VC constraint on free word order grammar with auxiliaries

I have 2 identical inferred grammar with the only difference of word order. One of them is v2, the other is free.

section=word-order
word-order=free
has-dets=yes
noun-det-order=det-noun
has-aux=yes
aux-comp-order=before
aux-comp=v
multiple-aux=yes

The v2 one parses this:
Bardba murndu
run-FUT 1.DU.INC.S-FUT
we will run

The free word order one fails in interactive unification with the comp-head rule. The VC constraint for one is plus, the other is minus. Are these word order choices supported my the Grammar Matrix and if so, what’s the expected result? Why would a sentence parses in the v2 grammar but not the free word order?

Both v2 and free word orders are partially supported in the Grammar Matrix. This means that you can expect a grammar which loads and parses some sentences with both these orders, but the range of sentences which parse may be more limited compared to other orders and you will likely encounter more bugs with free and v2 than you will with other orders.

As for your specific unification failure: the way to debug this is to set up a breakpoint in customize.py at line 617 (word_order.customize_word_order(mylang, ch, rules)) and then step through the execution and see what happens and why. Without doing that, I can see that [ VC + ] is set for head-complement rules by the following lines, for example:

        if (aux == 'vini-vc' and aux == 'vo-auxv' ) or wo == 'free':
            mylang.add('head-comp-phrase := [ HEAD-DTR.SYNSEM.LOCAL.CAT.VC + ].')
        if (aux == 'vfin-vc' and aux == 'ov-vaux') or wo == 'free':
            mylang.add('comp-head-phrase := [ HEAD-DTR.SYNSEM.LOCAL.CAT.VC + ].')

There may be other places which do this, too.

One way to reconstruct the logic behind these lines would be to look at the regression tests involving auxiliaries and verb clusters (vc). You can tell from the regression test names.

So if your combination of choices is valid, then you need to modify the logic in word_order.py such that your combination of choices is accommodated but all of the regression tests still pass.

1 Like