some linguistically odd head choices

[jeisner]

[item from the old TO-DO file dated 2002-04-07]

In

(NP (NP (DT the) (NN computer) (NN language) ) (VP (VBN called) (S (NP-SBJ (-NONE- *) ) (NP-PRD (NNP UNIX) ))))

we currently get Unix as the head of an S. Yuck! Many nouns show up as heads of S, in exactly this configuration - the superficial object of"called.

SBAR -> @ WHNP S/NP, e.g., "in which we trusted," is currently headed by "in". Should probably be headed by the WHNP, with "in" marked as moved, as in "which we trusted in."

[jeisner]

[item from the old TO-DO file dated 2002-04-07]

Copular VPs should probably be headed by the predicate part. Think of the copula as a function word feature that turns an adjective or noun into a VP by giving it appropriate inflectional features (not necessarily tense, though). This should help with tough-movement.

[jeisner]

[item from the old TO-DO file dated 2002-04-07]

Munge corpus to let "less" and "than" be sisters, or parent and child, in "less X than Y". (Currently, X is chosen as the head child, so "less" and "than" can't see each other.) Similarly for other paired comparatives.

[jeisner]

[item from the old TO-DO file dated 2002-04-07]

• When the head of a constituent is a function word (preposition or complementizer), consider splicing the subconstituents into the parent, so that the arguments of the function word become arguments of the parent. This should also be done for verb groups -- where I chose to mark "to have slept" as headed by "slept," I should have marked it as headed by "to." Then under this convention, "to," "have," and "slept" would all have been generated by the parent (say, "seemed to have slept").
  • This does lose a little bit of what looks like phrase structure, because, for example, some PPs also contain adverbial modifiers: "from 9.45% a week earlier" (or, "in October"), "rather than another industry," etc.
    • But that should be handleable in terms of edit ops. Just insert the modifier next to the preposition or its object ... ?
    • Think of whole generative model as successive edits?? Start with root word, add dependents, etc. ... flesh out context-free frame that way.
      • Oh - that suggests that a sequence of edit ops can't necessarily be modeled as a transducer or FSA - we have to be careful that it's not allowed to recurse, e.g., rewrite X as AXB and repeat. Rather like models of phonology.
  • Some desirable benefits: "information on whether ..." should allow a correlation between "whether" and "information," since "whether" isn't a typical object of "on"!
    • "from X to Y" is currently marked in Treebank as PP -> PP PP, and in flattened version as from NP PP. We might prefer "from X to Y" to all be kids of the verb.
  • Much better -- use conventional heads, but when inserting lexemes, treat function words as transparent. This does mean that we can't use the n^3 algorithm, though.
    • Well, this isn't so good, because "want John to go" has a problem: if "to go" is headed by "to," then we can't make "John" the subject of "go" without crossing links.