CAMR: A transition-based AMR Parser

CAMR is a transition-based, tree-to-graph parser for the Abstract Meaning Representation of a sentence. It is a product of an on-going collaboration between the Chinese Language Processing Group at Brandeis University and cemantix.org

Reference:

@InProceedings{wang-xue-pradhan:2015:NAACL-HLT,
  author    = {Wang, Chuan  and  Xue, Nianwen  and  Pradhan, Sameer},
  title     = {A Transition-based Algorithm for AMR Parsing},
  booktitle = {Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  month     = {May--June},
  year      = {2015},
  address   = {Denver, Colorado},
  publisher = {Association for Computational Linguistics},
  pages     = {366--375},
  url       = {http://www.aclweb.org/anthology/N15-1040}
}
@InProceedings{wang-xue-pradhan:2015:ACL-IJCNLP,
  author    = {Wang, Chuan  and  Xue, Nianwen  and  Pradhan, Sameer},
  title     = {Boosting Transition-based AMR Parsing with Refined Actions and Auxiliary Analyzers},
  booktitle = {Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)},
  month     = {July},
  year      = {2015},
  address   = {Beijing, China},
  publisher = {Association for Computational Linguistics},
  pages     = {857--862},
  url       = {http://www.aclweb.org/anthology/P15-2141}
}
@InProceedings{wang-EtAl:2016:SemEval,
  author    = {Wang, Chuan  and  Pradhan, Sameer  and  Pan, Xiaoman  and  Ji, Heng  and  Xue, Nianwen},
  title     = {CAMR at SemEval-2016 Task 8: An Extended Transition-based AMR Parser},
  booktitle = {Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016)},
  month     = {June},
  year      = {2016},
  address   = {San Diego, California},
  publisher = {Association for Computational Linguistics},
  pages     = {1173--1178},
  url       = {http://www.aclweb.org/anthology/S16-1181}
}

Updates

09-2016 We have included the retrain instruction.

08-2016 We have released new model trained on bigger dataset.

Dependencies

First download the project:

  git clone https://github.com/Juicechuan/AMRParsing.git

Here we use a modified version of the Stanford CoreNLP python wrapper, Charniak Parser and Stanford CoreNLP toolkit. To setup dependencies, run the following script:

  ./scripts/config.sh

Note: For Mac users, there are some problems when installing the Charniak Parser python module bllipparser. We recommend to use Linux system in order to utilize the Charniak Parser. Also you may need swig to successfully install bllipparser.

Parsing with Pre-trained Model

The input data format for parsing should be raw document with one sentence per line.

Preprocessing

To preprocess the data, run:

  python amr_parsing.py -m preprocess [input_sentence_file]

This will give you the tokenized sentences(.tok), POS tag and name entity (.prp) and dependency structure (.charniak.parse.dep) (generated by Charniak parser and Stanford Dependency converter).

Note: The parser will try to read the preprocessed file with the above suffix. So if the preprocessing is not complete, do remove all the cached file and then re-run this step.

Parsing

Download the following model:

LDC2014T12: trained on training set of LDC2014T12.

SemEval2016: trained on training set of SemEval 2016 Task 8.

Uncompress the model file, then use the following command to parse the sentence:

  python amr_parsing.py -m parse --model [model_file] [input_sentence_file] 2>log/error.log

This will give your the parsed AMR file(.parsed) in the same directory of your input sentence file.

Note: these models doesn't incorporate the Semantic Role Label feature, so it will give you slightly lower results reported in the papers. We are working on integrating the SRL system into the pipeline.

Retrain the parser

If you would like to retrain the parser, you first have to obtain the alignment in the following section.

Alignment

If you have annotated AMR file, you could first run the preprocessing step:

python amr_parsing.py -m preprocess --amrfmt amr [input_amr_file]

This will generate a tokenized AMR file (.amr.tok) (which has :tok tag in the comments). Then you can run the following command to get the aligned AMR file(.aligned)

./scripts/jamr_align.sh [input_amr_tok_file]

Note: We use JAMR to get the alignment between sentence and its AMR annotation. You need to download and set up JAMR.

Retrain

To retrain the model with the provided feature template basic_abt_brown_feats.templates , use the following command:

  python2.7 amr_parsing.py -m train --amrfmt amr --verblist --smatcheval --model [path_to_save_model] -iter [iter_number] --feat ./feature/basic_abt_brown_feats.templates [path_to_train_amr_file] -d [path_to_dev_amr_file] > [log_file] 2>&1 &

The parser will be trained using basic feature template mentioned in our NAACL 2015 paper.

Note:

  • Since in the preprocessing step we don't include the Semantic Role Labeling feature, it wouldn't replicate the optimal results.
  • If you want to retrain the model with different features, you will have to get your hands dirty with the feature template file with examples defined in temp/ folder and modify the feature functions in model.py/newstate.py.