search

huji-nlp / ucca

Universal Conceptual Cognitive Annotation (UCCA)
https://universalconceptualcognitiveannotation.github.io/
GNU General Public License v3.0
20 stars 21 forks source link

Support for coordinated main relations #57

Closed danielhers closed 5 years ago

danielhers commented 5 years ago

UCCA-App now contains a category Coordinated_Main_Rel.. As of https://github.com/danielhers/ucca/commit/5406b6981c437d246e4a84faadb082312a83f46a, it is ignored when converting from JSON to XML, but the proper solution is to convert it as follows:

  1. Split the Centers inside the main relation to be the main relations of separate Parallel Scenes
  2. Move the Connector inside the main relation to be a Linker between the new Parallel Scenes
  3. Attach any element of the scene as remotes of the second Parallel Scene

@omriabnd

danielhers commented 5 years ago

Guidelines:

If two or more coordinated Processes or two or more coordinated States share exactly the same scene elements (As, Ds, Ts), then we unite them under one parent unit which is assigned two categories: Coordinated Main Relation (CMR) and also P or S accordingly. Internally, we use multiple Cs to mark each Main Relation separately. so: John [wrote_C and_N recorded_C]_P+CMR a song on Tuesday

On the other hand, if the coordinated Main Relations don't share one of the Scene elements, we separate them into different scenes. [John_A [woke up]_P]_A,]_H [went_P [to school]_A (John)_A]_H and [met_P Mary_A there_A (John)]_H

danielhers commented 5 years ago

Example: https://github.com/UniversalConceptualCognitiveAnnotation/UCCA_English-EWT/blob/master-images/001961-0002.svg [[Rusted out]_C and_N unsafe_C]_S cars_A should be [[[Rusted out]_S cars_A]_H and_L [unsafe_S cars_A*]_H]_S

nschneid commented 5 years ago

@danielhers Could you please update the images to reflect the CMR normalization?

danielhers commented 5 years ago

@nschneid Done. It happens automatically but it just took a while for the Travis CI job to finish.

nschneid commented 5 years ago

Great. I did some spot checks and they mostly look good—at least, the ones where the graph wasn't too tangled to read. :) Found one thing that I'm opening as a new issue.