SOV grammars with wh-fronting and clause-initial topic/focus

[olzama]

...do not work because of the subj-head rule constrained to be SLASH-empty:

    if wo in ['sov', 'svo', 'osv', 'ovs', 'vso', 'vos']:
        if pos == 'clause-initial' or infostr_type == 'topic-first':
            if wo == 'sov':
                ph_types['comp-head-phrase'] = 'comp-head-phrase := [ SYNSEM.L-PERIPH -, HEAD-DTR.SYNSEM.L-PERIPH - ].'
                ph_types['subj-head-phrase'] = 'subj-head-phrase := [ SYNSEM [ LOCAL.CAT.MC +, NON-LOCAL.SLASH.LIST < > ], ' \
                                               'HEAD-DTR.SYNSEM.L-PERIPH - ].'

the above is currently lines 478--484 in information_structure.py. This makes wh-fronting impossible in most cases: if the subject-head rule is SLASH-empty, this means a wh-complement cannot be extracted, from an SOV sentence.

This is about the general interaction of the two libraries (information_structure.py and wh-ques.py), which do a lot of similar work (fronting constituents), and that work should be better shared and consolidated between them (and tested!)

[olzama]

https://github.com/delph-in/matrix/issues/533

[emilymbender]

I'm trying to work out what work [ SLASH.LIST < > ] is doing there and can't quite see it. I can imagine that in this configuration we would want to block ambiguity in SOV sentences between "unmarked" SOV and subject-topicalization. But: putting [ SLASH.LIST < > ] on the head-subject rule wouldn't seem to help with that, since the head-subject rule wouldn't fire in the subject-topicalization analysis.

Again, I'd be curious to see what changes in the regression tests if this is removed.

[olzama]

OK, so, for this one, removing the constraint actually does not lead to any regressions (apart from in Paresi-Haliti, but those include improvements, in addition to some ambiguity which I would take, at this point), but I suspect this is because a bunch of information structure tests are currently being skipped.

[olzama]

I suppose we could say that, since those tests represented unfinished work, whether or not they fail due to this change is less relevant than the fact that this change lets us have those additional Paresi-Haliti parses?

[emilymbender]

Looking over at the skipped tests, it seems they all involve affixes marking information structure and so likely aren't the main ones that would be affected by removing this constraint on the SLASH in head-subj phrases.

So, I think it's safe to go down this path (removing the SLASH constraint, keeping LIGHT) as illuminated by the other tests.