#Embedded Clauses

Subordinate clauses in Yiddish display the same V2 word order as
declarative main clauses and are introduced by the complementizer
'az', which, in some circumstances can be omitted:

		Source: a:233
		Vetted: t
		Judgment: g
		Phenomena: {embedded declarative}
		Zi meynt, az ikh bin meshuge.
		Zi meyn-t, az ikh bin meshuge.
		3SG.NOM.F think-3SG, COMP 1SG.NOM be;1SG crazy. 
		She thinks that I have gone crazy.

		Source: a:233
		Vetted: t
		Judgment: g
		Phenomena: {embedded declarative}
		Zi meynt, ikh bin meshuge.
		Zi meyn-t, ikh bin meshuge.
		3SG.NOM.F think-3SG, COMP 1SG.NOM be;1SG crazy. 
		She thinks I have gone crazy.

Additionally, constituents within the embedded clauses can be
topicalized:

		Source: a:232
		Vetted: t
		Judgment: g
		Phenomena: {embedded declarative}
		Ikh veys, az moyshe kumt morgn.
		Ikh veys, az moyshe kum-t morgn.
		1SG.NOM know;3SG COMP Moyshe come;3G tomorrow.
		I know that Moyshe is coming tomorrow.
 
		Source: a:232
		Vetted: t
		Judgment: g
		Phenomena: {embedded declarative}
		Ikh veys, az morgn moyshe kumt.
		Ikh veys, az morgn moyshe kum-t.
		1SG.NOM know;3G COMP tomorrow Moyshe come;3G.
		I know that it's tomorrow Moyshe is coming.

		Source: a:232
		Vetted: t
		Judgment: g
		Phenomena: {embedded declarative}
		Ikh veys, az es moyshe kumt morgn.
		Ikh veys, az es moyshe kum-t morgn.
		1SG.NOM know;3G COMP 1SG.NOM Moyshe come;3G tomorrow.
		I know that it's Moyshe that's coming tomorrow.

Since subordinate clauses are not yet handled by the Grammar Matrix,
no implementation has of yet been done.

While our reference grammar does not say anything about subordinate
yes-no interrogatives, from looking through the data in the Corpus of
Modern Yiddish it would appear that the interrogative marker 'tsi',
which translates as 'whether' is used as a complementizer instead of
'az' similar, to English. As to the word order in embedded
interrogatives, it appears that both SVO and VSO are possible, with
VSO having the effect of contributing more interrogative force to
the utterance. 

		Source: author based on examples in c
		Vetted: f
		Judgment: g
		Phenomena: {embedded interrogative}
		Ikh freg, tsi iz es efsher
		Ikh freg, tsi iz es efsher
		1SG.NOM ask;3SG, whether;COMP be;3SG 3SG.NOM.N possible.
		I ask whether it is possible?

		Source: author based on examples in c
		Vetted: f
		Judgment: g
		Phenomena: {embedded interrogative}
		Ikh freg, tsi es iz efsher
		Ikh freg, tsi es iz efsher
		1SG.NOM ask;3SG, whether;COMP 3SG.NOM.N be;3SG possible.
		I ask, whether it is possible

		Source: author based on examples in c
		Vetted: f
		Judgment: u
		Phenomena: {embedded interrogative}
		Ikh freg, az iz es efsher
		Ikh freg, az iz es efsher
		1SG.NOM ask;3SG, that;COMP be;3SG 3SG.NOM.N possible.
		I ask whether it is possible.

The subset of verbs in our lexicon that can embed clauses follow:

		ze := trans-verb-lex &
		  [ STEM < "ze" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_see_v_rel" ].

		shrayb := trans-verb-lex &
		  [ STEM < "shrayb" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_write_v_rel" ].
		
		leyen := trans-verb-lex &
		  [ STEM < "leyen" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_read_v_rel" ].
		
		boydek := strict-trans-verb-lex &
		  [ STEM < "boydek" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_check_v_rel" ]. 
		
		ruf := trans-verb-lex &
		  [ STEM < "ruf" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_call_v_rel" ].

		lib := trans-verb-lex &
		  [ STEM < "lib" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_love_v_rel" ].

		lern := trans-verb-lex &
		  [ STEM < "lern" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_learn_v_rel" ].

		farshtey := trans-verb-lex &
		  [ STEM < "farshtey" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_understand_v_rel" ].

		meyn := trans-verb-lex &
		  [ STEM < "meyn" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_believe_v_2_rel" ].
	
		mey := trans-verb-lex &
		  [ STEM < "mey" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_intend_v_1_rel" ].
		
		freg := trans-verb-lex &
		  [ STEM < "freg" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_ask_v_2_rel" ].
		
