📄 NP-External Arguments: A Study of Argument Sharing in English

Summary

The paper looks at a gap in predicate-argument annotation for nouns. Nominal predicates such as walk, decision or attack often have arguments that sit outside the noun phrase the noun heads. In John took a walk, John is the walker, but syntactically he is the subject of took. The authors work within the NomBank effort, which annotates noun arguments in the Wall Street Journal portion of the Penn Treebank, complementing PropBank for verbs. They catalogue the syntactic environments where these NP-external arguments appear, chiefly support-verb constructions and related configurations in which a governing word passes its argument down to the noun. They argue that such argument sharing is lexically conditioned, so the relevant combinations behave like multiword expressions and should be recorded in lexical resources rather than left to general rules. The result is a classification and annotation policy for these cases. It matters because systems that extract who-did-what from nominalizations miss many participants unless they follow these sharing paths.

Key points

• The paper was published in the proceedings of the second ACL-hosted workshop on multiword expressions, which focused on integrating their processing, in July 2004, pp. 96–103.
• Its context is NomBank: argument annotation for nouns over the Penn Treebank WSJ corpus, aligned with PropBank roles.
• The central phenomenon is nominal arguments realized outside the noun's own NP. Support verbs (take a walk, make a decision) are the prime case.
• Argument sharing is treated as a property of particular word combinations, which links it to multiword-expression research and to lexicons such as NOMLEX.
• Caveat: the supplied PDF extraction was almost entirely unreadable. Specific counts and fine-grained category details could not be verified from it, so none are reported here.

Interesting ideas

• I1. A support verb adds little meaning of its own and mainly hands its subject (or another argument) to the event noun, so the noun is treated as the real predicate.
• I2. Argument sharing can pass through a chain of governing words (a verb, then a noun, then another noun), with each link licensing the transfer.
• I3. Whether a combination shares arguments is recorded lexically, as with MWEs, rather than guessed from syntax alone.
• I4. Annotators mark noun arguments outside the noun's syntactic constituent, so the role layer is deliberately decoupled from tree constituency.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): When building frames, detect support verb + event noun pairs and fill the noun's VerbNet-style roles from the verb's syntactic arguments, instead of creating a vacuous "take" frame.
• I2 (frames): Follow UD paths such as nsubj → obj → nmod across licensed links to find a noun's agent, and log each step so the inferred role stays explainable.
• I3 (world): Store a compiled list of support-verb/noun pairings in the knowledge layer next to the verb classes, so frame building looks up a licence rather than applying a blanket heuristic.
• I4 (learn): Use support-verb sentences as targeted evaluation items and negative controls, checking that the agent of walk is recovered in took a walk but not wrongly assigned in non-support uses of take.