PRANAV ANAND CO-PI ON LARGE NSF-FUNDED RESEARCH PROJECT
It was recently announced that an interdisciplinary research team which includes Pranav Anand as Co-PI was to be awarded a large multi-year research award from the National Science Foundation. WHASC spoke to Pranav about the project, what it would involve, and where it fits in the larger scheme of things.
WHASC: First off: Many congratulations to you and your colleagues, Pranav, on getting funding for this project. It seems like a very large one, involving quite a few people and quite a few different groups. Could you give us a sense of who is involved and what the principal goals are?
PA: The goal for the project, in brief, is to try to understand how argumentative contexts develop. Much prior research on dialogue has emphasized collaborative environments: product meetings, helpdesk-style sites, task-oriented interaction. We are now at a watershed moment in the language sciences because we are simply awash in data from all corners. A very large portion of that interaction is decidedly non-collaborative, and the central aims of this project are to discover what aspects of what we learned in collaborative interactions transfer. The team is interdisciplinary, involving two psychologists who specialize on language (Jean E. Fox Tree and Steve Whittaker), two computer scientists who work on NLP (Craig Martell, Lyn Walker), and a linguist who looks at higher-level pragmatics (me). How we are adjudicating the work is a somewhat complicated dance, but suffice to say that we all have hands in each of the pots.
WHASC: What are your own hopes for the project? How does it interact with your own research program?
PA: I have been studying higher-level pragmatic moves for the past few years, and one central outcome of this investigation has been the relative narrowness of the language we look at in formal work on pragmatic interactions. It’s hard to know how fetishizing we are, since we don’t really have a good catalog of the range of uses across genres and contexts, but, as I said above, we know that things are much broader than we lay them out in our simple theorizing. My own goal is relatively modest: if we take this genre on its own terms and study its properties, and then try to reflect them back on what we assume from the more rational, information-sharing contexts we tend to assume, where are the points of friction? There is an assumption in much pragmatic work that the collaborative interaction is somehow primary, and that all other uses of, say, questions or imperatives, arise from some complicated pragmatic negotiation. But I wonder if our own approaches might blind us to a more basic plasticity.
WHASC: What will be the main activities?
PA: We will be doing a lot of summarizing and annotating of some online argumentative debates we have gathered over the past few years, and then building models to try to match those annotations. There is lots of interesting semantics and pragmatics all over the place — these conversations are about lots of issues floating in and out in complicated anaphoric ways, about varieties of disagreement, about the way conversations can be structured. Under the hood, we will be trying to piece together some of that, but I suspect a lot of the interesting semantic issues will be delayed until late in the process; they might simply be too hard at present.
WHASC:How long will the project last?
PA: It’s a 3 year grant.
WHASC: What do you see the most important outcomes being?
PA:I think whatever comes of the annotation will be valuable. We still don’t know how to describe this kind of data adequately. One surprising embarrassment of the field is that ‘discourse relations’ is a term that really only describes monologic conversational organization, while ‘dialog acts’, which talks about two or more parties, is almost always at the level of conversational turn pair. It’s surprising that we don’t have a theory of relations across dialogic turns. We will need to construct one.
WHASC: This is a project at the interface between linguistics and computer science. How do you see such interdisciplinary work developing? What are the opportunities? What are the pitfalls?
WHASC: Gosh, that’s hard to answer. I think one central issue is what these two communities consider interesting and publishable. For linguists, it is the surprising corner cases that merit real attention. Our modern abandonment of large scale grammars or fragments rests, I think, on the fact that so much of those projects turns out to be drudgery, from a linguistic point of view. But in NLP what is interesting is what practical results can be achieved with the least effort. And this is because the goal here is always to build an actual functioning product, and for that one needs to have some kind of answer to every niggling question. If that can be handled with less upfront specification, that is very interesting, but the corner cases, which amount to 1% or less, are so rare to not warrant any attention until we are at something resembling ceiling for the more prosaic material.
What this means in terms of cross-fertilization is unclear. It has been remarked by many that the knowledge transfer has tended to go from formal linguistics to NLP: phones, trees, morphemes, etc. But those tools arose from foundational issues for the field. I’m not sure what modern methods have passed over. In the other direction, I see one clear benefit being the range and diversity of annotated material that is now available to the practicing linguist. It still is true that little of that is workaday material for linguists, but I think we’ll see that change relatively soon. But really deep, lasting changes have been rare. For example, Massimo Poesio‘s work on definite noun phrase realization is a careful, exhaustive, theoretically-informed examination on the uses of definite descriptions. He (and subsequent work) found that well over half of definites are discourse-novel. But I doubt this is widely known in formal circles, despite its clear relevance.
WHASC: Thanks for your time, Pranav, and best of luck with the project. It sounds really exciting.