We created a subtype of `verb-lex` called `verb-comp-lex`. The
`CONT.HOOK.INDEX.SF` of the complement received the constraint
_prop-or-ques_.

		verb-comp-lex := verb-lex & basic-one-arg &
		  [ SYNSEM.LOCAL.CAT [ HEAD verb & [ MOD < > ],
		                       VAL.COMPS < #comp > ],
		    ARG-ST < #comp &
		             [ LOCAL [ CAT [ HEAD verb & [ FORM finite ],
		                           VAL [ SUBJ < >,
		                                 COMPS < > ],
					   MC + ], 
							   CONT.HOOK.INDEX.SF prop-or-ques ] ] > ].

The bifurcated the type into two subtypes `verb-decl-comp-lex` and
`verb-int-comp-lex` which had their `CONT.HOOK.INDEX.SF` constrained
to the appropriate values. Right now the verbs are using the less
specific `verb-comp-lex` type since our corpus resource went offline,
but we plan to remedy this in the future.

		verb-decl-comp-lex := verb-comp-lex &
		  [ ARG-ST < [ LOCAL.CONT.HOOK.INDEX.SF prop ] > ].

		verb-int-comp-lex := verb-comp-lex &
		  [ ARG-ST < [ LOCAL.CONT.HOOK.INDEX.SF ques ] > ].

We then took the verbs that we identified above and added additional
lexical entries which inherit from one of the three aforementioned
types that we defined like the following:

		ze_v1 := trans-verb-lex &
		  [ STEM < "ze" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_see_v_rel" ].

		ze_v2 := verb-comp-lex &
		  [ STEM < "ze" >,
		    SYNSEM.LKEYS.KEYREL.PRED "_see_v_rel" ].

We didn't complete this for all the verbs but we plan to fix that over
the weekend.

We used a `complementizer-lex-item` that the matrix generated for
us. We then made subtypes for interrogatives and declarative sentences
being embedded and constrained the `CONT.HOOK.INDEX.SF` on each of
them to `question` and `proposition` respectively.

		complementizer-lex-item := raise-sem-lex-item & basic-one-arg &
		  [ SYNSEM.LOCAL.CAT [ HEAD comp &
		                            [ MOD < > ],
		                       VAL [ SPR < >,
		                             SUBJ < >,
		                             COMPS < #comp > ] ],
		    ARG-ST < #comp &
		             [ LOCAL.CAT [ MC +,
		                           HEAD verb,
		                           VAL [ SUBJ < >,
		                                 COMPS < > ] ] ] > ].

		decl-comp-lex := complementizer-lex-item &
		  [ ARG-ST < [ LOCAL.CONT.HOOK.INDEX.SF prop ] > ].

		int-comp-lex := qpart-lex-item &
		  [ ARG-ST < [ LOCAL.CONT.HOOK.INDEX.SF ques ] > ].

We then used the indexed MRS view to check that our the indexes of our
predicates on the correct indices. Here are the relevant sentences
from our test suite:

		Source: a:233
		Vetted: t
		Judgment: g
		Phenomena: {embedded interrogative}
		Ikh veys nit vemen du host gezen.
		Ikh veys nit vemen du host ge-ze-n.
		1SG.NOM know;3G NEG WHO;DAT 2SG.NOM have;2SG PP-see-PP. 
		I know not whom you have seen.
		
		Source: a:232
		Vetted: t
		Judgment: g
		Phenomena: {embedded declarative}
		Ikh veys, az es moyshe kumt morgn.
		Ikh veys, az es moyshe kum-t morgn.
		1SG.NOM know;3G COMP 1SG.NOM Moyshe come-3G tomorrow.
		I know that it's Moyshe that's coming tomorrow.

# Further refinement of adverbs and adjectives

In Yiddish, uninflected adjective roots function as adverbs. When
modifying nouns, adjectives are mandatorily inflected for case, number
and gender. Rather than having both adjective root forms and identical
adverb forms in the lexicon (as our grammar previously did), we
decided to model this by having a single base adjective lexeme which
was used for entering adjectives lex entries into the lexicon. We did
this by creating a new lexical type, base-adjective-lex, which
contains all constraints that are common to both full adjectives and
adverbs:

base-adjective-lex := single-rel-lex-item & intersective-mod-lex & norm-ltop-lex-item &
 [ SYNSEM [ LOCAL.CAT.VAL [ SPR < >,
                            SUBJ < >,
                            COMPS < >,
                            SPEC < > ],
            LKEYS.KEYREL event-relation ],
   INFLECTED.ADJ-FLAG - ].

