Word Order: Y2 Hyperbaton in Matthew 1:20

[Jonathan Robie]

You need a representation that can handle two things: dependencies (which element depends on which) and word order (which element precedes which). Constituency trees (which is the base representation of your syntax trees, IIRC) basically assume that dependency and linear order go together, which works well for English but not so much for Greek. I suppose you can hack the representation to account for hyperbaton, but my preference would be to use a representation originally designed for Greek instead of one for English.

As you know, I disagree, but let's not have a bike shed argument about dependency trees versus constituency trees.

Dag Haug has written well on the virtues of each for representing Greek. Both his trees and ours include some aspects of each. I'll encourage you to take a look at what we've done once we've done it. I'm quite convinced that neither pure constituency nor pure dependency is adequate here. And I think that's based on significant experience at this point.

[Stephen Carlson]

Actually, my first sentence is clear that mere dependencies are not enough.

[Jonathan Robie]

Actually, my first sentence is clear that mere dependencies are not enough.

Sorry, I must have misunderstood.

Our current trees represent both word order and constituency. Our new trees will do the same. We are changing the way in which they do that. In the current format, word order is represented using sequence tokens and constituents are held intact. In the new format, document order strictly follows word order, and constituents can be interrupted. I'm fine tuning the ways in which a constituent can be interrupted.

We use traditional concepts like phrase and clause, and within a clause we identify the roles of constituents by their relationship to the verb. I would argue that constituents are a third thing that is good to represent.

[Stephen Carlson]

Our current trees represent both word order and constituency. Our new trees will do the same. We are changing the way in which they do that. In the current format, word order is represented using sequence tokens and constituents are held intact. In the new format, document order strictly follows word order, and constituents can be interrupted. I'm fine tuning the ways in which a constituent can be interrupted.

From Dag's paper that you linked to: "Intuitively, we can look at constituency as combining information about dependencies and word order, i.e. a constituent is a continuous domain of words related by dependency relations." The problem to me seems that the way constituency combines information about dependences and word order isn't really working for you, so you're augmenting your trees with word order. Fine, but this is suggestive of a poor fit between constituency and non-configurational languages like Greek, a poor fit that's been known for a while. I suppose one can jury-rig a constituency-based system with bailing wire and duct tape to get it to work for Greek, but you'll be spending your whole time fighting against the formalism. Here, you're separating out the word order information, but are still using a word-order based formalism (constituency) to handle the remaining dependence, despite their discontinuous nature.

We use traditional concepts like phrase and clause, and within a clause we identify the roles of constituents by their relationship to the verb. I would argue that constituents are a third thing that is good to represent.

In my opinion, once you're talking about discontinuous constituents (which is necessary for Greek), you've abandoned the benefits that constituency is supposed to deliver.

[Jonathan Robie]

We use traditional concepts like phrase and clause, and within a clause we identify the roles of constituents by their relationship to the verb. I would argue that constituents are a third thing that is good to represent.

In my opinion, once you're talking about discontinuous constituents (which is necessary for Greek), you've abandoned the benefits that constituency is supposed to deliver.

We'll find out. We can always convert to dependency from constituency, going the other direction is much more difficult. We're trying to optimize for (1) easily explaining a difficult sentence, and (2) queries on a wide variety of questions. We may well try more than one approach before we're done, or even support more than one model.

[Stephen Carlson]

Good luck. Sometimes the best way to learn is to do.

[Jonathan Robie]

Good luck. Sometimes the best way to learn is to do.

Exactly. I have to try it this way first.

[jtauber]

In my opinion, once you're talking about discontinuous constituents (which is necessary for Greek), you've abandoned the benefits that constituency is supposed to deliver.

When dependency is applied to languages like English, there are often extra constraint mechanisms put in place to try to prevent discontinuous structures where not allowed. Similarly when constituency approaches are applied to languages like Greek (or Dyirbal!) extra mechanisms need to be invented to handle the discontinuities. In fact, even with English, extra mechanisms are often invented even with constituency approaches (movement, GPSG SLASH). One of the innovations of GPSG was separating dominance relations from linear ordering relations (although that was more motivated by perspicuity than handing discontinuities I think).

BTW, having come off a semester of Pindar, I laugh at the hyperbaton in the NT

[Stephen Carlson]

separating dominance relations from linear ordering relations

If I understand correctly what's happening in the Minimalist Program, this too is now being done among generative linguists.

[jtauber]

separating dominance relations from linear ordering relations

If I understand correctly what's happening in the Minimalist Program, this too is now being done among generative linguists.

GPSG doesn't* do it at the actual representation level, though, only the rules, whereas transformational grammar to a certain extent has always done it through movement.

The innovation in GPSG was handling certain things with just single-level representation and phrase structure rules that transformational grammar needed movement or multi-levels of representation for.

*(or should I say didn't as no one does GPSG any more to my knowledge, it being mostly taken over by HPSG)