Add clause types?

[nschneid]

CGEL defines categories of clauses not reflected in the trees. I will attempt to summarize them here.

Would it be worth adding these subtypes in the trees? I bet it could be done with rules. (Not suggesting we do this right now, leaving this note to consider later)

Top level of the taxonomy

• Main clause (ch. 10)
• Subordinate clause (ch. 11)
  • finite
    • content
    • relative (ch. 12): these are distinguished already as "Clause_rel"
    • comparative (ch. 13)
  • non-finite (ch. 14)
    • infinitival
    • participial

Here are the subtypes of main clauses, some of which are also subtypes of content clauses ([+content]):

Main clauses: Major types

• declarative [+content]
  • Not just clauses that are complements of verbs: IIUC the adjunct in "Wake me when the sun comes up" is analyzed as a PP whose complement is a declarative subordinate content clause. This is a typical advcl in UD.
• closed interrogative [+content or infinitival, p. 985]
• open interrogative [+content or infinitival, p. 985]
• exclamative [+content]
• imperative

Main clauses: Minor types (pp. 944–945)

• optative e.g. subjunctive
• clause with subordinate form
• conditional fragment
• verbless directive
• parallel structure

Also

We would presumably need

• fragment

for corpus sentences that are not complete grammatical utterances.

[nschneid]

One approach is described in https://github.com/nert-nlp/cgel/blob/main/datasets/twitter_parsed/README.md#clause-types with heuristics in https://github.com/nert-nlp/cgel/blob/main/clausetype.py. The main features are ClauseLevel (main or subordinate), ClauseType, and ClauseFinite=Yes. Because content clauses intersect with some of the main clause types, this results in combinations like ClauseLevel=Sub|ClauseType=Content,Decl.

Another way to carve things up would be to borrow the Stype feature (for declarative, imperative, interrogative.open, interrogative.closed, exclamative, interjective...) and add SubClType for just the subordinate clauses: content, relative, imperative, infinitival, participial. Thus no need for mixed values except for infinitival relatives. And then have features to express clause-level analytic tense/voice properties not already covered by the predicate (e.g., finite, perfect, progressive).

[BrettRey]

I see no benefit of including the finite/non-finite distinction. Also, I don't see a need for fragments. A conditional fragment, for instance, is just a PP. There's no need for clause to be the top node. Similarly, a verbless directive like on to the next step is just a PP. The parallel structure ones might be worth a label, but I'd put them In with the nonce constructions. I don't want to call them clauses.

[nschneid]

Fragments: What do you propose we do for conversational sentences/utterances where some material that would normally complete the clause is omitted? One example is pro-drop: "Kinda want to...", "Got any ideas?". Or: "Back in a minute" where both the subject and verb are understood but the two overt phrases don't form a standard constituent on their own. I tentatively marked those as fragments in combination with the clause type that is understood (Decl or ClosedInt). But perhaps this notion of fragment is really an utterance-level rather than clause-level phenomenon.

[BrettRey]

• Kinda want to could be a clause or a VP
• Got any ideas could be a clause or a VP
• Back in a minute PP

Yes, utterance level.