Docs: discourse vs. advmod

[nschneid]

A couple of discrepancies:

• actually: The discourse doc page lists actually as an example, but in the English treebank, all actually tokens are (actually) attached as advmod. Should the distinction be clarified in the guidelines?

• of course: In the English treebank, there are 4 tokens where of course is attached as discourse. I'm guessing at least 2 of them should be advmod, which is how most of course tokens are handled, as well as all tokens of however. For the two sentence fragments where there is no verb, I think discourse is reasonable.

[nschneid]

cf. #306

[dan-zeman]

In UD 2.4 English EWT, _actually_ is still treated as advmod, except for one parataxis; see this query: http://hdl.handle.net/11346/PMLTQ-WZIF. @manning @sebschu : Should we modify the examples in the documentation or the data?

[dan-zeman]

UD 2.6 English EWT, still no change: http://hdl.handle.net/11346/PMLTQ-0YVL.

[amir-zeldes]

In GUM 'actually' is also advmod. Are we sure we want this to be 'discourse'? There is a variety of sentence 'stance' adverbials which are a bit different from classic VP adverbs, but there are also different uses of actually (e.g. adnominal and often stressed "actually a rock"), and I'm not sure we'd get high agreement on this. I would be pretty OK with removing "actually" from the examples in the guidelines and keeping it as advmod, but I'm convinceable. FWIW, here's a frequency list of words labeled discourse in EWT vs. GUM with f>1 (some may be errors; I'd say other words similar to 'actually', like 'indeed' are not represented):

===
well    25
no  15
oh  15
yes 9
Yeah    8
mean    6
yeah    6
right   6
say 5
like    5
fucking 4
Hey 4
hey 4
man 4
ha  3
ho  3
Yep 3
please  2
=p  2
OK  2
No  2
ok  2
lol 2
you 2
Well    2
look    2
okay    2
sorry   2

===
please  330
yes 67
:)  59
well    58
no  41
like    26
lol 21
ok  16
hi  15
oh  15
hey 13
thanks  12
okay    10
yeah    10
anyway  9
:(  9
wow 9
sorry   9
:-) 8
:d  7
mean    7
right   7
(:  5
of  5
;)  4
^_^ 4
say 4
yep 4
haha    4
p.s.    4
sure    4
though  3
anyways 3
=)  3
uh  3
huh 3
pls 3
plz 3
yea 3
dhin    3
hmmm    3
o   2
:p  2
^^  2
eh  2
er  2
lo  2
ps  2
so  2
ya  2
aye 2
btw 2
god 2
idk 2
thx 2
fine    2
hell    2
nope    2
hello   2
ummmm   2

[nschneid]

From a PDTB perspective, there are many discourse adverbials, which are a subtype of discourse connectives: therefore, actually, however, instead, indeed, in other words, for example, on the other hand, ...

But it looks like most of the ones listed above are greetings, please/thanks, yes/no, filled pauses, emoticons/laughter/swearing, etc. These are not really semantic connectives. So I'd be happy to use advmod for the adverbial connectives like "actually", "though", "anyway(s)".

Not sure about "(I) mean", "so", "well", "look", "say": These feel to me like they mainly have a floor-taking function (we could add "listen" and "now" too); the speaker is leaning in to the fact that they are speaking now, or emphasizing a point, and not necessarily choosing these words based on the content of what they are trying to communicate. "(I) mean" and "well" do feel like they have, or could have, a contrast-like quality though, where the speaker is objecting to or hedging what the other person has said. "I mean" also appears to have a use similar to "that is". (For "so" I am talking about the let-me-start-talking sense, not the causal sense.)

[amir-zeldes]

For multiword expressions with internals syntax, like "I mean", we've been using the parataxis guideline, which allows you to keep the clause-like analysis internally:

parataxis(mainverb,mean)
nsubj(mean,I)

I think this is better than using discourse on 'mean', which seems like more of a pragmatic than a syntactic statement.

[dan-zeman]

I think it is better to attach _actually_ (and other sentence "stance" adverbs) as advmod. In fact I thought that this is what UD does and didn't know that _actually_ is cited under discourse until I saw this issue.

Perhaps we should simply remove the example from the documentation of discourse. BTW the formulation with this example seems to date back at least to Depling 2013, two years before first release of UD :-) But the UD English data, since UD 1.0 in January 2015, treated _actually_ as advmod (with one exception, which is parataxis).

[amir-zeldes]

OK, given the apparent consensus on this, I went ahead and updated the documentation.