Infinitival relatives

[nschneid]

Today we decided that UD English treebanks should acknowledge that some adnominal infinitival clauses are relative clauses, as described here: https://universaldependencies.org/en/dep/acl-relcl.html#infinitival-relatives

It will take some effort to revise the annotations, but I suspect we may be able to come up with some heuristics to speed it along.

[rueter]

@nschneid would you use this same relation for the noun in:

Joe has a house for sale.

And the verb in

Joe has a house to sell.

[nschneid]

Good question. Thinking in terms of the meaning, there is an interesting construction with "have" + infinitivals (e.g. "I have a plane to catch", which does not entail "I have a plane"). But for UD purposes, I suppose I would treat "a house to sell" as a constituent, in which case it is an infinitival relative, so acl:relcl. "Joe has a house for sale" definitely does NOT contain a relative clause, and arguably should be parsed with "for sale" as modifying the verb. Maybe xcomp, like "the room has a window open"?

[nschneid]

We also have to consider tough constructions. In "This is a tough problem to solve" as well as "This problem is tough to solve", the infinitival is considered to be licensed by the adjective, not the noun, so it is not a relative clause. (CGEL calls it a "hollow infinitival clause".) Cf. UniversalDependencies/docs#308, though I'm not sure where that landed—it seems to have morphed into a discussion of the argument/adjunct distinction.

[amir-zeldes]

Thinking in terms of the meaning, there is an interesting construction with "have" + infinitivals (e.g. "I have a plane to catch", which does not entail "I have a plane"). But for UD purposes, I suppose I would treat "a house to sell" as a constituent, in which case it is an infinitival relative, so acl:relcl.

Are you saying that "a plane to catch" is also acl:relcl?

[nschneid]

Yeah I am leaning that way.

[nschneid]

Also need a policy for sufficiency constructions: We don't have the money/wherewithal to do that, enough money to do that. Some borderline-sufficiency cases like "willingness". These don't seem like relative clauses.

A start at filtering out head nouns that license infinitival complements: http://universal.grew.fr/?custom=63542dcce03ed Many of the filtered-out heads are nominalized and/or modal.

Some tricky cases are things like an opportunity, reason, or way to do something (the noun can be construed as a circumstance of the infinitival clause event, thus an adverbial relative).

[amir-zeldes]

All of these borderline cases make me think we should just leave it as acl except in the cases where there is an explicit relative pronoun, or possibly a stranded preposition explicitly indicating a relative. I would say these can be acl:relcl:

• A place in which to live/acl:relcl
• A place to live/acl:relcl in

But if we also allow "a place to live", then I don't see where it ends, because you can virtually always find a relative paraphrase:

• The money to do that -> The money with which to do that
• A way to win -> A way by which to win

Once things get subjective and we allow explicitation by an unexpressed preposition, I think we are leaving syntax in favor of the semantics implied by the lexical items involved, and I worry that we'll see annotator disagreements (not to mention in practical terms, I wouldn't know how to introduce this distinction into existing data without a manual pass on all infinitival acls)

[nschneid]

PTB guidelines for infinitival relatives: pp. 64, 238-240

Since EWT has PTB annotation we can just write a script to extract them.

[nschneid]

Attempt at an ANNIS query (rules out tough-constructions where the infinitival is within an ADJP). There is a (possibly null) relative element with category WHNP, WHADVP, or WHPP, useful for constructing the enhanced dependencies.