search

cfmrp / mtool

Software to Manipulate Different Flavors of Semantic Graphs
http://mrp.nlpl.eu
GNU Lesser General Public License v3.0
51 stars 24 forks source link

Do we need to predict node values in SDP? #18

Closed wangxinyu0922 closed 5 years ago

wangxinyu0922 commented 5 years ago

As the properties in SDP are fixed to be ['pos','frame'] pair, so do we need to predict the corresponding values? As we can get "pos" value from companion data, we only need "frame"?

oepen commented 5 years ago

yes, both PoS and frame predictions will be scored, each as a node property in its own right. note, however, that not all nodes have frame values in PSD (essentially only verbs do, where the PSD frame values are in fact sense identifiers into the CzEngVallex dictionary).

wangxinyu0922 commented 5 years ago

图片 Hi, is this table that the what we need to predict in evaluation? For example, we don't need to predict node properties in EDS.

oepen commented 5 years ago

thanks for catching this one: the table was wrong in this respect! inasmuch as there are node properties in EDS, they will be scored in MRP evaluation. for MRP, we have in some cases dropped information from the original graph banks, for example :wiki links in AMR, implicit (unexpressed nodes) in UCCA, and most EDS node properties (things like tense and number). however, as you point out there are remaining node properties in EDS, notably the CARG in named entities, and that is certainly relevant for MRP parsers to predict.