scofield7419 / LAGCN-SRL

Codes for the AAAI 2021 paper: Encoder-Decoder Based Unified Semantic Role Labeling with Label-Aware Syntax

LAGCN SRL Pointer

This repository includes the code of the Semantic Role Labeling (SRL) Parser with label-aware graph convolutional network (LAGCN) based pointer networks of the AAAI 2021 paper: Encoder-Decoder Based Unified Semantic Role Labeling with Label-Aware Syntax.

---

Requirement Install

pip install -r requirements.txt

Datasets

Two popular dependency-based SRL datasets.

Download them and put at ./data folds.

• CoNLL09
• UPB

Syntax annotation parsing

To prepare the syntactic dependency features, deploy the CoreNLP:

wget https://nlp.stanford.edu/software/stanford-corenlp-latest.zip

nohup java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 8083 -timeout 15000 > 1.log 2>&1 &

See the example data in ./data/demo.

Experiments

Step 1. To train the parser, you need to include the pre-trained word embeddings in the embs folder and run the following script:

./scripts/run_parser.sh <model> <data>

To evaluate the best trained model on the test set, just use the official script to compute the F1 scores:

./scripts/eval.sh <best epoch> <data> <model>

Citation

@inproceedings{FeiGraphSynAAAI21,
  author    = {Hao Fei and Fei Li and Bobo Li and Donghong Ji},
  title     = {Encoder-Decoder Based Unified Semantic Role Labeling with Label-Aware Syntax},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  pages     = {12794--12802},
  year      = {2021},
}

License

The code is released under Apache License 2.0 for Noncommercial use only.