Reference Resolution in the Wild: On-line circumscription of referential domains in a natural, interactive problem-solving task

Abstract

Reference Resolution in the Wild: On-line Circumscription of Referential Domains in a Natural, Interactive, Problem-solving Task. Sarah Brown-Schmidt (sschmidt@bcs.rochester.edu) Department of Brain and Cognitive Sciences, University of Rochester Rochester, NY 14627 Ellen Campana (ecampana@bcs.rochester.edu) Department of Brain and Cognitive Sciences, University of Rochester Rochester, NY 14627 Michael K. Tanenhaus (mtan@bcs.rochester.edu) Department of Brain and Cognitive Sciences, University of Rochester Rochester, NY 14627 Abstract We examined how naive conversational participants circumscribed referential domains during the production and comprehension of referring expressions by monitoring participants' eye movements during a referential communication task. The results replicated some well- established results, e.g., incremental reference resolution, demonstrating the feasibility of studying real-time language comprehension in interactive conversation. We also observed a high proportion of underspecified referential expressions that were easily understood by addresses because of discourse and pragmatic constraints, including constraints developed as a result of the conversation. Background In characterizing work in language performance, Clark (1992) pointed out that the field has been largely divided into two traditions. One tradition, the language-as-action tradition, emphasizes interactive conversation as the most basic form of language use. According this tradition the principles of language performance and language design cannot be understood without taking into account the interactive collaborative processes that are embedded in conversation. A central tenet in work within this tradition is that utterances can only be understood within a particular context, which includes the time, place and participant's conversation goals. Thus researchers within this tradition have focused primarily in investigations of interactive conversation using natural tasks, typically with real-world referents. A second tradition, the language-as-product tradition, focuses primarily on the processes by which listeners decode (and speakers encode) linguistic utterances. Psycholinguistic work on language comprehension within in this tradition typically examines moment-by-moment processes in real-time language processing using fine- grained reaction time measures. The rationale for using these measures is that comprehension processes are closely time-locked to the linguistic input which, for spoken language, unfolds over time. Until recently, the real-time response measures in the psycholinguist's toolkit required the use of de-contextualized language, typically pre- recorded sentences presented in impoverished contexts. This constraint ruled out real-time investigations of natural, interactive conversation. Moreover, a dominant theoretical perspective within the product tradition was that initial core processes (e.g., lexical access and syntactic processing) were informationally encapsulated from contextual influences (e.g., Fodor, 1983). Recently, the advent of light-weight head-mounted eye- tracking systems has made it possible to investigate real- time comprehension in more natural tasks, such as tasks where participants follow spoken instructions to manipulate objects in a task-relevant visual world (Tanenhaus, Spivey-Knowlton, Eberhard & Sedivy, 1995). Fixations to task-relevant objects are closely time-locked to the unfolding utterance, providing a continuous real-time measure of comprehension processes at a temporal grain fine enough to track the earliest moments of lexical access, parsing and reference resolution (Tanenhaus, Magnuson, Dahan & Chambers, 2000). A growing body of research employing eye-tracking techniques demonstrates clear effects of contextual constraints. For example, syntactic ambiguity resolution is influenced by referential constraints provided by the visual context, including the number of potential referents and their properties (Tanenhaus et al., 2000). Moreover, some recent work using confederates in constrained tasks suggests that under some circumstances information provided by knowledge of the speaker's perspective and intentions can affect even the earliest moments of comprehension (Hanna, However, a major limitation of previous work is that all of the language used has come from scripted language, ruling out spontaneous collaborative processes that are likely to underlie circumscription of referential domains,