We then created two kinds of lexical rules which derive full
adjectives and the adverbs respectively from
base-adjective-lex. adverb-lex-lex rule is a non-inflecting that
lexical rule which derives adverbs, and adj-lex-rule-super is the
super type of all the inflecting lexical rules which derive the
specific adjective forms, applying the appropriate case, gender and
number.

adverb-lex-rule := const-lex-rule & ad-lex-rule-super & 
  [ SYNSEM.LOCAL.CAT.HEAD adv & [ MOD < [ LOCAL.CAT [ HEAD verb,
                                                      VAL.SUBJ < > ] ] > ] ].

adj-lex-rule-super := infl-lex-rule & ad-lex-rule-super &  
  [ SYNSEM.LOCAL.CAT [ HEAD adj & [ MOD < [ LOCAL.CAT [ HEAD noun,
                                                        VAL.SPR cons ] ] > ],
                       POSTHEAD - ] ].

Besides from constraining the head of the modified element
appropriately, the adverb-lex-rule constrains the VAL to have the SUBJ
list be empty, so that the adverb only attaches to the S node. This
had the effect of removing the spurious ambiguity found in the first
of the following examples (from 3 to 1 parses) and also of eliminating
the overgeneration for the next two examples.
 
Source: a:193
Vetted: t
Judgment: g
Phenomena: {adverb}
Dos kind zingt shayn.
Dos kind zing-t shayn.
DEF.ACC.N child sing-1SG nicely.
The child sings nicely.

Source: author based on a:223-224
Vetted: t
Judgment: u
Phenomena: {adverb, word order}
Der man shayn leyent dos bukh.
Der man shayn leyen-t dos bukh.
DEF.NOM.M man nicely read-1SG DEF.ACC.N book.
The man reads the book nicely.

# adverb in wrong location
Source: a:223-224
Vetted: t
Judgment: u
Phenomena: {adverb, word order}
Der man leyent shayn dos bukh.
Der man leyen-t shayn dos bukh.
DEF.NOM.M man read-1SG nicely DEF.ACC.N book.
The man reads the book nicely.

Below is an example of one of the lexical rules for inflecting
adjectives that inherits from adj-lex-rule-super:

sg-masc-nom-adj-lex-rule := adj-lex-rule-super & 
  [ SYNSEM.LOCAL.CAT.HEAD.MOD.FIRST.LOCAL [ CAT.HEAD.CASE nom,
                                            CONT.HOOK.INDEX.PNG [ GEND masc,
                                                                  NUM sg ] ] ].


# Yes-No Questions

At the beginning of this lab yes-no questions were not implemented,
since, as previously discussed, the Matrix did not give us a helpful
analysis of yes-no questions in Yiddish. Yes-no questions in Yiddish
are marked through VSO word order with the optional use of the marker
tsi preceding the clause. We initially tried to implement this by
creating new binary phrasal types to explicitly handle the top level
spans that correspond to VSO word order. The types we created were
q-head-comp-phrase and q-head-subj-phrase. The former handles verbs
which take at least one argument and the latter handles verbs which do
not take any arguments.  Both inherited from q-head-initial-head-nexus,
which was a new type we created, similar to head-initial-head-nexus,
except that it stipulates that the phrase must be a main clause.

q-head-initial-head-nexus := head-initial &
  [ SYNSEM.LOCAL.CAT.MC + ].

q-head-comp-phrase := basic-head-1st-comp-phrase & q-head-initial-head-nexus & interrogative-clause &
 [ HEAD-DTR head-subj-phrase ].

q-head-subj-phrase := basic-head-subj-phrase & q-head-initial-head-nexus & interrogative-clause & 
  [ HEAD-DTR.SYNSEM.LOCAL.CAT.VAL.COMPS < > ].

The problem with this approach is that it was permitting other
non-verb-initial sentences (such as OVS fronted declarative sentences
to parse as interrogatives). The problem with this approach was that
for q-head-subj-phrase we had to find a way to let it accept only
those phrases that were verb initial as the head. This includes
intransitive verbs which are then just simple verb lexemes, optionally
transitive verbs, which would have been head-opt-comp-phrases, and
also the OV part of OVS sentences, which combine with
comp-head-phrase.  This was leading down a horribly messy path, so we
instead backed out of this approach and tried the approach used by
Anstke in her Germanic CLIMB grammars.

