12/11/10
Post-IAA notes on conventions:

1. Capitalization
   Preserve whatever capitalization and hyphenation is in examples.num.txt

2. Multi-words
   Verb-particles: verb+particle-pos
   All others:     token1-pos|token2-pos

3. Auxiliaries have|be|do|will|would|going+to
   Include all auxiliaries down to first non-aux token, but not more.
   If the non-aux token is a verb-particle, then include particle as usual

4. Measure/classifier Ns
   Include both tokens:  classifier-pos|noun-pos
   -- applies to partitives as well (some of, many of)

5. Dependency labels
   Nominal modification target of PPs and adjectives: use ARG1
   Verbal modification target of PPs: use MOD
   External argument of passive verb: use ARG

6. Clausal modifier attachment
   When not clear, include both main and embedded verbs: verb1-pos|verb2-pos

7. Complementizers
   Include both the complementizer token and its complement's head token

------------------------------------------------------------------------------
12/2/10

Format of gold standard file

i-id@phenomenon@target@triple-string@matched

Example:
1001000700030@barerel@1@in-14 ARG2 competition-9@

Use surface forms of tokens in triple-string (see `testitems.num.txt')

Triple ordering conventions:
predicate ARGn argument   dogs barked        barked-1 ARG1 dogs-0
predicate ARGn argument   dogs chase cats    chase-1 ARG2 cats-2
modified MOD modifier     We sang in Paris   sang-1 MOD in-2

Special cases:

For |have, be, do|, include both aux and main verb, using vertical bar as
 separator:                                 role-17 MOD has-19|embodied-20
For verb-particles and other multiwords, use plus sign as separator, with
 only first element's token position:       let+on-2 MOD in-4
For positive contracted auxiliaries, token begins with leading apostrophe:
                                            games-29 MOD 've-31|created-32
For negative contracted auxiliaries, include |n't| as part of token:
                                            man-4 MOD couldn't-6
For `classifier' nouns, include both classsifier and `content' noun tokens:
                                            to-20 ARG2 strips-13|metal-15
For coordination, list each conjunct as separate dependency (not conjunction)
 e.g. |the-0 murder-1 they-2 are-3 convicted-4 of-5 and-6 lynched-7 for-8|:
                                            of-5 ARG2 murder-1
                                            murder-1 MOD convicted-4
                                            for-8 ARG2 murder-1
                                            murder-1 MOD lynched-7
For gerunds, only record first complement of verb
                                            allocating-16 ARG2 domains-20
For passives, include both the verb and `by', and use role ARG
                                            taught-5|by-6 ARG officers-7
For prepositional passives, include verb, prep and by, and use role ARG:
                                           paid-6|for-4|by-8 ARG government-11

12/2/10

For CP complements, record both the C  and the head verb of the
embedded verbal projection (as well as have|be|do):
    said-2 ARG2 that-3|had-5|left-6
    tried-2 ARG2 to-3|swim-4

For arg/adj interleaving, there are cases where the relevant arg
dependency is between a semantically empty auxiliary (be, will, etc)
and its verbal complement.  I've been using ARG for this one and NOT
listing the auxiliary and complement on the same side of the
dependency.

For NP complements of be, I've been using ARG2

For appositives in arguments, use only the head of the first NP.

Apply the rule about including both have|be|do and the contentful
lexical verb, even with negated forms, such as haven't|aren't|don't.

Treat "number" as a `classifier noun'.

12/3/10

--- Looking at the first three examples for barerel, it seems that this rule:

For |have, be, do|, include both aux and main verb, using vertical bar
as separator:
  role-17 MOD has-19|embodied-20

applies only when the dependency involves the auxiliary directly
(e.g., as a MOD, as opposed to dependencies between the verb and its
complements).  I'm going to update the annotations for argadj to
reflect this understanding.

--- I have been treating will|would|going to like have|be|do, but not
can, could, might

--- Using just ARG for surface subjects of passives as well.  (I've
noted a couple of reduced subject relatives, but only as the second
instance in an example... so now I'm guessing that these don't really
count, but I am annotating them anyway.)

--- In constructions like "all the people", putting both all and the
head noun in the dependency.

saw-3 ARG2 all-7|people-9

--- I'm using ARG2 for what I think are argument marking prepositions,
aside from passive by.

--- I'm not sure I've been using the prepositional passive rule
entirely correctly, I think because my intuition is that the
preposition in these cases isn't relevant to the dependency involving
the by phrase.  There aren't too many of these, however, and I'm
confident we can fix them in adjudication.

--- I'm treating two-part proper names as MWEs.
--- Following this example from targets.txt:

1003061600060@absol@1@taking-10 ARG1 Garth-9|Tander-9@

now treating proper names not at MWEs but as disjunctions in the 
dependency.


--- I put both "obligated" and "to" on the left hand side of the
dependency triple in this one, but I wasn't happy about it (since "to"
really belongs with "carry").

1008045600050 much-0 like-1 terrestrial-2 radio-3 stations-4 that-5
carry-6 air-7 america-8 programming-9 XM-10 167-11 is-12 not-13
obligated-14 to-15 and-16 does-17 not-18 carry-19 the-20 network-21
24/7-22

Including be|have|do|will|going to but not other raising verbs together 
with contentful predicate on left-hand side of ARG1 dependencies.

Some of the ned examples are open adjuncts.  Rather than try to fill
in the ARG1, I've been giving a MOD dependency with the main verb
as the dependent. (e.g., 1001120201360)

Ann#1 has this in targets.txt:
1000065200110@absol@1@being-18 ARG1 use-17@

But that being is not the identity be, so I did this:
1000065200110@absol@1@being-18|in-19 ARG1 use-17@

--- For "with" absolutives, I'm putting both the with and the
head of the absolutive predicate after the MOD in the MOD
dependency

--- For headless NPs, using adjective as head: "The most recent" -> "recent"

--- I am assuming that absolutives only modify verbal projections.

--- Because of this example, I am including modifier uses
of -ing forms as vger examples:

1000102100350@unknown@formal@none@1@S@Others advocated the opposite convention, including IBM for writing [[Algol]] programs.@@@@1@11@nil@oe@24-11-2010

... but I suspect that this example is spurious and the modifier
uses are not meant to be included.  Not treating -ing forms in
absolutive predicates as gerunds, though.  Example of that
is starting-14 here:

1000102200040 it-0 has-1 been-2 in-3 development-4 since-5 2000-6 with-7 actual-8 work-9 on-10 building-11 the-12 game-13 starting-14 in-15 2003-16

--- Where no complement is available for a gerund, recording
a modifier of it instead.  If neither complement nor modifier
is available, then only one depedency is recorded (the role
the gerund is playing in the larger sentence).


