Better Transition-Based AMR Parsing with a Refined Search Space

DyNet implementation of the transition-based AMR parser. We provide the code for aligner and the parser. You can find all the details in the following paper:

• Better Transition-Based AMR Parsing with a Refined Search Space. (Zhijiang Guo and Wei Lu. EMNLP 2018)

See below for an example for this transition-based AMR parser:

Requirements

In order to train the parser, you will need to have the following installed:

• Python 3.5.2
• DyNet 2.0
• JAMR
• MGIZA++
• NLTK

Notice that we do not support DyNet version higher than 2.0.

MGIZA++ Installation (Ubuntu)

• Preparation:

sudo apt-get install cmake
sudo apt-get install libboost-all-dev

• Installing: clone the AMR-Parser first, then put the MGIZA++ inside the data folder

cd data
git clone https://github.com/moses-smt/mgiza.git
cd mgiza/mgizapp
cmake .
make
make install

NLTK Tagger Installation

• Installing (under the python3 environment):

import nltk
nltk.download('averaged_perceptron_tagger')
nltk.download('wordnet')

Usage

Hybrid Aligner

Get the AMR corpus (LDC2014T12/LDC2015E86/LDC2017T10). We can't distribute them because them paid. Put the train/dev/test folder of the corresponding AMR corpus (under the split folder) to the folder /data/amr/data/amrs/split. Then we run the preprocess script.

• For LDC2014T12 corpus, run the bash script:

./preprocess_14.sh

• For LDC2015E86, run the bash script:

./preprocess_15.sh

• For LDC2017T10 corpus, run the bash script:

./preprocess_17.sh

Aftering running the script, you can find preprocessed AMR files under the amr/tmp_amr folder, which contains train, dev and test three folders. Before using our Hybrid Aligner, we need the output files of JAMR Aligner.

You can get the aligned train file simply by using the JAMR aligner:

. scripts/config.sh
scripts/ALIGN.sh < amr.txt > train.txt

Then we put three aligned files into one folder called jamr_ouput inside the data folder. Then you can run the bash script of the Hybrid Aligner to get the transition sequences for training the AMR parser:

./align.sh

Finally, we get four new files: train.transtions, dev.transitions, test.transitions and train.txt.pb.lemmas. train.txt.pb.lemmas is the look-up table for predicate and lemma concept (look at the paper for more details.)

AMR Parser

Now we are ready for train our own AMR parser.

Train a new AMR parser

python3 train.py

Link to the word vectors that we used in the ACL 2015 paper for English: sskip.100.vectors. The training process should be stopped when the development SMATCH score does not substantially improve anymore (around epoch 14-16).

Parsing with a trained parser

If you want to use a pretrained parser, just put the pretrain model under the result folder. Then run the script (here we use the pretrained model for the AMR2014 corpus as an example):

python3 train.py --load_model 1 --model result/pretrain14.model --dev_file data/test.transitions --gold_AMR_dev data/amr/tmp_amr/test/amr.txt 

Pretrained models can be downloaded:

• pretrain14.model: Parser trained on AMR2014 corpus.
• pretrain15.model: Parser trained on AMR2015 corpus.

Related Repo

The code of the unsupervised aligning is adapted from this repository.

Citation

If you make use of this software, please cite the following:

@InProceedings{D18-1198,
author = "Guo, Zhijiang and Lu, Wei",
title = "Better Transition-Based AMR Parsing with a Refined Search Space",
booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
year = "2018"
}