A simple repair mechanism can alleviate computational demands of pragmatic reasoning: simulations and complexity analysis

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 177–194 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 177 A simple repair mechanism can alleviate computational demands of pragmatic reasoning: simulations and complexity analysis Jacqueline van Arkel1 jacquelinevanarkel@gmail.com, m.woensdregt@let.ru.nl m.dingemanse@let.ru.nl, m.blokpoel@donders.ru.nl Marieke Woensdregt2 1Faculty of Science and Engineering, University of Groningen, Groningen, the Netherlands 2Centre for Language Studies, Radboud University, Nijmegen, the Netherlands 3Donders Centre for Cognition, Radboud University, Nijmegen, the Netherlands Mark Dingemanse2 Mark Blokpoel3 Abstract How can people communicate successfully while keeping resource costs low in the face of ambiguity? We present a principled theo- retical analysis comparing two strategies for disambiguation in communication: (i) prag- matic reasoning, where communicators reason about each other, and (ii) other-initiated re- pair, where communicators signal and resolve trouble interactively. Using agent-based sim- ulations and computational complexity analy- ses, we compare the efficiency of these strate- gies in terms of communicative success, com- putation cost and interaction cost. We show that agents with a simple repair mechanism can increase efficiency, compared to pragmatic agents, by reducing their computational bur- den at the cost of longer interactions. We also find that efficiency is highly contingent on the mechanism, highlighting the importance of ex- plicit formalisation and computational rigour. 1 Introduction Natural languages are rife with ambiguity (Wasow et al., 2005), yet people seem to communicate effi- ciently regardless. How can people communicate successfully in the face of ambiguity while keeping resource costs low? There seem to be at least three strategies communicators have at their disposal. First, contextual information can be used to disam- biguate the speaker’s intended meaning (Piantadosi et al., 2012; Sperber and Wilson, 1986; MacDonald et al., 1994), though context-sensitive computations are notorious in computational cognitive science for the astronomical demands they make on com- putation time (Fodor, 2000; Haselager, 1997; van Rooij et al., 2011). Second, pragmatic reasoning allows taking into account the speaker’s goal (e.g. ‘being informative’) (Grice, 1975; Sperber and Wil- son, 1986; Goodman and Frank, 2016), but this alone is not always enough to fully disambiguate meaning (Schegloff, 1992). Finally, communica- tors can leverage the interaction itself by explicitly requesting clarification (e.g. by asking ‘Huh?’ or ‘Who?’) in a process known as other-initiated re- pair (Schegloff et al., 1977; Purver et al., 2018). This provides a possible way for communicators to reduce their computational burden through inter- action, potentially increasing communicative effi- ciency (Dingemanse, 2020). To investigate the computational plausibility of this potential gain in communicative efficiency we present a theoretical analysis of other-initiated re- pair and pragmatic reasoning. Following Gibson et al. (2019), we define efficient communication as communication in which participants reach mu- tual understanding while requiring minimal effort in terms of resource costs (deconstructed here as the sum of computational and interactional cost). We compare a novel agent-based model of other- initiated repair with one of pragmatic reasoning (Goodman and Frank, 2016) for both their commu- nicative success and use of computational and in- teractional resources. Simulations are used …

Content truncated to 500 words