Page 1 of 3

Discourse Features and Treebanks

Posted: April 18th, 2017, 2:04 pm
by Jonathan Robie
I've started thinking about the best ways to reflect discourse features in a treebank. I have months of work to do on other things first, but I'd like to start thinking about this now.

So far, I think the following features from Levinsohn's data should probably be added to our Lowfat treebank in some way:
  • P1 / P2 membership for words and word groups, including types like "topic", "situation", etc.
  • Split focus / split constituents
Are there other features that I should add to the Lowfat treebanks?

I should mention that the hierarchy in Levinsohn's discourse feature data, depending on points of departure, is sometimes a different hierarchy than a purely sentence-based hierarchy. A discourse analysis hierarchy would be yet another hierarchy. So I suspect there will never be just one useful hierarchy for analyzing a set of sentences ...

Is this more or less correct?

Re: Discourse Features and Treebanks

Posted: April 19th, 2017, 1:21 pm
by MAubrey
Jonathan Robie wrote:
April 18th, 2017, 2:04 pm
Are there other features that I should add to the Lowfat treebanks?
That depends on whether you want them to stay 'low fat'!
Jonathan Robie wrote:
April 18th, 2017, 2:04 pm
I should mention that the hierarchy in Levinsohn's discourse feature data, depending on points of departure, is sometimes a different hierarchy than a purely sentence-based hierarchy. A discourse analysis hierarchy would be yet another hierarchy. So I suspect there will never be just one useful hierarchy for analyzing a set of sentences ...

Is this more or less correct?
There are approaches to information structure that are designed around syntax models that are more or less comparable to your trees. It's possible that you could do some sort of mapping to get from one to the other.

There's a draft version of Sag Wasaw & Bender's book Syntactic theory (here) that you might find useful (edit: this draft only has chapters 1-11 and doesn't get to information structure. My mistake.). Their framework is aimed at computational work. I know that Andi Wu had (some) reliance on their work at least in part.

Re: Discourse Features and Treebanks

Posted: April 20th, 2017, 12:08 am
by Stephen Carlson
What happens when they form different hierarchies? It's not clear to me that there is much to be gained by reconciling two different trees.

Re: Discourse Features and Treebanks

Posted: April 20th, 2017, 6:05 am
by Jonathan Robie
Stephen Carlson wrote:
April 20th, 2017, 12:08 am
What happens when they form different hierarchies? It's not clear to me that there is much to be gained by reconciling two different trees.
I would have to play with it to see if it makes sense. A common reference system and having both for the same text might be a better approach.

Re: Discourse Features and Treebanks

Posted: April 21st, 2017, 8:21 am
by Jonathan Robie
One more question on this topic.

Some people have suggested that syntax treebanks are a really good starting point for discourse analysis, which should be built as a layer on top of the syntax trees. Levinsohn's data looks like a different hierarchy, where points of departure often do not align with sentence boundaries, for instance, though both hierarchies use essentially the same clauses. A syntax treebank would be helpful for identifying the underlying discourse features, but he uses them to build a different hierarchy.

Do Runge's frames align with sentence boundaries and clause boundaries, or are they also a distinct hierarchy?

What about other approaches to discourse analysis?

Re: Discourse Features and Treebanks

Posted: April 21st, 2017, 11:47 am
by MAubrey
The trouble with building discourse features as a layer on top of syntax tree is that they're never going to map accurately unless you had already started with the discourse structure in mind at the beginning. Discourse features affect syntactic constituency in ways that aren't foreseeable if you're not already thinking about discourse when you build the trees. There are going to be mismatches.

Still, these are good questions. It's just a matter of how much revision you want to embark on.

Re: Discourse Features and Treebanks

Posted: April 21st, 2017, 2:50 pm
by Stirling Bartholomew
MAubrey wrote:
April 21st, 2017, 11:47 am
The trouble with building discourse features as a layer on top of syntax tree is that they're never going to map accurately unless you had already started with the discourse structure in mind at the beginning. Discourse features affect syntactic constituency in ways that aren't foreseeable if you're not already thinking about discourse when you build the trees. There are going to be mismatches.
Stephen Carlson wrote:
April 20th, 2017, 12:08 am
What happens when they form different hierarchies? It's not clear to me that there is much to be gained by reconciling two different trees.
I have reservations about what look like great grand children of generative syntax trees being wed to graphical reppresentations of information structure. I am not saying that someone hasn't tried to do this already. I don't keep up on all the latter day spinoffs from generative grammer. Jumping between dissimilar frameworks is possible if you understand both frameworks. But wedding the frameworks is another issue.

Re: Discourse Features and Treebanks

Posted: April 21st, 2017, 4:09 pm
by Jonathan Robie
I don't think Levinsohn's points of departure closely mirror anything I would call a sentence in any system I know. Here's an example where they clearly do not - consider the placement of the hierarchy starting with Σὺ τίς εἶ in relation to the sentence hierarchy.
Screen Shot 2017-04-21 at 3.51.14 PM.png
Screen Shot 2017-04-21 at 3.51.14 PM.png (158.2 KiB) Viewed 1822 times


That hierarchy is fundamentally different from the hierarchy we use to represent the same text in Lowfat. (FWIW, the gridlines here can be turned on or off, I have them on now ...)

That's what prompted this question in the first place.

Screen Shot 2017-04-21 at 4.02.04 PM.png
Screen Shot 2017-04-21 at 4.02.04 PM.png (71.83 KiB) Viewed 1822 times
Screen Shot 2017-04-21 at 4.02.27 PM.png
Screen Shot 2017-04-21 at 4.02.27 PM.png (61.05 KiB) Viewed 1822 times

Re: Discourse Features and Treebanks

Posted: April 22nd, 2017, 10:29 pm
by Stephen Carlson
Jonathan Robie wrote:
April 21st, 2017, 4:09 pm
I don't think Levinsohn's points of departure closely mirror anything I would call a sentence in any system I know.
No, almost never a sentence. Points of departure usually correspond to a phrase (e.g. a noun phrase for a referential point of departure, prepositional phrase for settings, conditional clauses, etc.).

The term "point of departure" is usually associated with the Prague School and goes back to Henri Weil's 1844 study of Greek word order (which has been translated into English and available on line). The terminology is in his 2000 coursebook, but I'm not sure if he's kept it.

Re: Discourse Features and Treebanks

Posted: April 23rd, 2017, 8:03 am
by Jonathan Robie
Yes, he still uses that term. My point is that his hierarchies do not correspond to any conventional notion of a sentence.
The first level of indentation reflects the presence in a proposition of any pre-nuclear or pre-verbal
constituents, including vocatives, pre-verbal subjects (1:6a) and other points of departure (also called
‘topicalised constituents’―3:6a), preposed focal constituents (1:8c) and pre-nuclear participial
clauses (DFNTG §11.1, as in 2:2a-b).
The second level of indentation is for nuclear clauses that begin with a verb (e.g. 1:2a).

Third and subsequent levels of indentation are for post-nuclear clauses (1:2b), as well as the second
part of long propositions that have had to be divided.
Which means it makes more sense to use a common reference system, but not try to combine it with a conventional syntax tree. I think that's equally true of syntax trees based on constituency vs. dependency.