Nerd Industries: Stuart Robinson's blog

WordNet may be the most widely used resource in computational linguistics. It's a great tool but far from perfect. Today I was reminded of this fact when trying to make use of its subcategorization frames. I started out trying to use the NLTK's interface to WordNet.

The NLTK's documentation of its WordNet interface provides a simple example of how to extract the subcat frames for the verb fly with some Python code:

>>> import nltk
>>> w = nltk.wordnet.V['fly']
>>> s = w.synsets()[0]
>>> s.verbFrameStrings
['Something fly', 'Somebody fly', 'Something is flying PP', 'Somebody fly PP']

So far, so good. But when I tried using this technique to extract the subcat frames for a longish list of verbs, things went south. I'll illustrate with a colorful verb popularized by MTV:

>>> import nltk
>>> w = nltk.wordnet.V['pimp']
>>> s = w.synsets()[0]
>>> s.verbFrameStrings
['Somebody pander']

I thought I would get the subcat frames for the verb pimp but I got the subcat frames for a different verb, pander. What's going on here?

The problem is that WordNet provides a subcat frame for a synset, rather than a particular verb:

>>> nltk.wordnet.V['pimp'].synsets()   
[{verb: pander, pimp, procure}]

Although only one subcategorization frame (intransitive) is provided for pander, pimp, and procure, there are others, which aren't shared by all three verbs:

• The PR firm is pimping (*pandering/*procuring) out their client's product.
• John McCain is pandering (*pimping/*procuring) to religious conservatives.
• I want MTV to pimp (*pander/*procure) my ride.

These examples show that the verbs in a synset don't necessarily share the same subcat frames. As a result, a system that provides subcat frames for synsets, rather than particular verbs, is going to be of limited utility.