UMIACS Computational Linguistics Colloquium, May 8, 2000

Predicting functional control from lexical semantics: Control verbs in Natural Language Processing


Mari Broman Olsen


Microsoft


UMIACS Computational Linguistics Colloquium

May 8, 2000,
1:30pm, AVW Room 2120


It has long been recognized that verbs like try and persuade obligatorily "control" the subject of their participial complements (1), and that verbs like ban and support do not (2).

  (1) a.  Kim tried smoking cigarettes.
      b.  Kim persuaded Mary to smoke cigarettes.
      c.  Kim continued being a member of the Communist Party. 
          [activity reading] 
      d.  Kim kept smoking.

  (2) a.  Kim banned smoking cigarettes.
      b.  Kim supported being a member of the Communist Party.
This contrast has traditionally been couched in terms of syntactic complementation patterns. That is, control verbs select an open complement (usually an XCOMP in LFG) and specify the control relation via an equational statement in their lexical entry (cf. Bresnan 1982, Mohanan 1983, inter alia). Verbs whose complements take arbitrary subjects select some sort of closed complement (usually NP, but in principle some sort of closed verbal complement), from which the non-control is taken to follow. Reifying the behavior of the verb classes in (1) and (2) by making a distinction in the syntactic category of their complements is, at best, not particularly enlightening. A better explanation hinges on the lexical semantic properties of typical complements to these types of verbs. Specifically, open complements introduce distinct events (as opposed to complements which are state-denoting or which do not introduce distinct event variables) and therefore resist being controlled in a variety of contexts; this resistance has been grammaticized in the case of the ban-type verbs, due to the nature of the verbs' meaning. That events resist control is evidenced in part by the existence of a distinct subclass of control verbs in which a confluence of both lexical semantic and syntactic considerations determine whether control is obligatory (as with other control verbs in (1)), or absent (as with verbs like (2)).

In these cases, the nature of the complement drives the nature of the control, and this relationship between the lexical semantic of the complement and the possibilities of control extends to the obligatory control verbs, as well. The relevance of the state/event distinction is well-known for main verbs; this data establishes it for variable control patterns in complements. I further demonstrate how the relevant properties may be implemented in an existing NLP system.


For the colloquium series schedule, see the UMD Computational Linguistics Colloquium Series web page at http://umiacs.umd.edu/~resnik/cl_colloquium/. If you are interested in meeting with the speaker, please contact Philip Resnik (resnik@umiacs.umd.edu).