search

coli-saar / am-parser

Modular implementation of an AM dependency parser in AllenNLP.
Apache License 2.0
30 stars 10 forks source link

Keep remote edges for decomposition + adapt heuristics #59

Closed namednil closed 4 years ago

namednil commented 4 years ago

Currently, we simply delete all remote edges. However, they actually make UCCA parsing much more interesting, so we should treat them as well.

alexanderkoller commented 4 years ago

Let's deprioritize the primary/remote prediction.

@mariomgmn check that we are predicting all edges in the output now (including remote ones). Then close.

luciaelizabeth commented 4 years ago

@mariomgmn let me know if I can help at all with the decomposition heuristics. There is a notebook in my office with how different edges can be decomposed with the AM algebra, but I have most of it in my head and would be happy to do a Skype call to discuss. Or shoot me an email with specific questions!

mariomgmn commented 4 years ago

We weren't previously but I just modified the alto corpus creation to take into account the remote edges, (but we are currently not distinguishing between remote and primary).

@luciaelizabeth I'm going to do the edge raising today soon. Afterwards maybe we could Skype to see if we can come up with better decomposition/raising/lowering heuristics.

luciaelizabeth commented 4 years ago

Sure, let me know!