Transferability of Syntax-Aware Graph Neural Networks in Zero-Shot Cross-Lingual Semantic Role Labeling

About

This is the implementation code for our paper: "Transferability of Syntax-Aware Graph Neural Networks in Zero-Shot Cross-Lingual Semantic Role Labeling". Feel free to use this repository to reproduce our results for research purposes. We do not provide any data in this repository, therefore please prepare the data in advance. We refer to the other repositories when implementing the models for our experiments:

1. diegma/neural-dep-srl for Syntactic Graph Convolutional Networks (SGCNs)
2. dmlc/dgl for Relational Graph Convolutional Networks (RGCNs)
3. AnWang-AI/towe-eacl for Attention-Based Graph Convolutional Networks (ARGCNs)
4. gordicaleksa/pytorch-GAT for Graph Attention Networks (GATs)
5. shenwzh3/RGAT-ABSA for Relational Graph Attention Networks (RGATs)
6. thudm/hgb for Simple Heterogeneous Graph Neural Networks (SHGNs)
7. deepakn97/relationPrediction for Knowledge-Based Graph Attention Networks (KBGATs)
8. wasiahmad/GATE for Graph Attention Transformer Encoders (GATEs)

Requirements

We provide the versions of libraries that we use:

1. huggingface-hub: 0.15.1
2. matplotlib: 3.7.1
3. numpy: 1.24.3
4. pandas: 2.0.2
5. prettytable: 3.7.0
6. python: 3.8.16
7. scipy: 1.10.1
8. seaborn: 0.12.2
9. tokenizers: 0.13.3
10. torch: 1.13.0+cu116
11. torch-scatter: 2.1.1+pt113cu116
12. tqdm: 4.65.0
13. transformers: 4.29.2

Corpus Preprocessing

1. Clone SRL annotations for target languages from Universal Proposition Bank (UPB) v2 and for English.
2. Run helper_scripts/fix_upb_2.py to fix shifted annotation problem caused by enhanced dependency tree annotations in treebanks. Make sure the paths written in the Python script are correct. The treebanks that contain enhanced dependency tree annotations are:
   1. Czech-CAC
   2. Czech-FicTree
   3. Czech-PDT
   4. Dutch-Alpino
   5. Dutch-LassySmall
   6. Finnish-TDT
   7. Italian-ISDT
   8. Spanish-AnCora
   9. Ukrainian-IU
3. Download annotations from Universal Dependencies (UD) v2.9.
4. Clone UPB tools to merge UPB annotations with UD annotations.
5. Setup UPB tools.
6. Run this command from UPB tools to merge UPB annotations and UD annotations and obtain complete SRL annotations for a certain treebank:

python3 up2/merge_ud_up.py --input_ud=<ud-treebank> --input_up=<up-treebank> --output=<merged-treebank>

6. Change the paths mentioned at the metadata_by_version_to_lang_to_treebank inside constants/dataset.py file to point to the correct paths. UP-2.0 must point to the complete SRL annotations using gold POS tags and dependency trees. UP-2.0-predicted must point to complete SRL annotations using predicted POS tags and dependency trees.

Traning

1. We provide a Python script in helper_scripts/create_script.py to generate a script to run the training and evaluation for each model. The codenames for models are as follows:
   1. gcn_syn for SGCNs
   2. gcn_ar for ARGCNs
   3. gat_plain for GATs
   4. lstm for BiLSTMs
   5. gat_het for SHGNs
   6. gat_two_att for RGATs
   7. gat_kb for KBGATs
   8. gcn_r for RGCNs
   9. trans_rpr for Self-Attention with Relative Position Representations (SAN-RPRs)
   10. gate for GATEs
   11. trans for Transformers
   12. trans_spr for Self-Attention with Structural Absolute Position Representations (SAN-SAPRs)
   13. trans_spr_rel for Self-Attention with Structural Relative Position Representations (SAN-SRPRs)
   14. trans_x_spr_rel for Transformers with SRPRs (Trans-SRPRs)
   15. trans_rpr_x_spr for SAN-SAPRs with SAN-RPRs (SAPR-RPRs)
   16. trans_x_spr_rel_x_dr for Trans-SRPRs with DR (Trans-SRPR-DRs)
   17. trans_rpr_x_spr_x_dr for SAPR-RPRs with DR (SAPR-RPR-DRs)
   18. trans_rpr_x_spr_x_ldp for SAPR-RPRs with LDPs (SAPR-RPR-LDPs)
   19. trans_rpr_x_spr_x_sdp for SAPR-RPRs with SDPs (SAPR-RPR-SDPs)
2. Fill the codename of the desirable model in model_list variable to generate the script for training a certain model.
3. Fill the desirable parameters for the model in params variable. The explanation for each parameter can be found in the Appendix of the paper.
4. Run the Python script in helper_scripts/create_script.py. Make sure the paths written in the Python script are correct.
5. The script to train and evaluate the model will be available at scripts/tests.
6. Run the script with bash <script-name> to start the training process.
7. After the training and evaluation finish, the logs and models will be available at directories with _logs and _models suffixes.

Licence