This approach was to create a single unary-phrase rule which takes
phrases that would otherwise be accepted as spanning parses except
that their MC feature was na. This had the effect of accepting verb
initial phrases, as the MC feature is only set from na to + when it as
combined with something to the left of the verb -- but without this
initial element, the verb is then found at the left.

q-clause-rule := unary-phrase & interrogative-clause &
  [ SYNSEM.LOCAL.CAT [ MC +,
                       HEAD #head,
                       VAL #val ],
    ARGS < [ SYNSEM.LOCAL.CAT [ MC na,
                                HEAD #head & verb & [ FORM finite,
                                                      MOD < > ],
                                VAL #val & [ SUBJ < >,
                                             COMPS < >,
                                             SPR < >,
                                             SPEC < > ] ] ] > ].

In addition to inheriting from unary-phrase, q-clause-rule also
inherits from interrogative-clause which is found in the Matrix core
and provides the appropriate ques value for the SF feature. The rule
simply has the effect of requiring the single argument to be a phrase
which has the same constraints used to determine if a phrase can be a
root, with the exception of MC +, which it requires to be na, and then
this is then changed to + while everything else is passed up as is.

This rule was quite successful, with all interrogatives from the
testsuite that do not have the initial question marker tsi now
parsing. The following are two examples of such sentences:

Source: author
Vetted: f
Judgment: g
Phenomena: {word order, Matrix yes-no questions}
Zingt dos kind?
Zing-t dos kind?
sing-1SG DEF.NOM.N child?
The child sings?

Source: author
Vetted: f
Judgment: g
Phenomena: {word order, Matrix yes-no questions}
Leyent der man dos bukh?
Leyen-t der man dos bukh?
read-1SG DEF.NOM.M man DEF.ACC.N book?
The man reads the book?

One additional problem we found that was already present in our
grammar was that verb initial interrogative sentences that had an
adverb either preceding them or following them (as illustrated below)
were being parsed successfully as declarative sentences. These were only
discovered as problem cases as they are new additions to the testsuite.

Source: author
Vetted: f
Judgment: g
Phenomena: {word order, Matrix yes-no questions}
Gut zingt dos kind?
Gut zing-t dos kind?
well sing-1SG DEF.NOM.N child?
The child sings well?

Source: author
Vetted: f
Judgment: g
Phenomena: {word order, Matrix yes-no questions}
Zingt dos kind gut?
Zing-t dos kind gut?
sing-1SG DEF.NOM.N child well?
The child sings well?

This appeared to be because neither adj-head-int-phrase and
head-adj-int-phrase passed the value of MC from the head daughter up,
meaning that a phrase that was MC na, could be parses as a declarative
through the application of one of the adverb rules which left the MC
underspecified as luk. To fix this, we created two new phrasal rules
that inherited from head-adj-int-phrase and adj-head-int-phrase
respectively, and also passed up the MV value:

adj-head-int-phrase-x := adj-head-int-phrase &
  [ SYNSEM.LOCAL.CAT.MC #mc,
    HEAD-DTR.SYNSEM.LOCAL.CAT.MC #mc ].

head-adj-int-phrase-x := head-adj-int-phrase &
  [ SYNSEM.LOCAL.CAT.MC #mc,
    HEAD-DTR.SYNSEM.LOCAL.CAT.MC #mc ].
 
This was successful, however for both these sentneces there is still
one extra spurious parse, as the q-clause rule and relevant adj rule
can apply in either order. We have yet to correct this, but we propose
to correct this by enforcing that the unary-phrase apply as the top
most span -- since the adverb is part of the question -- possibly by
requiring the head daughter of the adj phrasal rules to not have SF be
ques.

## Yes-No Interrogative marker

As already mentioned, Yiddish allows for the use of the interrogative marker
tsi before verb initial inter


# Copula AP predicates

Predicative adjectives in Yiddish work very much similarly to English,
in that they are licensed through the use of a copula construction,
where the subject is connected to the adjectival predicate through the
present tense form of the auxiliary , which happens to even use the
same auxiliary verb -- zaynen (to be). A distinctive feature of
predicative adjectives is that base adjective remains uninflected.

Source: author
Vetted: f
Judgment: g
Phenomena: {cop}
Der man iz gut.
Der man iz gut.
DEF.SG.NOM.M man COP good.
The man is good.

