2018 Friday Poster 6649

Friday, November 2, 2018 | Poster Session I, Metcalf Small | 3pm

Modeling the role of common ground in pragmatic word learning
M. Bohn, M. Tessler, M. Frank

Language is understood and learned in context. To resolve ambiguity, speakers and listeners make context sensitive inferences based on expectations they have about their communicative partner. In the Rational Speech Act (RSA) framework, hypothetical interlocutors make inferences about the interpretation of literal utterances based on the assumption that their partner communicates in a helpful but parsimonious (rational) way. This approach has been very productive in making accurate predictions about how humans use and interpret ambiguous utterances (see e.g. Goodman & Frank, 2016). But expectations about communication are also person specific: they pertain to the common ground shared between interlocutors (Clark, 1996,  Tomasello, 2008). From a modeling perspective, this aspect of pragmatic inference has been largely neglected. Here we extend earlier work by incorporating speaker specific expectations into probabilistic models of language comprehension.

We conducted a series of experiments with adults (methods, analysis and model predictions pre-registered – now in progress with children). In Experiment 1, participants (N = 40) expected speakers to produce utterances in a way that maximizes their informativity in context (paradigm Fig. 1, left; results fig. 2, left, p = 0.003), replicating earlier findings (Frank & Goodman, 2014). In Experiment 2, we manipulated speaker specific expectations about language use. Children expect particular speakers to refer to things that are new to them (Akhtar, et al., 1996) as well as to things that they previously preferred (Saylor, et al., 2004). We implemented these manipulations by changing the way the speaker interacted with the potential referents prior to her utterance, flagging them as novel or preferred (Fig. 1, right). Participants (N = 80) identified the preferred/novel object as the target more often if the speaker remained the same (Fig. 2, ppreference < .001, pnovelty= .003).

In Experiment 3, we combined the general and speaker specific manipulations from Experiment 1 and 2. Prior to data collection, we formalized the informativeness inference studied in Experiment 1 in a RSA-type probabilistic model. Next, we used the results of Experiment 2 to incorporate speaker specific expectations and derived quantitative model predictions about how speaker specific expectations should influence the general expectation that speakers are informative. Finally, we compared the model predictions to empirical data from adults (N = 120). We found a very tight fit between the two (Fig. 3, rmodel – data = .96, p < .001). The assumptions about the cognitive processes underlying pragmatic inference represented in the RSA model accurately captured adults’ behavior. An alternative model, relying on speaker specific expectations only, provided a similar but slightly worse fit and did not capture important qualitative crossovers when cues were in conflict. In a pre-registered follow-up experiment, we manipulated the strength of speaker specific expectations to further investigate the scope of the model.

Taken together, our study shows how empirical and modeling approaches can be used in conjunction to gain a more precise and explicit understanding of language comprehension in context. Here, this approach specified the importance of common ground for pragmatic inference. It also lays out a potential roadmap for how to study pragmatic inference developmentally.

References

Akhtar, N., Carpenter, M., & Tomasello, 645 (1996). The role of discourse novelty in early word learning. Child Development, 67, 635– 645.

Clark, H. H. (1996). Using language. Cambridge University Press.

Frank, M. C., & Goodman, N. D. (2014). Inferring word meanings by assuming that speakers are informative. Cognitive Psychology, 75, 80-96.

Goodman, N. D., & Frank, M. C. (2016). Pragmatic language interpretation as probabilistic inference. Trends in Cognitive Sciences, 20, 818-829.

Saylor, M. M., Sabbagh, M. A., Fortuna, A., & Troseth, G. (2009). Preschoolers use speakers’ preferences to learn words. Cognitive Development, 24, 125–132.

Tomasello, M. (2008). Origins of human communication. MIT press.