This repository provides software to support participants in the shared task on Meaning Representation Parsing (MRP) at the 2019 Conference on Computational Natural Language Learning (CoNLL). Please see the above task web site for additional background.
mtool implements the official MRP 2019 cross-framwork metric, as well as
a range of framework-specific graph similarity metrics, viz.
- MRP (Maximum Common Edge Subgraph Isomorphism);
- EDM (Elementary Dependency Match; Dridan & Oepen, 2011);
- SDP Labeled and Unlabeled Dependency F1 (Oepen et al., 2015);
- SMATCH Precision, Recall, and F1 (Cai & Knight, 2013);
- UCCA Labeled and Unlabeled Dependency F1 (Hershcovich et al., 2019).
The ‘official’ cross-framework metric for the MRP 2019 shared task is a generalization of the framework-specific metrics, considering all applicable ‘pieces of information’ (i.e. tuples representing basic structural elements) for each framework:
- top nodes;
- node labels;
- node properties;
- node anchoring;
- directed edges;
- edge labels; and
- edge attributes.
When comparing two graphs, node-to-node correspondences need to be established (via a potentially approximative search) to maximize the aggregate, unweighted score of all of the tuple types that apply for each specific framework. Directed edges and edge labels, however, are always considered in conjunction during this search.
./main.py --read mrp --score mrp --gold data/sample/eds/wsj.mrp data/score/eds/wsj.pet.mrp
{"n": 87,
"tops": {"g": 87, "s": 87, "c": 85, "p": 0.9770114942528736, "r": 0.9770114942528736, "f": 0.9770114942528736},
"labels": {"g": 2500, "s": 2508, "c": 2455, "p": 0.9788676236044657, "r": 0.982, "f": 0.9804313099041533},
"properties": {"g": 262, "s": 261, "c": 257, "p": 0.9846743295019157, "r": 0.9809160305343512, "f": 0.982791586998088},
"anchors": {"g": 2500, "s": 2508, "c": 2430, "p": 0.9688995215311005, "r": 0.972, "f": 0.9704472843450479},
"edges": {"g": 2432, "s": 2439, "c": 2319, "p": 0.95079950799508, "r": 0.9535361842105263, "f": 0.952165879696161},
"attributes": {"g": 0, "s": 0, "c": 0, "p": 0.0, "r": 0.0, "f": 0.0},
"all": {"g": 7781, "s": 7803, "c": 7546, "p": 0.9670639497629117, "r": 0.9697982264490426, "f": 0.9684291581108829}}
Albeit originally defined for one specific framework (EDS, DM and PSD, AMR, or UCCA, respectively), the pre-MRP metrics are to some degree applicable to other frameworks too: the unified MRP representation of semantic graphs enables such cross-framework application, in principle, but this functionality remains largely untested (as of June 2019).
The Makefile in the data/score/ sub-directory shows some example calls for the MRP scorer.
As appropriate (e.g. for comparison to third-party results), it is possible to score graphs in
each framework using its ‘own’ metric, for example (for AMR and UCCA, respectively):
./main.py --read mrp --score smatch --gold data/score/amr/test1.mrp data/score/amr/test2.mrp
{"n": 3, "g": 30, "s": 29, "c": 24, "p": 0.8, "r": 0.8275862068965517, "f": 0.8135593220338982}
./main.py --read mrp --score ucca --gold data/score/ucca/ewt.gold.mrp data/score/ucca/ewt.tupa.mrp
{"n": 3757,
"labeled":
{"primary": {"g": 63720, "s": 62876, "c": 38195,
"p": 0.6074654876264394, "r": 0.5994193345888261, "f": 0.6034155897500711},
"remote": {"g": 2673, "s": 1259, "c": 581,
"p": 0.4614773629864972, "r": 0.21735877291432848, "f": 0.2955239064089522}},
"unlabeled":
{"primary": {"g": 56114, "s": 55761, "c": 52522,
"p": 0.9419128064417783, "r": 0.9359874541112735, "f": 0.938940782122905},
"remote": {"g": 2629, "s": 1248, "c": 595,
"p": 0.47676282051282054, "r": 0.22632179535945227, "f": 0.3069383543977302}}}
For all scorers, the --trace command-line option will enable per-item scores in the result
(indexed by graph identifiers).
