Back in November, I discussed a paper by Florian Wolf and Ted Gibson that criticized the RST ("Rhetorical Structure Theory") Discourse Treebank and the ideas about how to represent discourse coherence that it embodies. Wolf and Gibson documented an alternative approach to annotating discourse structure and announced an alternative annotated corpus. In December, I discussed a response by Daniel Marcu. Florian Wolf has just emailed to let me know that he and Gibson have written and posted a response to Marcu's response.
I've learned a lot by reading this back-and-forth. As I wrote in connection with the first Wolf and Gibson paper,
I think that these things -- both the RST Treebank and the Wolf/Gibson corpus -- are wonderful steps forward. Two alternative approaches to the same (hard) problem offer not just examples and arguments, but also alternative corpora (of overlapping material!), annotation manuals, annotation tools and so on.
The core question -- whether trees or more general graph structures are more appropriate for representing relations among units in discourse -- is especially difficult to resolve, in my opinion, because it's not easy to agree, in particular cases, about exactly what the units and relations should be. I share with just about everyone else the impression that such units and relations exist, and that it's possible to say things about them that are true (e.g. that segment 2 is an attribution for segments 0 and 1 in the passage below) or false (e.g. that segment 3 is an attribution for segments 1 and 2).
0. Farm prices in October edged up 0.7% from September
1. as raw milk prices continued their rise,
2. the Agriculture Department said.
3. Milk sold to the nation's dairy plants and dealers averaged $14.50 for each hundred pounds,
4. up 50 cents from September and up $1.50 from October 1988,
5. the department said.
That much is something that all parties in this debate agree about.
However, I find it harder to decide about some of the other specific issues that are in play here. For example, Marcu disagrees with W&G about whether there is a "similarity" relation between segments 2 and 5. I'm not sure what I think -- certainly the segments are similar, but I'm not clear whether that "similarity" is part of the cognitive structure of the discourse in the same sense that the "attribution" relation is. But this may be only because my first analytic instinct is to think in terms of some sort of generative model, for which tree structures a natural choice. Such an approach may lack traction in the case of discourse, where one might imagine that the underlying process is often not tightly constrained by the (perhaps partial and emergent) structures that are undeniably present in the result.
Another case where I lack clear intuitions is the question of whether segments 3-5 are an "elaboration" just of segment 1, as W&G prefer, or perhaps of segments 0-2, as Marcu suggests.
For that matter, we might ask about other possible sub-units in this little passage -- after all, there are three chunks describable as "up Q from DATE" -- perhaps these should be identified and linked by a sort of similarity relation?
This is not an attempt at a reductio argument, but rather a serious observation that a stretch of text, like a picture or a piece of music, generally embodies many simultaneous, often-overlapping and variably-salient relationships. In some cases it makes sense to strip out a particular layer of description and observe its adherence to some quite regular (often hierarchical) structure. This works for many artistic forms (sonnets, blues, movie plots) and some aspects of linguistic structure. In other cases, the process doesn't seem to converge, but instead leads us into an apparently bottomless process of interpretation and commentary. The case of "rhetorical structure" under its various names is one that has been stuck in the middle, so to speak, for millennia. The way forward is through exactly the kind of careful analysis and discussion that Marcu, Wolf, & Gibson and others are carrying on. So I agree with both of them.
Posted by Mark Liberman at March 4, 2004 12:53 PM