These suggestions came out of the post-NAACL AMR-CCG meeting held at Boulder. There is one suggested principle to adhere to, one tool modification suggestion, and also suggestions for additional annotation.
The principle:
In order to facilitate machine learning approaches and leveraging previous annotations, as much as possible AMR should copy existing guidelines or be compatible with them or give a good reason why not. [24 in favor, 2 against]
The tool modification suggestion:
Could Smatch be enhanced to get subscores for core roles, non-core roles, negation, entities, etc?
There are 3 main categories of suggestions for additions:
1) Things that it would be very helpful to add to AMRs that could possibly be done automatically or semi-automatically as a separate layer in a post-processing step. Could be feature values.
Tense and Aspect
Plurality, including distinguishing between singular and collective readings.
Definiteness - overlaps with logical form and reference.
Topic/focus
Sense tags for nouns and more fine-grained verb senses
2) Things that it would be very helpful to add that could possibly be done manually during token alignment.
Marking elided VP's, implicit arguments, predicates and relations that have been made explicit as "inferred."
3) Something we are currently doing that needs to be fixed.
logical form - see separate issue by Tim O'Gorman.
Potential pilots on some of these topics:
1) Distinguishing "inferred" relations
2) Implicit arguments
3) Annotating scope, negation and quantifiers using "negation" shared task guidelines as a separate layer
These suggestions came out of the post-NAACL AMR-CCG meeting held at Boulder. There is one suggested principle to adhere to, one tool modification suggestion, and also suggestions for additional annotation.
The principle:
In order to facilitate machine learning approaches and leveraging previous annotations, as much as possible AMR should copy existing guidelines or be compatible with them or give a good reason why not. [24 in favor, 2 against]
The tool modification suggestion:
Could Smatch be enhanced to get subscores for core roles, non-core roles, negation, entities, etc?
There are 3 main categories of suggestions for additions:
1) Things that it would be very helpful to add to AMRs that could possibly be done automatically or semi-automatically as a separate layer in a post-processing step. Could be feature values.
Tense and Aspect Plurality, including distinguishing between singular and collective readings. Definiteness - overlaps with logical form and reference. Topic/focus Sense tags for nouns and more fine-grained verb senses
2) Things that it would be very helpful to add that could possibly be done manually during token alignment.
Marking elided VP's, implicit arguments, predicates and relations that have been made explicit as "inferred."
3) Something we are currently doing that needs to be fixed.
logical form - see separate issue by Tim O'Gorman.
Potential pilots on some of these topics:
1) Distinguishing "inferred" relations 2) Implicit arguments 3) Annotating scope, negation and quantifiers using "negation" shared task guidelines as a separate layer