MAubrey wrote: ↑
April 25th, 2017, 12:13 am
The problem is that the best (in my opinion) syntactic frameworks for modeling information structure are so very much non-generative that they would cohere at all with what Andi Wu & Randal Tan have done or what Jonathan & Micheal have derived from it--hence my talk of mismatches above.
I am very much out of date - in the 1980s, I was a PhD student in cognitive psychology, and took a little linguistics, but I have not kept up, and am trying to learn some of these things now. Since then, I have been fairly deeply involved in information modeling as a computer scientist.
Let me try a few guesses at what you are saying, and see how close I can get. Our treebanks model individual sentences. They do not attempt to model information flow. Discourse features also seem to occur at roughly the same level, hence constructs like P1 P2 V X, where the position of constituents with respect to the verb is meaningful.
One of the grand themes of cognitive psychology is that the world is full of information, we could never function if we tried to take it all into account, so we get very good at paying attention to the right information. You could never walk down the street if you had to first process every bit of information available to your senses as you go. You only become fluent at reading when you no longer need to pay attention to each individual letter to read a word. And a great deal of meaning is never encoded in sentence signals in the first place but is based on common frameworks of understanding. Once upon a time, Whitehead and Russell came up with the bright idea of encoding the great works of philosophy mathematically so we could compute new knowledge from them. They hit a bunch of problems, Gödel's Theorem and Turning's Halting Problem among them, but there were other problems that were much more mundane. Two people encoding the information understood the sentences differently, even in their native language, and wrote down different sets of propositions. More importantly, the information needed to reason from a sentence was often not found in the text at all, it was found in the set of understandings and assumptions people have about the world.
In Studies in the Way of Words, Grice gives the example of a man who comes to your door and knocks, carrying a gas can. You assume that the gas can is empty and he came to your door because he ran out of gas. You tell him there is a gas station around the corner. He assumes you are not willing to give him a ride, and that this is the closest gas station (there isn't a closer one in the other direction). He also assumes you believe the gas station is probably open.
Our treebanks are at a fairly low level in all of this. They represent the information found in individual sentences. A human being who is reading these sentences may not pay attention to all of this information, but they do represent the structure of a sentence in a meaningful way. As humans read, they select the information they care about and relate it to what they already know, discarding what they do not care about or do not understand. A woman living in Galilee might read Mark 16 very differently than a man, for instance, relating to the experience of being a woman compelled to show love in a situation that is potentially dangerous, when she is not in control of most of what is going on. We're not that ambitious. We simply try to reflect the structure of individual sentences. Discourse grammar might indicate a few more things that we could add to our representation of a sentence, but it is probably better to treat Levinsohn's data as a separate analysis of the same text.
I am not up to date on linguistic theories, and I really don't understand what level each of them functions on - what part of the problem they solve for understanding the way people extract meaning from text. I'm very much a data modeler, I like theories embodied in a set of regular rules applied to a corpus in a way that is verifiable. That makes Levinsohn and Runge attractive to me, they have each done this work. I would generally expect each analysis to answer a different set of questions, and having multiple analyses is a good thing.
So ... suppose I have treebanks and I have Levinsohn's data. Tell me about information flow, what part of this problem those models solve and how they solve it. Give me a concrete example or two of how it works and how it relates to Role and Reference Grammar. Links to blog posts or articles or whatever would be fine.
I have Dik's book and will read it along with Levinsohn. This year, much of my work will focus on improving the treebanks.
ἐξίσταντο δὲ πάντες καὶ διηποροῦντο, ἄλλος πρὸς ἄλλον λέγοντες, τί θέλει τοῦτο εἶναι;