This repository contains code implementation to train an AMR parser using triple linearization.

Pre-requisites

Clone the repository and install the required packages using the following command:

git clone https://github.com/RikVN/AMR.git
cd AMR
git clone https://github.com/snowblink14/smatch
cd ../
pip install -r requirements.txt

Add followings packages to python path

export PYTHONPATH=$PYTHONPATH:/[project_dir]/AMR
export PYTHONPATH=$PYTHONPATH:/[project_dir]/AMR/smatch

Download the data

• AMR 3.0

  • AMR3.0 is available on LDC. You need a proper license to download the data.

  • Structure of the data

    • Once the data is downloaded, they should be structured in the following way:

      - data
          - AMR
            - en
              - train
                - en-amr.pm      # Structured AMR graph, delimited with a blank line
                - en-amr.en   # Sentences corresponding to the AMR graphs
              - eval
                - en-amr.pm
                - en-amr.en
            ...
            - fr 
              - train
                - fr-amr.pm
                - fr-amr.fr
            ... 
    

Preprocess + linearize data

Preprocess the data using the following command:

cd preprocess
chmod +x preprocess.sh
./preprocess.sh

See the preprocess.sh file for more details to change the data path.

Train the model

Train the model using the following command:

./train.sh

See the train.sh file for more details to change training parameters.