Generic Relations

[marcverhagen]

The vocabulary is not well positioned to deal with any kind of relation that is not a semantic role or a dependency. For ReVerb (see issue https://github.com/lappsgrid-incubator/Services/issues/5), which creates relations without typing them, we simply want some annotation type to represent things like "call(ET,home)".

There is a Relation type that defines the label property (used for syntactic categories and names of dependencies in subtypes of relation). I suggest a new subtype named GenericRelation with two properties:

relation - identifier pointing to an object of type Annotation
arguments - list of identifiers pointing to objects of type Annotation

The Annotation objects could be tokens, chunks or markables.

[marcverhagen]

By the way, I first thought of having a type named BinaryRelation with three properties:

relation - identifier pointing to an object of type Annotation
arg1 - identifier pointing to an object of type Annotation
arg2 - identifier pointing to an object of type Annotation

But I felt that since we would then want ternary and n-ary relations as well I might just bite the bullet.

[reckart]

Won't creating such a generic relation type be tempting to derive all other relations from it? How about naming the type"Fact" or something similar?

[marcverhagen]

It may. I am also a bit uneasy about having GenericRelation and Dependency be siblings. As a third issue, being a subset of Relation a GenericRelation needs a label, and there is no meaningful label except for something bland like "relation", which would apply to all instances of GenericRelation.

I like Fact as a name, but I also like how Relation or GenericRelation connects to common use of the term in many NLP tools.

[marcverhagen]

Here is how this may work out in LIF, what is printed here is a view for the ReVerb output of "She swam to Paris.":

{
  "metadata": {
    "contains": {
      "Token": {"producer": "org.anc.lapps.reverb.relation:2.0.0" }}
      "Markable": {"producer": "org.anc.lapps.reverb.relation:2.0.0" }}
      "GenericRelation": {"producer": "org.anc.lapps.reverb.relation:2.0.0" }}},
  "annotations": [
     {"id":"t0", "start":0, "end":3, "@type":"Token", "features": {"pos":"PRP"}},
     {"id":"t1", "start":4, "end":8, "@type":"Token", "features": {"pos":"VBD"}},
     {"id":"t2", "start":9, "end":11, "@type":"Token", "features": {"pos":"TO"}},
     {"id":"t3", "start":12, "end":17, "@type":"Token", "features": {"pos":"NNP"}},
     {"id":"t4", "start":17, "end":18, "@type":"Token", "features": {"pos":"."}},
     {"id":"m1", "@type":"Markable", "features": {"targets":["t1"]}},
     {"id":"m2", "@type":"Markable", "features": {"targets":["t2","t3"]}},
     {"id":"m3", "@type":"Markable", "features": {"targets":["t4"]}},
     {"id":"r1", 
       "@type":"GenericRelation", 
       "features": {
         "label":"swim to", 
         "relation":"m2", 
         "arguments": ["m1","m3"]  }}]
}

[marcverhagen]

Perhaps we can use something like UntypedRelation for this, where untyped refers to the absence of a label or just the string (for example "swim to") as the label. This would dampen people's enthousiasm for over-using or sub typing this.

[ksuderman]

Added in ec2378ea48adfd2db1e8c869f46af236ef4810c3