Ling 567 Lab 7 Writeup
Nigel Kilmer and Woodley Packard

1.  Modals

Potentiality in Frisian is expressed by the auxiliary verb "ken".  The auxiliary appears in second position, undergoes obligatory inflection and takes a non-finite VP complement (similar to the tense marking auxiliaries):

Source: author
Vetted: f
Judgment: g
Phenomena:
ik ken glees iit
ik ken glees iit
I can glass eat
I can eat glass

Improper agreement:

Source: author
Vetted: f
Judgment: u
Phenomena:
hi ken glees iit
hi ken-st glees iit
he can-2SG glass eat
he can eat glass

Improper word order:

Source: author
Vetted: f
Judgment: u
Phenomena:
ik iit glees ken
ik iit glees ken
I eat glass ken
I can eat glass

To account for this phenomenon, we needed the "ken" lexeme and its type.  We treated "ken" as a contentful raising auxiliary verb:

; frisian.tdl
modal-auxiliary-aux-lex := subj-raise-aux & trans-first-arg-raising-lex-item-1 & norm-sem-lex-item &
  [ SYNSEM.LOCAL.CAT.VAL.COMPS.FIRST.LOCAL.CAT.HEAD.FORM modal-inf ].

; lexicon.tdl
ken := modal-auxiliary-aux-lex &
  [ STEM < "ken" >,
    SYNSEM.LKEYS.KEYREL.PRED "_can_v_rel" ].

The [ FORM modal-inf ] constraint prevents the main verb from being finite or a participle.  To make the modal infinitive available, we needed a new lexical rule:

; frisian.tdl
modal-infinitive-verbal-inflection-lex-rule := const-lex-rule & pernum-inflection-lex-rule-super &
  [ SYNSEM.LOCAL [ CAT.HEAD.FORM modal-inf ] ].

; lrules.tdl
modal-inf := modal-infinitive-verbal-inflection-lex-rule.

Our inventory of FORM values previously was simply {finite, nonfinite}.  In order to prevent our past participles from appearing as the complements of modal auxiliaries, we replaced 'nonfinite' with 'participle' and 'modal-inf' (our resource grammar frequently refers to two different types of infinitives: the modal infinitive and the complementational infinitive, the latter of which is used in a construct similar to "to eat" in English).

The analysis of the modal auxiliary "ken" seems to be functioning correctly, including generation.

2. Negation

Negation in Frisian is accomplished by the adverb "ek". 
For sentential negation, "ek" typically appears sentence-finally.
There are numerous examples of "ek" appearing in other positions, but its attachment and scope are not entirely clear in those examples. It seems that "ek" can be used to negate adverbial and adjectival constituents, but not noun phrases - there is a negative determiner glossed as the English determiner 'no' that is used instead.
The matrix negation library gave a functional implementation of full sentential negation for Frisian, which works correctly for "It doesn't hurt me.":

Source: author
Vetted: f
Judgment: g
Phenomena: neg
hat dēr mi siir ek
it give-3SG me hurt not
it doesn't hurt me

Note that the ditransitive verb `dö' "give" is part of the translation of "to hurt", not an auxiliary brought in to aid the syntax of negation.

3. One Corpus Sentence

The sentence from the test corpus that we have chosen to get parsing for this lab is:

Source: a:4
Vetted: s
Judgment: g
Phenomena: {wo, emb-d}
"Ik wel de Skepper spreek" swaaret hi
"Ik wel de Skepper spreek" swaar-t hi
"I will to creator SPREEK" swear-PST.PTCP he
he swore, "I will SPREEK to the creator"

We haven't been able to find a reliable translation of 'spreek'. It is probable that it is a cognate with English 'speak', but we already have the word 'snaki' which is translated as 'speak' in our lexicon, and dictionaries for related languages such as Dutch translate the word as 'weight', so it's unclear.

First, we had to add lexical entries for the preposition 'de', the common noun 'Skepper', the verb 'spreek', and the clausal-complement verb 'swaar'. There were no problems in adding these lexical items.

Our grammar source did not have any examples of embedded clauses without complementizers, except in the case of quotative verbs, so we previously only had an analysis of embedded clauses with complementizers. However, while it isn't attested by our source that complementizers are optional for verbs that take non-quotative clausal complements, it seems more likely that complementizers are always optional and that the difference between quotative and non-quotative clausal complement verbs is non-syntactic. To model the optionality of complementizers, the type emb-clause-verb-lex has been changed to allow either a verb or complementizer as the head of its complement:

emb-clause-verb-lex := main-verb-lex & clausal-second-arg-trans-lex-item &
  [ SYNSEM.LOCAL.CAT [ VAL.COMPS < #comp & [ LOCAL [ CAT.HEAD +vc,
			 	    		     CONT.HOOK.INDEX.SF prop-or-ques ] ] > ],
    ARG-ST < [ ], #comp > ].

After modifying this, the sentence would not parse due to the edge limit being reached during parsing. Increasing the maximum edge limit to 10000 allowed it to parse, with 8 trees. Four of these have clearly incorrect applications of the inv-lr rule and we are working on removing these.

The sentence also generates mostly correctly, but the auxiliary 'wel' is not included in any of the generated sentences.

4. Additional Grammar Fixes

We also fixed an issue with verbs that take clausal complements allowing their subject to have objective rather than subjective case. This was done by adding a constraint on main-verb-lex requiring that the subject have [ LOCAL.CAT.HEAD.CASE sub ].

5. Coverage

We added a few items to our test suite for lab 7.  All numbers are relative to this revised test suite.

Using our final grammar from lab 6, our coverage is 67 / 109 = 61.5% on positive items, and we overgenerate on 4 / 55 = 7.3% of our negative items.

After implementing everything necessary to parse the glass eating / hurting sentences and the corpus test sentence we chose, our coverage rose to 72 / 109 = 66.1% of positive items.  We also ruled out three ungrammatical items and significantly reduced ambiguity by disabling the narrow focus rules provided by the information structure library, since they seemed to be producing MRSes that were subsumed by the ordinary analyses.  This left us with overgeneration on 1 / 55 items = 1.8%.  This remaining ungrammatical item features verb-initial word order in an embedded clause, and we aren't yet sure what the best way to eliminate it is, given that word order in embedded clauses is otherwise free.
