when comparing our (forthcoming) MCES scorer with SMATCH, it emerged that there is a difference in the count of correct tuples for two test AMRs from the SMATCH distribution:
# ::snt The boy is a hard worker.
(p / person
:domain (b / boy)
:ARG0-of (w / work-01
:manner (h / hard)))
# ::snt The boy is a hard worker.
(w / worker
:mod (h / hard)
:domain (b / boy))
SMATCH correctly establishes three node correspondences: the ‘b’ and ‘h’ nodes from both graphs, and ‘p’ from the first graph with ‘w’ from the second. on this view, the two graphs have the same top node, share two node labels (‘boy’ and ‘hard’), as well as one edge (:domain). however, SMATCH appears to represent the top properties as triples that include the node label, which results in an undesirable double penalty. i wonder whether this treatment in SMATCH is background to the critique by Anchieta et al. (2019)?
when comparing our (forthcoming) MCES scorer with SMATCH, it emerged that there is a difference in the count of correct tuples for two test AMRs from the SMATCH distribution:
SMATCH correctly establishes three node correspondences: the ‘b’ and ‘h’ nodes from both graphs, and ‘p’ from the first graph with ‘w’ from the second. on this view, the two graphs have the same top node, share two node labels (‘boy’ and ‘hard’), as well as one edge (:domain). however, SMATCH appears to represent the top properties as triples that include the node label, which results in an undesirable double penalty. i wonder whether this treatment in SMATCH is background to the critique by Anchieta et al. (2019)?