make an environment with python 3.7

activate environment

pip install -r requirements.txt

Link to NAACL 2022 paper DOCAMR: Multi-Sentence AMR Representation and Evaluation

https://aclanthology.org/2022.naacl-main.256.pdf

To create docAMR representation from gold AMR3.0 data and the coref annotation in xml format:

python doc_amr.py 
--amr3-path <path to AMR3 data> 
--coref-fof <file-with-list-of-xml-annotations-files> 
--out-amr <output file> 
--rep <representation>

<path to AMR3 data> should point to uncompressed LDC data directory for LDC2020T02 with its original directory structure.

<file-with-list-of-xml-annotations-files> is one of the _coref.fof files included in this repository.

Default value for --rep is 'docAMR'. Other values can be: 'no-merge','merge-names','merge-all'. Use --help to read the descriptions of these representations.

To create docAMR representation from dcoument AMRs with no nodes merged

python doc_amr.py
       --in-doc-amr-unmerged <path to document-level AMRs un no-merge format>
       --rep <representation>
       --out-amr <output file>

To create docAMR representation from dcoument AMRs with pairwise edges between a representative node in the chain and the rest of the nodes in the chain:

python doc_amr.py
       --in-doc-amr-pairwise <path to document-level AMR with pairwise coref edges>
       --pairwise-coref-rel <relation label indicating coref edges>
       --rep <representation>
       --out-amr <output file>

default value for --pairwise-coref-rel is same-as

Use docSmatch the same way as the standard Smatch.

python docSmatch/smatch.py -f <amr1> <amr2>

It assumes that :snt relations connect sentences to the root. Moreover, it assumes that the numeric suffix of :snt is the sentence number and that the matching sentence numbers in the two AMRs are aligned.

You can also get a detailed score breakdown for the accuracy of coreference prediction:

python docSmatch/smatch.py -f <amr1> <amr2> --coref-subscore

This will ouput the normal smatch score as 'Overall Score', as well as a 'Coref Score' indicating the quality of cross sentential edges and nodes.