More SemLink 2 questions

[chaitanyamalaviya]

Hi, I had a few more questions about SemLink-2.

1) Why are there much fewer entries in SemLink2 pb-vn mappings? It seems like that a large number of these entries that existed in SemLink 1.2.2 are just missing in SemLink-2. More than 3900 entries are just mapped to empty lists. For eg, "force", "fortify" and many others are mapped to empty lists. Is this intentional?

2) Some verb lemmas have different PropBank rolesets and the same VN-class but the argument mappings only exist for one of these cases in SemLink 1.2.2. Why is this the case? Here is an example:

3) For any given sentence, would it reasonable to map PropBank arguments to VerbNet arguments using the pb-vn type map in SemLink (after having retrieved the correct verb sense)?

Thanks for your help!

[kevincstowe]

1. This was an error in the generation of the pb-vn mappings (which is in the propbank.py script). I'm not exactly sure what it was doing, but it was dropping a number of instances. I did some rewriting and it appears to be working correctly - the pb-vn2.json file should now be more complete. There will be cases where mappings are removed (they appear incorrect and we can't automatically fix them), but there should be much fewer now.
2. I don't know what this would be - some artifact of the semlink 1.2 process. It should be fixedin Semlink 2 - let me know if you see this issue and I'll look into it. Hopefully with the above fixes you won't need to use 1.2.2 anymore!
3. Yes, I think that's our intended use case: find the correct PB sense, map to VN using the pb-vn2.json, which should then allow for argument mapping from PB roles to VN thematics roles.

[chaitanyamalaviya]

Thanks for fixing the mappings. As for 2., I think this problem still exists unless I'm missing something. Here is an example: {"abandon.01": {"vn_class": "51.2", "ARG0": "theme"}, "abandon.02": {"vn_class": "51.2"}. I wonder if abandon.02 should also have the "ARG0": "theme" mapping?

[kevincstowe]

Okay I see. In PropBank, they actually are only marking these mappings on the one roleset. So for abandon.01, we get:

    <roles>
    <role descr="abandoner" f="ppt" n="0">
    <vnrole vncls="51.2" vntheta="theme"/>
    </role>
    <role descr="thing abandoned, left behind" f="dir" n="1"/>
    <role descr="attribute of arg1" f="prd" n="2"/>
    </roles>

while abandon.02 has the same VNC mapping doesn't have the role mapping:

    <roles>
    <role descr="abandoner" f="pag" n="0"/>
    <role descr="thing abandoned, left behind" f="ppt" n="1"/>
    <role descr="preferred item" f="gol" n="2"/>
    </roles>

This seems wierd, but may be intentional, giving functionality for a very specific case: two rolesets map to the same VN class, but their arguments don't map to the same roles. This seems strange, though. If this functionality isn't intended, I can merge all the role mappings for a given roleset/vnc pair together.

[MarthaSPalmer]

Kevin, it seems like a mistake to me, that should be fixed.

Martha

On Aug 9, 2021, at 7:12 AM, Kevin Stowe @.***> wrote:



Okay I see. In PropBank, they actually are only marking these mappings on the one roleset. So for abandon.01, we get:

while abandon.02 has the same VNC mapping doesn't have the role mapping:

This seems wierd, but may be intentional, giving functionality for a very specific case: two rolesets map to the same VN class, but their arguments don't map to the same roles. This seems strange, though. If this functionality isn't intended, I can merge all the role mappings for a given roleset/vnc pair together.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/cu-clear/semlink/issues/4#issuecomment-895210360, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ABB327WK77KISKOBTGEMRW3T37HYNANCNFSM466LEJYA. Triage notifications on the go with GitHub Mobile for iOShttps://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Androidhttps://play.google.com/store/apps/details?id=com.github.android&utm_campaign=notification-email.

[chaitanyamalaviya]

Thanks for taking a look! It appears that in the latest version of pb-vn2.json, ~60.9% entries are empty (6510 of 10687 entries). Just want to make sure if this is intentional?

Thanks again!

[kevincstowe]

I've implemented the above, that now two rolesets w/ the same class mapping will now have the same role mappings as well.

I had initially included all possible verbs in PB along with empty entries, for some reason, but I've now taken them out. There are now just the 4177 valid entries.

[chaitanyamalaviya]

Thanks for the changes, Kevin! Just want to mention that of the 4979 entries in the latest version (counting all the mapped vn-classes for any PB role), there are actually 924 empty entries. Is this always because there is no suitable mapping? For eg, abdicate.01 has no mappings, but I would think that ARG0 could just be Agent for this class.