This page describes how to annotate English sentences with AMR graphs to check the output of your parser or create more training data (perhaps in a new domain).

• Read the guidelines to learn how AMR works: AMR Guidelines
• Watch the video on how to use the AMR editor: AMR editor video (more videos)
• To actually annotate:
  • Go to AMR editor
  • First, decide the “main” event/verb in the sentence and put that at the top. If sentence is a conjunction of main verbs or is actually a multi-sentence then a conjunction or multi-sentence concept goes at the top.
  • You'll want to proceed top-down in the AMR. For each content word in the AMR, decide if it's an event, thing or property. Add the lemma to the AMR graph using the editor. If it is an event concept, then click on it and a pop-up will open with the PropBank frame file. Click on the definition you want to use. You'll need to look at the definitions of the arguments of the event to correctly annotate them.
  • There are many linguistic phenomena that are not covered in the annotation guidelines. If you don't know how to handle a certain linguistic phenomena, there are two approaches you can employ:
    • Look up the phenomena in the AMR Dict and decide what to do
    • Search in an AMR corpus to see how the other annotators have handled the phenomena (sometimes annotations may be inconsistent, and this should be resolved through discussion)
  • If you find that there is a reoccurring phenomena in your domain that isn't handled well in AMR or PropBank, you may need to add more predicates. To do this, you can expand the PropBank lexicon to include predicates in your domain or add more core AMR predicates. Examples: Spatial AMR, Dialogue-AMR
  • If you find a problem with the AMR data, you can submit a bug report
• To annotate non-English sentences, you'll need a PropBank-style resource of event predicates in that language.

• Abstract Meaning Representation