In Armenian, there is no marking of questions except for intonation.
Interestingly, this is shown in the writing system with a question
intonation diacritic that goes above the stressed vowel in the
question word -- however, that would break the LKB's morphology (and
how), so I didn't try to include it.  The result is that all
grammatical sentences will now parse twice: once as a statement and
once as a question.

In embedded clauses, the facts in my limited references are murkier,
and I'm honestly not confident that what I've implemented actually
reflects the real language.  However, as best I can tell, here are the
facts.  Embedded clauses come after the verb, and they're marked with
one of two complementizers: "the" (that's an aspirated [t]) or "or".
If "or" is used, the embedded clause must be a statement, not a
question; if "the" is used, it may be either.  The embedded clause may
be in the subjunctive for some verbs -- this marks uncertainty or
doubt.  The subjunctive in Armenian is marked by the absence of the
usual "ke" particle, which I implemented by adding mood to my
ke-marking rule, and then splitting that rule into an inflecting
indicative rule and a non-inflecting subjunctive rule.  Matrix clauses
can also be in the subjunctive and I have previously-ungrammatical
examples without the particle in my test.all file, but matrix
subjunctives are not necessary for the assignment, so I restricted
matrix clauses to be indicative in roots.tdl to make the coverage stay
the same.

See test.items for glossed examples (recalling that determiners come
after nouns, and the ke- particle is present in the indicative mood).

There is a cross classification here: the verb "gitnar" (to know)
takes clauses introduced by either "the" or "or", but which must be in
the indicative.  The verb "xntrer" (to ask) only takes clauses that
are introduced by "the", but they can be either indicative or
subjunctive.  To handle this, I made a type clausal-verb-lex that has
all the information in common to both verbs.  The subtype
nom-clausal-verb-lex requires a nominative subject -- I forgot to do
that at first, and all my sentence were parsing twice for each noun in
the nominative or accusative, since they're identical for nouns.  The
subtype nom-indic-clausal-verb-lex requires an indicative clause
(verbs like "gitnar"), and the subtype nom-interrog-clausal-verb-lex
takes only interrogative clauses (verbs like "xntrer").

For the complementizers, I defined comp-lex, which takes one argument,
a clause with [ FORM fin ], but which doesn't specify a question or a
statement.  The complementizer "the" is a bare comp-lex.  I created a
subtype called decl-comp-lex that requires the clause to be a
proposition; the complementizer "or" is one of these.

I added labels for C and CP so my trees would look reasonable.  Also,
since I was getting dual parses of most clauses, I split the sentence
label into two labels: the S label appears if a sentence is a
propsition; on all other sentences (questions), the S? label appears.

This handled the syntax, and also the semantics for matrix clauses,
but when I parsed sentences with embedded clauses, none of the
semantics was visible from the embedded clause.  I looked at
that-comp-lex and whether-comp-lex in the lab's English example.
Since the Armenian complementizers (especially "the") have to be more
flexible about the kind of clause they take, I used that-comp-lex as a
model, merging it into my comp-lex.  With some fiddling, this got the
semantics working for embedded clauses.

However, I was still seeing overgeneration in several different
places.  In particular, I was seeing overgeneration like I'd seen in
previous labs, where the verb form was either correct (person-number
inflected) or else in one of the two incorrect non-finite forms
(infinitive and negative participle).  The fix for this was the same
as in that lab: make sure the various head-subj and head-comp rules
require a fully-inflected (KE-MARKED and PN-MARKED) verb, and that the
head-subj rules require a [ FORM fin ] verb.

I also noticed that I expected two parses of every sentence, but I was
only seeing one for optional-subject sentences (or three parses in a
few cases where I expected four).  This was because I had a head-subj
rule to make interrogative sentences, but no corresponding
interrogative head-opt-subj rule.  By analogy with the old optional
subject rule, I added interrogative-head-opt-subj-phrase to fix the
syntax, and it works.  However, the semantics for both regular and
interrogative optional subject sentences aren't right.  After about
and hour of poking at it, I've decided I don't understand, so it
remains broken.