JAMR v0.2

New in this release:

• setup script to download external dependencies (including the version of the Illinois NER system we used)
• scripts for running the aligner and training the parser
• hand alignments for 200 sentences of LDC2013E117 data
• tab-separated output format of the nodes and edges in the AMR graph
• parser prints alignments from graph to the tokenized sentence
• more documentation, and various bugfixes

This release has been tested on Ubuntu 12.04.

The performance on LDC2013E117_DEFT_Phase_1_AMR_Annotation_R3 with the train/dev/test splits used in the paper is:

  *** Test Evaluation (gold concept ID) ***
Precision: 0.837
Recall: 0.758
Document F-score: 0.795
  *** Test Evaluation (all stages) ***
Precision: 0.666
Recall: 0.545
Document F-score: 0.600

To run the parser, download the source and see README.md.
You must agree to the license of the Illinois NER system if you use it with JAMR.

JAMR v0.1

This release has been tested with:

• Scala v2.11.1,
• Illinois NER system v2.4,
• cdec,
• WordNet v3.0.

on Ubuntu 12.04.4 and MacOS X 10.8.5. The performance on LDC2013E117_DEFT_Phase_1_AMR_Annotation_R3 with the train/dev/test splits used in the paper is:

jeff@supercomputer:~/working/jamr$ ./scripts/smatch_v1_0/smatch_modified.py --pr -f experiments/current/test.decode.allstages data/LDC-2013-Sep/amr-release-proxy.test
Precision: 0.652
Recall: 0.568
Document F-score: 0.607
jeff@supercomputer:~/working/jamr$ ./scripts/smatch_v1_0/smatch_modified.py --pr -f experiments/current/test.decode.stage2only data/LDC-2013-Sep/amr-release-proxy.test
Precision: 0.839
Recall: 0.760
Document F-score: 0.797

See README.md for information on how to run. Trained weight files and output on the test set are in JAMR_experiments-v0.1.tgz.