Improve AMR alignment / ConceptNet output

[namednil]

While our results using ConceptNet (+ CoreNLP) are still ok, the data seems to be much more noisy, resulting in a drop in supertagging accuracy and LAS.

What ways are there to improve this?

• can we count calls of ConceptNet with input, output pairs and see if there are frequent calls that have unexpected or suboptimal outputs and fix the most frequent ones by handwritten rules?

Is that feasible in a reasonable amount of time?

[alexanderkoller]

It is easy to count the calls, but the problem is that ConceptNet only provides options for the aligner. Which ones the aligner ends up choosing is much harder to count, and I don't have a clear idea how to do it.

Can we meet tomorrow (Mon) and talk about it?

[namednil]

I see.

Can we meet tomorrow (Mon) and talk about it?

Sure, I'm available from 16:00 (I have an exam before).

[namednil]

Would it help I compiled a small corpus where our decomposition got worse?