If you find our paper, code or framework useful, please reference this work in your paper:

@inproceedings{blloshmi-etal-2021-generating,
  title     = {Generating Senses and RoLes: An End-to-End Model for Dependency- and Span-based Semantic Role Labeling},
  author    = {Blloshmi, Rexhina and Conia, Simone and Tripodi, Rocco and Navigli, Roberto},
  booktitle = {Proceedings of the Thirtieth International Joint Conference on
               Artificial Intelligence, {IJCAI-21}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  pages     = {3786--3793},
  year      = {2021},
  doi       = {10.24963/ijcai.2021/521},
  url       = {https://doi.org/10.24963/ijcai.2021/521},
}

Pretrained Checkpoints

CoNLL-2009

Paper experiments:

Extra experiments:

CoNLL-2012

Paper experiments:

Extra experiments:

Evaluation Framework

• Please contact us by email.

Setup

Install

Create a conda environment with Python 3.8 and PyTorch 1.5.0 and install the dependencies requirements.txt.

Via conda:

conda create -n gsrl python=3.8
conda activate gsrl
bash ./download_artifacts.sh

To enable wandb logging**:

wandb login

**Also set log_wandb to True (currenly False) in configs files and fill in wandb-project and team information accordingly.

Add the CoNLL-2009 and CoNLL-2012 datasets inside data/ directory.

Modify the data paths in the configuration files in configs/ or follow our file structure.

E.g., the folder structure for CoNLL-2009 should look as below:

(gsrl)$ tree data/conll-2009 -L 2 data/conll-2009

conll-2009
    └── en
        │ ── dev
        │   └──CoNLL2009_dev.txt
        │── ood
        │   └── CoNLL2009_test_ood.txt
        ├── test
        │   └── CoNLL2009_test.txt
        └── training
            └── CoNLL2009_train.txt

Training & Evaluation

• All configuration and parameters to reproduce our main results are included in configs/ directory.

• Logs of wandb and model checkpoints are saved in runs/.

• Evaluation scripts are in scripts/ and their output is saves in out/.

• Vocabulary additions are included in data/vocab/. To allow reproducability do not change the files.

Span-based Semantic Role Labeling

1. To train a GSRL model with nested linearization:

python -m src.bin.train 
        --config configs/config-span-srl.yaml 
        --task-type span

Evaluate the model using the following command:

python -m src.bin.predict_srl 
       --datasets data/conll-2012/en/test/CoNLL2012_test.txt 
       --checkpoint runs/[checkpoint_name_here] 
       --task-type span 
       --beam-size 1 
       --eval-name nested-span-srl-result

2. To train a GSRL model with flattened linearization:

python -m src.bin.train 
       --config configs/config-span-srl.yaml 
       --task-type span 
       --duplicate-per-predicate

Evaluate the model using the following command:

python -m src.bin.predict_srl 
       --datasets data/conll-2012/en/test/CoNLL2012_test.txt 
       --checkpoint runs/[checkpoint_name_here] 
       --task-type span 
       --beam-size 1 
       --duplicate-per-predicate 
       --eval-name flattened-span-srl-result

Dependency-based Semantic Role Labeling

2. To train a GSRL model with nested linearization:

python -m src.bin.train 
       --config configs/config-dep-srl.yaml 
       --task-type dep

Evaluate the model using the following command:

python -m src.bin.predict_srl 
       --datasets data/conll-2009/en/test/CoNLL2009_test.txt 
       --checkpoint runs/[checkpoint_name_here] 
       --task-type dep 
       --beam-size 1 
       --eval-name nested-dep-srl-result

2. To train a GSRL model with flattened linearization:

python -m src.bin.train 
      --config configs/config-dep-srl.yaml 
      --task-type dep 
      --duplicate-per-predicate

Evaluate the model using the following command:

python -m src.bin.predict_srl 
       --datasets data/conll-2009/en/test/CoNLL2009_test.txt 
       --checkpoint runs/[checkpoint_name_here] 
       --task-type dep 
       --beam-size 1 
       --duplicate-per-predicate 
       --eval-name flattened-dep-srl-result

Extra

• To run without predicate identifiers in input, add --identify-predicate in both training and evaluation scripts above.

License

This project is released under the CC-BY-NC-SA 4.0 license (see LICENSE). If you use GSRL, please put a link to this repo.

Acknowledgements

The authors gratefully acknowledge the support of the ERC Consolidator Grant MOUSSE No. 726487 and the ELEXIS project No. 731015 under the European Union’s Horizon 2020 research and innovation programme.

This work was supported in part by the MIUR under the grant "Dipartimenti di eccellenza 2018-2022" of the Department of Computer Science of the Sapienza University of Rome.

We adopted modules or code snippets from the open-source projects:

• SPRING