
  • Each group also have: group_tasks which specifies the tasks (we use udep for dependency parsing and pb1 for propbank SRL parsing) to be performed for this group of datasets, input_folder which is the common folder of the files in this group (for conveniency), presample denotes how many instances we want to down-sample for a dataset.
  • By default, each group and each dataset inside a group will be sampled uniformly, the ratio can be changed with group_sample_rate and group_sample_alpha.
  • There are other configurations for the datasets, please refer to the class ZDatasetConf for more details.
  • Generally, by specifying the input files and tasks of each ZDataset group, we can control what to train and test.