Read on Twitter
Now available! Our #conll2020 paper on how a simple form of repair can ease the computational burden of pragmatic reasoning, essentially distributing cognition in interaction (preprint: https://psyarxiv.com/u2y4a/ ) 1/10 https://twitter.com/CCS_donders/status/1315716448371847168
First author Jacqueline van Arkel is a brilliant MSc student who carried out this work as an internship in our project w/ @mariekewoe & @MarkBlokpoel. It's not often that an internship report becomes a peer-reviewed #CoNLL conference paper! https://markdingemanse.net/elpaco/people/#students
The paper addresses what is perhaps the achilles heel of popular models of rational speech act theory: computational complexity. Often implemented with elegant toy lexicons, we wondered how well it would scale to larger ones, and whether interaction (esp repair) can help
We test a hypothesis inspired by empirical work on social interaction: that repair may help communicators to reduce their individual computational burden, in effect distributing the process of reaching mutual understanding over multiple turns (and multiple minds)
Using complexity analysis we compare two ways to deal with uncertainty: interactional, where agents are literal communicators but can use repair; & pragmatic, where agents use pragmatic reasoning (in 2 variants: frugal (levelling up if needed) and fancy (always complex)
Using simulations we compare these agent types in terms of communicative success & complexity across three lexicon sizes. Key finding: interactive repair can considerably ease the computational burden, achieving higher success raters for larger lexicons at lower cost
Note that we're not trying to model what people do in actual interaction — obviously we can and do use both repair and pragmatic reasoning. We model & formalize both strategies to understand them computationally and build better theories of when we might go for which one and why
Naturally the paper comes with open data, and indeed we see the computational framework as another important contribution — these methods pave the way for further theoretical work on how people balance cognitive and interactional resources in interaction https://osf.io/fxphv/
Just want to reiterate what an absolute pleasure it is to work with Jacqueline van Arkel, @mariekewoe and @MarkBlokpoel on these kinds of questions, combining empirical insights on interaction with agent-based models and formal computational work — thanks @NWO_SSH for funding us!
Anyway the paper is short and sweet, and the appendices offer plenty more details on the simulations and computational complexity analysis. Check it out at http://doi.org/10.31234/osf.io/u2y4a and don't miss out on seeing Jacqueline talk at #conll2020 (w/ #emnlp2020) next month
10/10
10/10
The paper is now officially published in the @emnlp2020 @conll_conf proceedings and available in the ACL Anthology — looking forward to #conll2020 #emnlp2020! https://www.aclweb.org/anthology/2020.conll-1.14/
Do check out this great thread by @mariekewoe explaining the paper, covering methods, key findings and some caveats #CoNLL2020 https://twitter.com/mariekewoe/status/1329104396798615552
