There are 8,283 instances of :degree in the AMR release. However, most of these are used as "stand-alone" modifiers where there is no comparison or other arguments of have-degree-91 evoked (especially "so, very, further, really." For example:
Given this, I suggest that we manually re-annotate all instances of :compared-to (1632), :degree-of, and :degree combined with :ord, and/or :score-on-scale, and then leave the other :degree cases as-is, under the assumption that they are correct and represent the non-reified version of :have-degree-91. We could then randomly sample some of the stand-alone :degree cases we are planning to leave as-is to validate this assumption.
I don't think any heuristics or mapping from old compared-to structures to new have-degree-91 are possible for adequately retrofitting any of the non-stand-alone instances (semi-)automatically. This is in part because there is overlap in these cases with what would now be :have-quant-91, and we also shifted what we think about as the point of comparison in comparisons and superlatives, for example (old on top, new on bottom):
There WILL be error cases that we overlook. Some of these stem from genuine errors in past annotations where :degree is used instead of :quant or :compared-to is missing even though there's a fairly clear comparison being invoked, for example:
Furthermore, if we don't inspect all instances of :degree, then we will be unlikely to find past instances of degree-consequence constructions (It's too wide to fit on my wrist) or The X-er The Yer constructions (The longer I'm around him, the more miserable I'll be). (These may be detectable by looking for certain lexical patterns, and this might be a fun project for a student?) These are relatively rare, and I think we can just look for them going forward. I think the benefit of assuming most stand-alone :degree instances are correct outweighs the potential costs.
There are 8,283 instances of :degree in the AMR release. However, most of these are used as "stand-alone" modifiers where there is no comparison or other arguments of have-degree-91 evoked (especially "so, very, further, really." For example:
Given this, I suggest that we manually re-annotate all instances of :compared-to (1632), :degree-of, and :degree combined with :ord, and/or :score-on-scale, and then leave the other :degree cases as-is, under the assumption that they are correct and represent the non-reified version of :have-degree-91. We could then randomly sample some of the stand-alone :degree cases we are planning to leave as-is to validate this assumption.
I don't think any heuristics or mapping from old compared-to structures to new have-degree-91 are possible for adequately retrofitting any of the non-stand-alone instances (semi-)automatically. This is in part because there is overlap in these cases with what would now be :have-quant-91, and we also shifted what we think about as the point of comparison in comparisons and superlatives, for example (old on top, new on bottom):
There WILL be error cases that we overlook. Some of these stem from genuine errors in past annotations where :degree is used instead of :quant or :compared-to is missing even though there's a fairly clear comparison being invoked, for example:
Furthermore, if we don't inspect all instances of :degree, then we will be unlikely to find past instances of degree-consequence constructions (It's too wide to fit on my wrist) or The X-er The Yer constructions (The longer I'm around him, the more miserable I'll be). (These may be detectable by looking for certain lexical patterns, and this might be a fun project for a student?) These are relatively rare, and I think we can just look for them going forward. I think the benefit of assuming most stand-alone :degree instances are correct outweighs the potential costs.