For MRP and SMATCH, the --limit option controls the maximum number of node pairings or
hill-climbing iterations, respectively, to attempt during the search (with defaults 500000
and 5, respectively).
Kuhlmann & Oepen (2016) discuss a range of structural graph statistics; mtool integrates their original code, e.g.
./main.py --read mrp --analyze data/sample/amr/wsj.mrp
(01) number of graphs 87
(02) number of edge labels 52
(03) \percentgraph\ trees 51.72
(04) \percentgraph\ treewidth one 51.72
(05) average treewidth 1.494
(06) maximal treewidth 3
(07) average edge density 1.050
(08) \percentnode\ reentrant 4.24
(09) \percentgraph\ cyclic 13.79
(10) \percentgraph\ not connected 0.00
(11) \percentgraph\ multi-rooted 0.00
(12) percentage of non-top roots 0.00
(13) average edge length --
(14) \percentgraph\ noncrossing --
(15) \percentgraph\ pagenumber two --
mtool can test high-level wellformedness and (superficial) plausiblity of MRP
graphs through its emerging --validate option.
The MRP validator continues to evolve, but the following is indicative of its
functionality:
./main.py --read mrp --validate all data/validate/eds/wsj.mrp
validate(): graph ‘20001001’: missing or invalid ‘input’ property
validate(): graph ‘20001001’; node #0: missing or invalid label
validate(): graph ‘20001001’; node #1: missing or invalid label
validate(): graph ‘20001001’; node #3: missing or invalid anchoring
validate(): graph ‘20001001’; node #6: invalid ‘anchors’ value: [{'from': 15, 'to': 23}, {'from': 15, 'to': 23}]
validate(): graph ‘20001001’; node #7: invalid ‘anchors’ value: [{'form': 15, 'to': 17}]
Among its options for format coversion, mtool supports output of graphs to the
DOT language for graph visualization, e.g.
./main.py --id 20001001 --read mrp --write dot data/sample/eds/wsj.mrp 20001001.dot
dot -Tpdf 20001001.dot > 20001001.pdf
When converting from token-based file formats that may lack either the underlying
‘raw’ input string, character-based anchoring, or both, the --text command-line
option will enable recovery of inputs and attempt to determine anchoring.
Its argument must be a file containing pairs of identifiers and input strings, one
per line, separated by a tabulator, e.g.
./main.py --id 20012005 --text data/sample/wsj.txt --read dm --write dot data/sample/psd/wsj.sdp 20012005.dot
For increased readability, the --ids option will include MRP node identifiers
in graph rendering, and the --strings option can replace character-based
anchors with the corresponding sub-string from the input field of the graph
(currently only for the DOT output format), e.g.
./main.py --n 1 --strings --read mrp --write dot data/sample/ucca/wsj.mrp vinken.dot
The --read and --write command-line options determine the input and output
codecs to use.
Valid input arguments include mrp, amr, ccd, dm, eds, pas, psd, ud,
and ucca; note that some of these formats are only partially supported.
The range of supported output codecs includes mrp, dot, or txt.
The optional --id, --i, or --n options control which graph(s)
from the input file(s) to process, selecting either by identifier, by (zero-based)
position into the sequence of graphs read from the file, or using the first n
graphs.
These options cannot be combined with each other and take precendence over each
other in the above order.
You can install mtool via pip with the following command:
pip install git+https://github.com/cfmrp/mtool.git#egg=mtool
- Yuta Koreeda koreyou@mac.com (@koreyou)
- Matthias Lindemann mlinde@coli.uni-saarland.de (@namednil)