A more updated version of this parser, supporting other languages, is available at: https://github.com/mdtux89/amr-eager-multilingual

AMR-EAGER [1] is a transition-based parser for Abstract Meaning Representation (http://amr.isi.edu/).

• Make sure you have Java 8
• Install Torch and torch packages dp, nngraph and optim (using luarocks, as explained here: http://torch.ch/docs/getting-started.html)
• Install the following python dependencies: numpy and pytorch (https://github.com/hughperkins/pytorch)
• Run ./download.sh
• Install JAMR aligner (https://github.com/jflanigan/jamr) and set path in preprocessing.sh

Note: the input file must contain English sentences (one sentence for line), see contrib/sample-sentences.txt for example.

Preprocessing:

./preprocessing.sh -s <sentences_file>

You should get the output files in the same directory as the input files, with the prefix <sentences_file> and extensions .out and .sentences.

python preprocessing.py -f <sentences_file>

You should get the output files in the same directory as the input files, with the prefix <sentences_file> and extensions .tokens.p, .dependencies.p.

Parsing:

python parser.py -f <file> -m <model_dir>

If you wish to have the list of all nodes and edges in a JAMR-like format, add option -n. Without -m the parser uses the model provided in the directory LDC2015E86.

Mac users: the pretrained models seem to have compatibility errors when running on Mac OS X.

We provide evaluation metrics to compare AMR graphs based on Smatch (http://amr.isi.edu/evaluation.html). The script computes a set of metrics between AMR graphs in addition to the traditional Smatch code:

• Unlabeled: Smatch score computed on the predicted graphs after removing all edge labels
• No WSD. Smatch score while ignoring Propbank senses (e.g., duck-01 vs duck-02)
• Named Ent. F-score on the named entity recognition (:name roles)
• Wikification. F-score on the wikification (:wiki roles)
• Negations. F-score on the negation detection (:polarity roles)
• Concepts. F-score on the concept identification task
• Reentrancy. Smatch computed on reentrant edges only
• SRL. Smatch computed on :ARG-i roles only

The different metrics are detailed and explained in [1], which also uses them to evaluate several AMR parsers. (Some of the metrics were recently fixed and updated)

cd amrevaluation
./evaluation.sh <file>.parsed <gold_amr_file>

To use the evaluation script with a different parser, provide the other parser's output as the first argument.

• Preprocess training and validation sets:

  ./preprocessing.sh <amr_file>
  python preprocessing.py --amrs -f <amr_file>
  

• Run the oracle to generate the training data:

  python collect.py -t <training_file> -m <model_dir>
  python create_dataset.py -t <training_file> -v <validation_file> -m <model_dir>
  

• Train the three neural networks:

  th nnets/actions.lua --model_dir <model_dir>
  th nnets/labels.lua --model_dir <model_dir>
  th nnets/reentrancies.lua --model_dir <model_dir>
  

  (use also --cuda if you want to use GPUs).

• Finally, move the .dat models generated by Torch in <model_dir>/actions.dat, <model_dir>/labels.dat and <model_dir>/reentrancies.dat.

• To evaluate the performance of the neural networks run

  th nnets/report.lua <model_dir>
  

• Note: If you used GPUs to train the models,you will need to uncomment the line require cunn from nnets/classify.lua.

• Smatch: http://amr.isi.edu/evaluation.html
• Tokenizer: https://github.com/redpony/cdec
• CoreNLP: http://stanfordnlp.github.io/CoreNLP/

[1] "An Incremental Parser for Abstract Meaning Representation", Marco Damonte, Shay B. Cohen and Giorgio Satta. Proceedings of EACL (2017). URL: https://arxiv.org/abs/1608.06111