David Reitter
The Penn State University
Thursday, June 2nd at 2:30 PM in AP&M 4301.
Computational approaches to modeling dialogue
Abstract: In this talk, I will discuss how computational linguistics and "big data" methods can help us model language processing. I will describe empirical results and cognitive models of language production in dialogue. A focus of this work is on adaption effects, which cause a speaker to align with the interlocutor's language.
"Syntactic priming" is a well-studied adaptation effect that causes speakers to make syntactic choices according to previously used constructions (or rules, or lexical categories); it can provide a window into psycholinguistic representations. I present a cognitive computational model of syntactic priming, which situates language production with a framework of general cognition.
Syntactic priming is also the basis of a recent, influential theory of coordination in dialogue. I have used corpus data to validate the syntax-task success link, a key prediction of the Interactive Alignment Model (Pickering&Garrod, 2004). Evidence now also comes from newer, much larger and much less curated sources. These allow us to look at key questions surrounding interactive alignment as a functional model of language use. Working with web-based online forums such as Reddit, we found that alignment is relatively automatic, rather than being a social signal. The extent of alignment is also greater than previously thought. As I will show, dialogue partners also converge in syntactic complexity and high-level intent.
A new information-theoretic analysis contributes to literature surrounding information density; we show that interlocutors take on different roles and systematically converge in their information density. This paves the way to integrating language-as-action theories of dialogue with mechanistic interactive alignment.