lil-lab / amr

Cornell AMR Semantic Parser (Artzi et al., EMNLP 2015)

Cornell AMR Semantic Parser

Requirements

Java 8.

Preparing the Repository

• Get all required resources: ./getres.sh (form the root of the repository)
• Compile: ant dist

Pre-trained Models

A pre-trained model is available to download here.

Parsing

Given a file sentences.txt, which contains a sentence on each line, and a model file amr.sp, both located in the root of the repository:

java -Xmx8g -jar dist/amr-1.0.jar parse rootDir=`pwd` modelFile=`pwd`/amr.sp sentences=`pwd`/sentences.txt

The output files will be in experiments/parse/logs. To see the full set of options (including increasing the logging level), run:

java -jar dist/amr-1.0.jar

Preparing the data (required only for training and testing)

To re-create our experiments, obtain the AMR Bank release 1.0 (LDC2014T12) form LDC. Extract the corpus to the directory corpus/amr_anno_1.0.

Then run the following:

• Compile the code: ant dist
• Prepare the environment: utils/config.sh
• Prepare the data: utils/prepdata-ldc.sh

Attribution

@InProceedings{artzi-lee-zettlemoyer:2015:EMNLP,
  author    = {Artzi, Yoav  and  Lee, Kenton  and  Zettlemoyer, Luke},
  title     = {Broad-coverage CCG Semantic Parsing with AMR},
  booktitle = {Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing},
  month     = {September},
  year      = {2015},
  address   = {Lisbon, Portugal},
  publisher = {Association for Computational Linguistics},
  pages     = {1699--1710},
  url       = {http://aclweb.org/anthology/D15-1198}
}

amr's Issues

Training Not Working

I've been having a lot of trouble training new models. During the the training process,
"Extracting features for level 2 inference
Done - Extracting features for level 2 inference"
is printed over and over until at some point my CPU usage drops to almost 0% and then it says im out of memory in my shell. Then if you wait an incredibly long amount of time, training continues. I have not been able to train any models successfully yet on any of the data in the amr 1.0 release (accept one model which I trained on just two sentences from the bolt treebank). Have you run into this issue and/or what specs do you suggest I have to train a model in this way. I'm running on 12 CPU cores with 32 Gb of memory. I tried looking at the source code to see if I could fix the problem there but the issue is happening in one of the precompiled jar files (i'm gonna look at spf now to see which jar is giving me issues and see if I can create a new jar file from there) Any suggestions? Thank you so much.

binary files in resources/propbank

After running getres.sh, there are binary files in the resources/propbank folders with names like "._bore.xml". On Linux, these will cause the PropBankReader to crash, since it attempts to parse all files in the directory.

The ._*.xml files appear to be OSX metadata. Maybe there's a difference in the way Files.walk(...) and/or Files.isRegularFile(...) treat them, e.g. ignore on OSX, while they're just normal files on Linux?

Anyway, a simple workaround is to do rm resources/propbank/._*.xml after running getres.sh

Train

How to create a new model using new dataset..?

AMR v.s SPF

Hi,

Not really an issue, but how does this repo relate to https://github.com/cornell-lic/spf ?

Thanks!

Pretrained Model

Was the pretrained model trained on AMR bank release 1.0?