#183 Segment Processing PR creates new annotation ids, breaking Orthomatcher 'matches' feature values

[Phil Gooch]

Running ANNIE with Orthomatcher within a Segment Processing PR causes incorrect 'matches' feature values to be created, as the annotation ids created by Orthomatcher are no longer valid in the new document created by Segment Processing PR

[Phil Gooch]

Steps to reproduce.

1. Load ANNIE, and add Pronominal Coreference to it.
2. Create a document containing the attached text (from bbc.co.uk)
3. Create a corpus from this document
4. Run ANNIE. Note that 'matches' feature values on pronouns and person names are valid and correct
5. Remove Document Reset PR from ANNIE and create a new pipeline containing Segment Processing PR, with analyser parameter set to ANNIE
6. Delete all annotations from the document, then add two new Section annotations around each paragraph
7. Run the new pipeline that contains the Segment Processing PR on this document. Note the invalid values of the 'matches' feature - none of the ids listed exist in the result document.

[Mark Greenwood]

I think there are a number of issues here, but essentially I don't think we are going to fix this, hence I'm closing the bug.

The main thing is that while the coref tools are useful hard wiring support for them into all the other tools is unlikely to happen, and certainly isn't sustainable.

The problem of course is that the coref (and the new relations API) use the annotation IDs to refer to annotation instances and that breaks if you start moving annotations around.

In this specific instance it is possible to work around the problem by changing the processing order. Instead of running all of ANNIE on each section in turn (and geting silly results) you could run everything but the coref on the full document and then just run coref over the results in each section.

[Mark Greenwood]

actually I'm not sure my previous statement is entirely true so I'm re-opening this one while I investigate further

[Phil Gooch]

Maybe the best approach is to run ANNIE except the Orthomatcher via the
Segment Processing PR pipeline, and then put the Orthomatcher PR after the
Segment Processing PR in that pipeline?

Phil

On Sat, Apr 12, 2014 at 8:44 AM, Mark Greenwood <markagreenwood@users.sf.net

wrote:

• status: closed-wont-fix --> open
• Comment:

actually I'm not sure my previous statement is entirely true so I'm
re-opening this one while I investigate further

---

• [bugs:#183] http://sourceforge.net/p/gate/bugs/183/ Segment Processing
  PR creates new annotation ids, breaking Orthomatcher 'matches' feature
  values*

Status: open
Group: HEAD
Created: Mon Nov 25, 2013 03:11 PM UTC by Phil Gooch
Last Updated: Fri Apr 11, 2014 09:12 PM UTC
Owner: nobody

Running ANNIE with Orthomatcher within a Segment Processing PR causes
incorrect 'matches' feature values to be created, as the annotation ids
created by Orthomatcher are no longer valid in the new document created by
Segment Processing PR

---

Sent from sourceforge.net because you indicated interest in
https://sourceforge.net/p/gate/bugs/183/

To unsubscribe from further messages, please visit
https://sourceforge.net/auth/subscriptions/

[Mark Greenwood]

Having made other changes to the segment processing PR I've retested this bug and it now seems to be fixed. However, there is still one outstanding issue in that while the matches feature on annotations now contains the correct IDs and so is valid, the document feature containing all the chains is not created and so the coreference editor appears to show no chains in the document. You should be able to work around this with a jape grammar or groovy script though.