Awesome

AMR parse(r) evaluation experiment

This is code/data for our paper Better Smatch = better parser? AMR evaluation is not so simple anymore

Data

Data for little prince:

• qualitylp/abparser/src.test.amr: pairs of candidate parses (first parse is of BART based parser, second parse by T5 based parser)
• qualitylp/abparser/tgt.test.amr: pairs of reference parses (first and second parse are the same)
• qualitylp/LABEL.txt: contains human annotation for the pairs. First number: preference (1: prefer first, 0: prefer second, 0.5: both are equal quality). Second number: is first parse acceptable? 1: Yes, 0: No. Third number: Is second parse acceptable? 1: yes, 0: No

Data for AMR3:

Is found in qualityamr3. It follows the same format as little prince data. Note: Release of reference graphs requires LDC license, therefore this repo does not contain qualityamr3/abparser/tgt.test.amr. For constructing the reference AMR3, please see also Notes below.

Run experiments

Please look in this directory:

evaluation-suite

Notes

• Removal of :wiki relations: Since parsers usually use external post-precessing for wiki linking, we do not consider wiki relations for AMR quality. We added a script for wiki removal in util/format_amr.py. When constructing the reference for AMR3, it has to be applied.

• Re-naming variables: We standardized variables with a running index and prevent occurence of a variable that is also a concept. See also the script in util/format_amr.py. This is internally also done by some, but not all metrics.

Citation

If you find the work interesting, consider citing

@inproceedings{opitz-frank-2022-better,
    title = "Better {S}match = Better Parser? {AMR} evaluation is not so simple anymore",
    author = "Opitz, Juri and Frank, Anette",
    booktitle = "Proceedings of the 3rd Workshop on Evaluation and Comparison of NLP Systems",
    month = nov,
    year = "2022",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.eval4nlp-1.4",
    doi = "10.18653/v1/2022.eval4nlp-1.4",
    pages = "32--43",
}