To implement this we used the following lextype for copula verbs:

copula-verb-lex := verb-lex & trans-first-arg-raising-lex-item-2 &
  [ SYNSEM.LOCAL [ CAT.VAL [ SUBJ < #subj >,
                             COMPS < #comps > ],
                   CONT.HOOK.XARG #xarg ],
    ARG-ST < #subj &
             [ LOCAL [ CONT.HOOK.INDEX #xarg,
                       CAT [ VAL [ SPR < >,
                                   COMPS < > ],
                             HEAD noun ] ] ],
             #comps &
             [ LOCAL.CAT [ VAL [ COMPS < > ],
                           HEAD +jp ] ] > ].

Since zaynen is an irregular verb, we did not try to hook
copular-verb-lex up to the existing verbal morphology. Instead we have
created separate entries. Currently they are all copular-verb-lex
entries and hence appropriate person/number constraints are not placed
upon these forms -- so, like our existing auxiliaries, these currently
overgenerate. We plan to correct this next week in the Great Auxiliary
Deforestation.

One slight complication was that predicative adjectives are mot
inflected. Given our existing analysis, where full adjectives are only
derived through inflectional morphology, there was no way for the
copula to pick up uninflected predicative adjectives. We solved this
problem, by creating another const-lex-rule that inherits from
ad-lex-rule-super, deriving a specifically predicative adjective.  In
order to prevent this adjective from modifying nouns (which would be
ungrammatical due to not being inflected), we changed its definition
to have an empty MOD lost. Since this no longer made this lexeme a
modifier, we had to take intersective-mod-lex inheritance off from
base-adjective-lex and push it down into the adj-lex-rule-super and
adverb-lex-rule. The updated hierarchy for adjectives and adverbs is
as follows:

base-adjective-lex := single-rel-lex-item & norm-ltop-lex-item &
 [ SYNSEM [ LOCAL.CAT.VAL [ SPR < >,
                            SUBJ < >,
                            COMPS < >,
                            SPEC < > ],
            LKEYS.KEYREL event-relation ],
   INFLECTED.ADJ-FLAG - ].

ad-lex-rule-super := add-only-no-ccont-rule & 
  [ INFLECTED.ADJ-FLAG +, 
    DTR base-adjective-lex ].
    
adj-pred-rule := const-lex-rule & ad-lex-rule-super &  
  [ SYNSEM.LOCAL.CAT [ HEAD adj & [ MOD < > ] ] ].

adverb-lex-rule := const-lex-rule & ad-lex-rule-super & intersective-mod-lex &
  [ SYNSEM.LOCAL.CAT.HEAD adv & [ MOD < [ LOCAL.CAT [ HEAD verb,
                                                      VAL [ SUBJ < >,
                                                            COMPS < > ] ] ] > ] ].

adj-lex-rule-super := infl-lex-rule & ad-lex-rule-super &  intersective-mod-lex &
  [ SYNSEM.LOCAL.CAT [ HEAD adj & [ MOD < [ LOCAL.CAT [ HEAD noun,
                                                        VAL.SPR cons ] ] > ],
                       POSTHEAD - ] ].


A further complication that we have currently not handled, is that
there are some adjectives that can only be used as predicative
adjectives. We will have to work out how these are marked as such in
the lexicon. 

As for copula PPs, these function the same way as predicative adjectives:

Source: author
Vetted: f
Judgment: g
Phenomena: {cop}
Der man iz unter dos hoyz.
Der man iz unter dos hoyz.
DEF.SG.NOM.M man COP under DEF.SG.ACC.N house.
The man is under the house.

We implemented the preposition lex type as follows below, however it is most
definitely still in need of some constraining as we are currently
getting four parses for the above sentence.


preposition-lex := basic-int-mod-adposition-lex &
  [ SYNSEM.LOCAL.CAT [ HEAD adp & [ MOD < [ LOCAL.CAT [ HEAD noun,
                                                        VAL.SPR cons ] ] > ],
                       POSTHEAD - ] ].             


Unfortunately we ran out of time to complete copula NPs. These will be
attacked next week!

We were also somewhat lax with paying attention to the generation. While
we have tested some sentences that we have been adding as we progressed,
they have not been exhaustively tested and we have not yet added the 
needed entries to the trigger rules file. Again, this will be done for 
next week.

 
# Coverage numbers for Lab 6:

These are our coverage numbers:

		Total items:            215
		Positive items:         114
		Negative items:          74
		Parsing positive items:  63
		Parsing negative items:  10
		Coverage:                44.7%
		Overgeneration:          13.5%

We see some increased coverage with the addition of interrogative
sentence handling and non-verbal predicates . Comparing the lab6
baseline with the final results, we see that coverage has gone up from
41.1 to 44.7 analyses.

