ctakes-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Miller, Timothy" <Timothy.Mil...@childrens.harvard.edu>
Subject RE: types for hybrid relations
Date Tue, 10 Feb 2015 12:36:40 GMT
I'm not sure what in particular you're interested in so prepare for an info dump.

Section 1 of this paper is good background on the problem setting (mention pair vs. entity-mention
models in coreference)
https://urldefense.proofpoint.com/v2/url?u=http-3A__www.aclweb.org_anthology_P08-2D1096.pdf&d=BQIF-g&c=qS4goWBT7poplM69zy_3xhKwEW14JZMSdioCoppxeFU&r=6K2jncop0hjH-CSVJRe1t5Ehv0V75znADU0wtfGz_1w&m=MWh5uyeBcgcOAqVmBqOxmUXfFSm7gfb4iMfjUXhbJtA&s=MK7xeG4xtmFVdRKEUqeOKr0Cktw7fHl_aLI5oR_yexQ&e=


As for the ctakes aspect of it, there is a division in the typesystem between "spanned" types
-- these derive from the UIMA Annotation class -- and unspanned types. Spanned types can point
to character offsets in the original document where unspanned types do not. Generally we use
spanned types for mentions of entities and events, while we (intend to) use unspanned types
to represent abstract entities/events. These basically stand in for the real-world events/entities
that are referred to by mentions.

One way of thinking about this issue is that relations could be more general. The BinaryTextRelation
is a relation between two RelationArguments, which wrap Annotation, but if RelationArgument
wrapped a TOP it could be more general.

Alternatively, we could create a ElementMentionRelation that is a new type of relation with
one RelationArgument argument and one Element argument (Element is the shared parent of Entity
and Event in the typesystem).

Finally, since in reality the coreference module creates a type called CollectionTextRelation,
which is technically a "text relation" and not an Element (Event or Entity), we could create
a RelationMentionRelation which would allow for linking a RelationArgument (Event or entity
mention) with a relation.

Any votes for one or more of the following:

A) Generalize BinaryTextRelation
B) Create ElementMentionRelation (and then map coref chains to Elements)
C) Create RelationMentionRelation
D) I'm not doing anything until I clear this mountain of snow off of my car

Tim


________________________________________
From: John Green [john.travis.green@gmail.com]
Sent: Tuesday, February 10, 2015 6:21 AM
To: dev@ctakes.apache.org
Cc: dev@ctakes.apache.org
Subject: Re: types for hybrid relations

Im interested in hearing more about this.


John
—
Sent from Mailbox

On Mon, Feb 9, 2015 at 12:22 PM, Miller, Timothy
<Timothy.Miller@childrens.harvard.edu> wrote:

> The typesystem has a few different basic relations:
> Relation: The base type, it has no information about how many arguments
> or what type of arguments it uses.
> BinaryTextRelation: Between 2 RelationArgument objects, which are
> wrappers for UIMA Annotation type (spanned arguments).
> CollectionTextRelation: Between a set of RelationArgument objects
> (Annotation)
> ElementRelation: Between 2 Element objects, which are non-spanned types,
> with pointers to mentions.
> AttributeRelation: Between an Element and an Attribute, which is a type
> of Element.
> However, as far as I can tell there is no relation type which would
> allow for a link between an Annotation and an Element. This use case
> comes up in certain models of coreference resolution, where you attempt
> to link new mentions back to clusters instead of to individual mentions.
> I am interested in trying out models of this type and was going to
> extend RelationExtractorAnnotator but I think the typesystem doesn't
> have what we need for this case. Someone please correct me if I'm wrong,
> but I would propose to modify the typesystem to make such relations
> possible.
> Thanks
> Tim

Mime
View raw message