Standardization of the Complement Adjunct Distinciton

Meyers, Macleod, Grishman 1994

Meyers et al propose a set of guidelines for consistently distinguishing complements from adjuncts. The goal is to find the complements of verbs, as a step in building a large lexicon (COMLEX). To evaluate their succes, they measured the inter-annotator agreement for annotating the complements in a text.

I find many of their criteria suspect (see below), but they do get reasonable inter-annotator agreement. So they're consitantly annotating something -- it's just not clear if it's something that we want to have annotated.

The guidelines define three set of criteria, which should be consulted in order to decide whether a given candidate phrase is a complement or an adjunct:

  • If the phrase satisfies any of the sufficient complement-hood criteria, then it is a complement.
  • Otherwise, if the phrase satisfies any of the adjunct-hood criteria, then it's an adjunct.
  • Otherwise, if the phrase satisfies any of the complement-hood rules of thumb, then it's a complement.

In practice, they had the annotators check all criteria for each phrase, and flag phrases that seemed to satisfy conflicting criteria. They also instructed the annotators about several special cases, where the standard decision-process should not be applied (e.g., instrumental PPs). It was unclear how many of these "special cases" they defined.

Sufficient complement-hood criteria.

If any of these apply, then it's a complement.

  • Is the phrase obligatory for a particular sense of the verb? E.g., in "John woke up Mary," Mary is obligatory, since "wake up" in "John woke up" is a different sense.
  • Can the phrase be the subject of a passive? (Or stranded by a pseudo-passive?)
  • Does the phrase have an "argument (=complement) theta role." They didn't provide a complete list of argument theta roles, but they include: theme, source, goal, patient, recipient, experience, proposition, question.
  • If the phrase can be omitted, is it implied when omitted?
  • Does the verb impose selectional restrictions on it? Meyers et al propose 2 tests for this:
    • Does the verb influence the meaning of the phrase if a "semantically unspecified phrase" (such as "it") is used?
    • If there is a semantic anomaly, is it resolved by changing the meaning of the phrase (and not the verb)?

Adjunct-hood criteria.

If any of these apply (and no previous criteria apply) then it's an adjunct.

  • Does the phrase occur with most verbs with roughly the same frequency and meaning?
  • Is it a typical adjunct phrase? These include purpose clauses; clauses headed by "before," "after," "while," "because," "although," "if," or "by"; instrumental "with" phrases; "by means of" phrases; and benefactive, place, manner, and time adverbials or PPs.
  • Does the phrase impose a selectional restriction on the verb?
  • Is it an AdvP/PP which can be questioned with "why" or "how"?
  • Can it be naturally fronted (without extra stress)?
  • Is it unable to violate island constraints?

Complement-hood rules of thumb.

If any of these apply (and no previous criteria apply) then it's a complement.

  • Does the phrase co-occur frequently with its verb?
  • Is it a typical complement? These include NPs, PPs headed by "to," and clauses (other than relatives, "whether," and "if" clauses).
  • Can it participate in "complement alternations"? They didn't provide a complete list, but this includes dative and spray/load.
  • Does the phrase occur between a head and a complement?
  • Can it violate island constraints?

Criticism of Criteria.

Many of the criteria proposed by Meyers et al seem troublesome, or outright wrong.

  • Theta roles: It is notoriously difficult to assign theta roles consistantly. Also, it's unclear how we should decide which theta roles are "complement" theta roles and which are "adjunct" roles. E.g., Meyers et al claim that instrumental phrases are adjuncts; but many others consider them complements.
  • Implied when omitted: Some such phrases are clearly adjuncts. E.g., "John ate [somewhere]." Without the location adjunct, it's clearly implied that there was some location where John ate; but we don't want to call it a complement. How do we draw the line between cases like this and cases like "John ate [something]" (which the rule is intended to capture)?
  • Selectional restrictions: Neither of the tests for selectional restrictions that they give seems workable to me.
    • Semantically unspecified phrases: Verbs influence the meaning of both complements and adjuncts. E.g., in "Mary removed the nail with it," clearly "it" is +solid.
    • Semantic anomaly: Semantic anomaly can be resolved either way, depending on context. E.g., in "John teased the book," it's easy to imagine that it's some idiomatic meaning of "tease;" and in "John washed the dishes with a hammer," the first image that comes to my mind is a foam hammer (and not some strange new meaning of "wash").
  • Frequency criteria: these seem vague, and it's unclear how to measure them.


Meyers et al had 4 graduate students annotate 205 sentences, covering 35 verbs. The interannotator agreement was 91%, measured as average inter-annotator recall:


Excluding 44 examples (17% of the examples!) raised the score to 93%. These are fairly good scores for inter-annotator agreement, but again it's difficult to say whether this definition of the argument/complement distinction is the one we want to annotate. In fact, the fact that so many criteria are used, and that they are often in conflict, brings into question whether this is a single distinction, or a conflation of different (but partially overlapping) distinctions.


@misc{ meyers1994,
  author = "Adam M. and C. Macleod and R. Grishman",
  title = "Standardization of the Complement Adjunct Distinction",
  text = "Meyers, Adam, Catherine Macleod and Ralph Grishman (1994). Standardization
    of the Complement Adjunct Distinction, Proteus Project Memorandum 64, Computer
    Science Department, New York University.",
  year = "1994",
  url = "citeseer.nj.nec.com/adam94standardization.html" }