Making Inferences: The Case of Scalar Implicature Processing

Making Inferences: The Case of Scalar Implicature Processing Judith Degen (jdegen@bcs.rochester.edu) Michael K. Tanenhaus (mtan@bcs.rochester.edu) Department of Brain and Cognitive Sciences, University of Rochester Rochester, NY 14627-0268, USA Abstract Scalar implicature has served as a test case for investigating the nature of inference processes in language comprehension. Specifically, the question of whether or not scalar implicatures are computed by default has been extensively investigated in recent years. We argue that the question of default is overly simplistic and propose instead to think of scalar implicature computation as a problem of optimal cue combination within a constraint-based framework. We provide evidence from three experiments supporting the view that multiple constraints of differing strength operate in parallel to provide probabilistic support for or against an implicature. Keywords: experimental pragmatics; scalar implicature; eye- tracking; subitizing Introduction Successful communication requires comprehenders to infer a speaker’s intended meaning from an underspecified utterance. While some information is transmitted via an utterance’s se- mantic content, additional meaning is computed by taking into account pragmatic information about the discourse con- text, as in the sample discourse in (1). Alex: Did you submit your paper? Thomas: Some of the sections are written. Thomas didn’t submit his paper. Some, but not all, of the sections are written. Here, Alex might infer that Thomas, being a cooperative speaker, intends to convey both that he hasn’t yet submitted his paper and that some, but not all, of the sections are written. The former of these inferences is what Grice (1975) termed a particularized conversational implicature (PCI), while the latter is a prototypical example of a generalized conversa- tional implicature (GCI). According to Grice, both of these inferences arise because comprehenders assume that speak- ers conform to certain conventions of rational communicative behavior. The crucial difference between GCIs and PCIs lies in the role that context plays: PCIs are assumed to arise in virtue of special features of the context, while GCIs are as- sumed to arise unless blocked by context. Applied to (1): the inference that Thomas did not submit his paper is tied tightly to Alex’s question and would not have arisen if Thomas’s ut- terance was, for example, an answer to the question whether Thomas had written the introduction yet. However, the infer- ence that some but not all of the sections are written is taken to arise independently of the context. This particular kind of inference, that arises in virtue of a speaker not uttering a rel- evant stronger alternative, is called a scalar implicature. In the example there is a stronger statement the speaker could have made but didn’t (e.g. All of the sections are writ- ten). Under the assumption that a speaker is being as informa- tive as possible, a weak implicature, that the speaker doesn’t know whether all of the sections are written, is licensed. If the hearer further assumes that the speaker is competent with respect to the truth of the stronger statement, the implicature that the speaker believes that some, but not all of the sections are written, is licensed. In recent years the representation and processing of scalar implicatures has emerged as perhaps the defining problem in experimental pragmatics – a subfield of cognitive science that seeks to combine theoretical proposals from linguistics, com- puter science and philosophy with state-of-the-art experimen- tal methods. Importantly for the experimental investigation of the phe- nomenon, scalar implicatures are cancelable, that is, they are defeasible inferences. There are cases in which the scalar im- plicature does not contribute to achieving the discourse goal (Horn, 1984, Sperber & Wilson, 1995, Levinson, 2000). In such cases, where all that is relevant or can be known, is the lower bound (in our example that at least some of the sections are written), the implicature does not arise. Following Katsos, Breheny, and Williams (2005), we will call these contexts lower-bound contexts, and contexts in which pragmatic in- ference is required to achieve the discourse goal upper-bound contexts. Emphasising the GCI-PCI distinction, Levinson (2000) ar- gues that GCIs are pragmatic default inferences that have evolved to maximize the speed of communication. It is only in special contexts that the inference has to be cancelled, where cancellation proceeds in a second, effortful step. In contrast to this default approach, a variety of approaches have viewed scalar implicature as a context-driven process (e.g. Hirschberg, 1991, Sperber & Wilson, 1995). Under these accounts, scalar implicatures are generated as part of the same process as PCIs. The strongest formulation of such accounts is a modular one, whereby pragmatic processing be- gins only once semantic processing is under way (Huang & Snedeker, 2009). That is, generating the implicature requires computing the literal meaning first. A notion of implicit can- cellation is not necessary under the context-driven account since the implicature does not arise in lower-bound contexts in the first place. The default and context-driven approach make different empirical predictions. Under the default model, generating a scalar implicature should be very rapid, indeed as rapid as computing an expression’s literal meaning. An increase in processing effort is predicted only for cases where the impli- cature is cancelled. In contrast, the modular context-driven