In order to support annotators in improving annotation consistency, the AMR Checker uses a number of lists, including some that I have built or augmented for AMR. Here's the primary such resource of lists:

http://www.isi.edu/~ulf/amr/lib/amr-auto-correction-rules.txt

You are invited to have a look at it and provide feedback.

Many of the entries were extracted from our growing corpus of AMRs. For my decisions (e.g. VERBALIZE vs. DO-NOT-VERBALIZE) I took into consideration of what LDC and other annotators have come up with. The list of 150 mappings from support verb constructions to AMR concepts also greatly benefitted a list that Claire kindly sent me.

Below are a few particular cases where I was not sure, as indicated by a comment-out (#) or a MAYBE- prefix. For each class of advice, I also include a few normal cases in order for you to get a quick sense of the class.

DO-NOT-VERBALIZE professor TO person :ARG0-of profess-01 DO-NOT-VERBALIZE white TO whiten-01

DO-NOT-VERBALIZE relationship TO relate-01

DO-NOT-VERBALIZE suspicious TO suspect-01

DO-NOT-VERBALIZE valuable TO value-01

DO-NOT-MAP international TO internation DO-NOT-MAP friendly TO friend

DO-NOT-MAP official TO officer

DO-NOT-MAP various TO variety

VERBALIZE afraid TO fear-01 VERBALIZE approval TO approve-01

VERBALIZE favorable TO favor-01

VERBALIZE hungry TO hunger-01

VERBALIZE scandalous TO scandalize-01

VERBALIZE speedy TO speed-01

VERBALIZE dinner TO dine-01

MAYBE-VERBALIZE terrorism TO terrorize-01

Why did we decide against a pianist being a piano player?

VERBALIZE drummer TO person :ARG0-of play-01 :ARG1 drum

VERBALIZE pianist TO person :ARG0-of play-01 :ARG1 piano

VERBALIZE trumpeter TO person :ARG0-of play-01 :ARG1 trumpet

DO-NOT-VERBALIZE chairman TO person :ARG0-of chair-01 DO-NOT-VERBALIZE editor TO person :ARG0-of edit-01 DO-NOT-VERBALIZE president TO person :ARG0-of preside-01

Not 100% sure about the following:

MAYBE-VERBALIZE commander TO person :ARG0-of command-02

Corpus usage: military commanders

MAYBE-VERBALIZE captor TO person :ARG0-of capture-01

Corpus usage: hostage taking, kidnapping

MAYBE-VERBALIZE container TO product :ARG0-of contain-01

Corpus usage: shipping containers, container cranes

MAYBE-VERBALIZE carrier TO ship :ARG0-of carry-01

Corpus usage: aircraft carrier

MAYBE-VERBALIZE contractor TO person :ARG0-of contract-02

Corpus usage: defense contractor, insurance contractor

MAYBE-VERBALIZE devotee TO person :ARG0-of devote-01

Corpus usage: Buddhism, charities

MAYBE-VERBALIZE terrorist TO person :ARG0-of terrorize-01

MAP Kirghizia TO Kyrgyzstan

MAP natural TO nature

MAP worthy TO worth

SPLIT moonlight TO light :source moon

SPLIT moonlight TO light-02 :ARG0 moon

Hi Ulf,

Less urgent for the current delivery. A couple of questions:

If we have other suggestions for the Checker lists, should we communicate by e-mail or list them under this issue? Some cases for mappings:
- right to > entitle-01
- "in return" > exchange-01
- in accordance with > comply-01
- is due to > suppose-02
- (so) what if > supose-01
- opposition > oppose-01
- vision > envision-01
- fraud > defraud-01

We also have some suggestions for types of issues that the Checker might help us spot and currently does not:

Extra content in AMR as opposed to source sentence
Check for :op "-" (as based on our decision, we will not represent the dash on a separate :op, however, when the annotators copy the NE, due probably to tokenization, the hyphenated word is not represented on a single :op and they overlook this).
Isolated constructions with no substructure but which should have a substructure (leftovers due to hesitations, which the annotators don't corrrect) e.g. (a / and) --> not followed by anything
check for coreference in case of repeated NEs It would be nice if the checker catches cases of co-reference that have not been annotated correctly, i.e. by using the variable. E.g.
Thai Government has indicated Thailand will join the pact .
check for all nationalities to be mapped to the country name e.g.
1. Would it be possible to have all these rules that you have created for the checker implemented as suggestions? These would be of great help to the annotators and would theoretically prevent them from making mistakes.
2. We have noticed some inconsistencies between the tool suggestions and the checker alarms.

One such example is "suspicious" (the tool suggestion is to use "suspect-01" . If the annotator uses "suspect-01" the Checker will identify this as a word not annotated.) Should we collect these instances and report them?

amrisi / amr-guidelines

Review of AMR Checker lists #27