This is the code for the EACL2021 paper, [Few-shot Semantic Parsing for New Predicates].

• Cuda 10.2
• conda env create -f environment.yml
• conda activate few-shot-semantic-parsing
• pip install transformers

./pull_dependency.sh

Download the pre-trained models from Google Drive, and copy them to the corresponding folders.

• Copy [ATIS pre-trained model for one-shot learning] to saved_models/atis/freq_0/ and unzip it.
• Copy [ATIS pre-trained model for two-shot learning] to saved_models/atis/freq_50/ and unzip it.
• More pre-trained models will be provided if needed.

This is to generate the sequences of actions that could construct the logical forms.

• Atis python preprocess_data/atis/generate_examples.py
• Geo python preprocess_data/geo/generate_examples.py
• Jobs python preprocess_data/jobs/generate_examples.py

You could either download the pre-trained models from the corresponding links or pre-train the parser yourself.

• ATIS one-shot: ./scripts/atis/one_shot/train.sh 0 1 pretrain
• ATIS two-shot: ./scripts/atis/two_shot/train.sh 0 1 pretrain
• Geo one-shot: ./scripts/geo/one_shot/train.sh 0 1 pretrain
• Geo two-shot: ./scripts/geo/two_shot/train.sh 0 1 pretrain
• Jobs one-shot: ./scripts/jobs/one_shot/train.sh 0 1 pretrain
• Jobs two-shot: ./scripts/jobs/two_shot/train.sh 0 1 pretrain

• ATIS one-shot: ./scripts/atis/one_shot/fine_tune.sh saved_models/atis/freq_0/pretrained_model_name.bin [0..4] [1..2]
• ATIS two-shot: ./scripts/atis/two_shot/fine_tune.sh saved_models/atis/freq_50/pretrained_model_name.bin [0..4] [1..2]
• Geo one-shot: ./scripts/geo/one_shot/fine_tune.sh saved_models/geo/freq_0/pretrained_model_name.bin [0..4] [1..2]
• Geo two-shot: ./scripts/geo/two_shot/fine_tune.sh saved_models/geo/freq_50/pretrained_model_name.bin [0..4] [1..2]
• Jobs one-shot: ./scripts/jobs/one_shot/fine_tune.sh saved_models/jobs/freq_0/pretrained_model_name.bin [0..4] [1..2]
• Jobs two-shot: ./scripts/jobs/two_shot/fine_tune.sh saved_models/jobs/freq_0/pretrained_model_name.bin [0..4] [1..2]

If you have any questions, please open an issue on Github or contact me via zhuang.li@monash.edu. I am also happy to collobrate with others on the semantic parsing task based on this work. If you find this code useful, please cite:

@inproceedings{li2021few,
  title={Few-Shot Semantic Parsing for New Predicates},
  author={Li, Zhuang and Qu, Lizhen and Huang, Shuo and Haffari, Gholamreza},
  booktitle={Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume},
  pages={1281--1291},
  year={2021}
}