StanfordGraph

Arboreal MRS Discussion

SIG at Stanford 2016 summit

===

Mike Goodman: DMRS is still not tree-like enough. Notation like Stephan's, but with POST. "-of" means an inverted edge

Guy Emerson: There are two things we could be talking about here: a different notation for DMRS, or a rooted graph - different theoretical status.

Stephan Oepen: At the outset, two types of dependency graph: ED, and DMRS. POST, and undirected edges (being addressed). And occasional creative differences, like "nearly every". Issue is to convert these graphs to singly-rooted DAGs, either based on what constitutes a non-scopal modifier, and the other by walking the graph. Is there is a unique solution for either method? It's possible that the order in which order you reach nodes changes the output.

Guy: An undirected cycle?

Woodley Packard: "persuaded Kim to leave", several ways of serialising it...

===

Various options discussed during SIG:

(1) persuade

• 2 Kim [] 3 leave
  • 1 []

(2) persuade

• 2 Kim

  • 1-of leave []

  3 []

Same DMRS as (1), but different if considering 1-of as a directed edge in the other direction

(3) persuade

• 3 leave

  • 1 Kim []

  2 []

Same DMRS as (1), and also the same when "-of" links are considered flipped; different serialisation from (1)

(4) Kim []

• 2-of persuade
  • 3 leave
    • 1 []

When considering "-of" links as flipped, this graph is cyclic: re-entrancy points to a dominating node

(5) leave [1]

• 1 Kim [2]

persuade

• 2 [2] 3 [1]

This is the same graph as (1), but as a serialisation, this is no longer singly rooted.

===

Guy: When traversing the graph, we could choose to visit edges depth-first in order (ARG1, ARG2, ARG3), to get (1); or we could flip as few as possible, to get (1) or (3). Or we could minimise cycles in the graph (not 4)

Stephan: Is there a canonical solution?

Mike: We could just pick a node and let things fall off it. If we want to order nodes based on some function, that's something we could do, but we could have a canonical default option.

Woodley: Asking for subgraphs to be the same in two different contexts may be asking for more than you can have. For example, with re-entrancies.

"Abrams persuaded Kim to leave"

"Kim left"

"Kim, who left, slept"

So do we care about making these look the same?

Mike: I can imagine I would like these to be the same, but it's nice to know that they're different. I'd like to transform them to be the same (could be useful for MT), but as default, they would be different.

Guy: This comes back to whether we are serialising DMRS graphs or modifying them...

Woodley: but we have "leaves" in the serialisation. So why are we serialising? Mike is doing it to compare tree fragments, so it does matter what the serialisation is.

Stephan: We found directed cycles in AMR, if edges are flipped as annotated - but if we normalise the graphs, then the cycles go away. (CL forthcoming). So it comes back to whether this about notation, or about different graphs.

Stephan: AMR people have a notion of "pulling up" a node, e.g. for focus. We similarly have "top"

Woodley: serialisation singly-rooted or graph singly-rooted... see (5)

===

Mike: Another possible transformation is to change some nodes to arcs.

Guy: Like Stephan's bilexical dependencies (DM)

Stephan: Similar to what we did for abstract predicates. Also Stanford dependencies. Won't be lossless if the preposition is modified or conjoined (nearly in / in or on)

Mike: still lossless if we don't convert those cases...

Guy: It's impossible whenever there are multiple outgoing edges

Mike: And only for binary relations

Alex Kuhnle: Including transitive verbs?

Stephan: They are formally equivalent, but the corner cases are much more common. Prepositions are a small class.

Mike: Number example on the slide... nominalisations...

Stephan: We're now discussing graph transformations that make your job easier, without losing important information

===

Mike: Can we unify our representations?

"nearly all" - EDS, add in ARG1; DMRS EQ link

Stephan: I like that we agree on the topology; Ann has given a principled way of arriving at this result

Mike: Any other distinctions? Both solved scopal arguments.

Stephan: Not a theoretical distinction, but the EDS code is more robust, to deal with ill-formed MRSs.

===

Mike: Compare algorithms?

Stephan: use depth-first traversal with canonical order (unprincipled)

===

Agreed we don't need a special name for a serialisation of MRS. Agreed not to call it "AMR" or "AMRS", but we can acknowledge inspiration.

Penman?

Mike and Stephan will go think about it... and also discuss canonical serialisation.