Richard Gerrig - US grants

A listener's task in everyday conversation is to determine what speakers mean by uttering particular sentences on particular occasions. The speaker's meaning of a sentence can vary considerably among occasions of use. For example, an utterance of "It's hot in here" may communicate (among an unlimited number of possibilities) simply the information that a room is hot, a request that a window should be opened, or a reminder that fuel should be conserved. In each case, a listener must recover the specific meaning that the speaker intended. The experiments proposed here are designed to examine some of the properties of speakers' meanings. They fall under two headings: 1. Priming by speaker's meaning. Suppose a speaker says, "The cows have come home" and means by that, "For once, Bruce isn't whining about money." What is the psychological status of this speaker's meaning? Specifically, in what sense is the indirect communication of this meaning equivalent to a direct statement? Priming techniques will provide data on the properties of representations of these speakers' meanings. 2. The resolution of speaker-sensitive ambiguities. Listeners must often examine the knowledge they share with a speaker to recover his or her intended meanings. For example, to understand the ambiguous utterance "I did what I said I was going to do", a listener must access from memory an appropriate shared episode. How do listeners exploit shared knowledge to resolve such ambiguities? A variety of techniques will be used to determine by what processes listeners are able to be sensitive to the dependence of meaning on speakers' intentions.

This project probes the nature of psychological representations of discourse and applies the results to a computational model of "who knows what" in text about multiple agents. Many kinds of inferences and pragmatic effects depend on knowing who has access to which information, as well as who shares common ground. Computational and psychological models of text understanding have largely ignored the ramifications of multiple agent sources on the representation and use of textual information. Our objectives are: A) to examine how people process and represent the distinct knowledge states of agents depicted or quoted in both constructed and naturally occurring texts, B) to investigate how readers' representations of texts are affected by the perspectives they bring to these texts, as well as by how sources are presented, and C) to apply these empirical results to the construction of a partly-automated system for extracting, representing, and coding "who knows what" in a large corpus of naturally occurring texts - news stories from The Wall Street Journal. This research will contribute to practical goals concerning the modeling and extraction of information from text and reported speech, as well as to theoretical goals concerning how people process and represent texts about multiple agents.

During remote collaboration, partners have access to much less information than during face-to-face collaboration. This project uses psychological experiments to examine how knowing where a partner is looking affects performance and strategies on collaborative tasks. Phase 1 asks: When partners are physically co-present, how aware are they of where the other is attending, and how do they achieve shared attention? Phase II applies these basic results to remote collaborations; partners wear eye-trackers that transmit gaze information to each other's computer displays. A space of tasks and representations is explored: Tasks are varied in systematic ways (e.g., some lend themselves to parallel activity, while others require consensus for each step), and different representations of the same gaze information are compared. The goal is to understand which representations work best for which tasks. With technological advances making eyetracking easier, less cumbersome, and more affordable, a gaze-based computer interface may someday join the ranks of ubiquitous input devices like the mouse. If this technology is to be integrated into the "every citizen interface," it is necessary to understand how people use the information in gaze to achieve a joint focus of attention. This could provide the foundations for new technology for computer-mediated collaboration.

This project takes an innovative and multidisciplinary approach to characterizing how spoken language production and interpretation are coordinated in dialog. It examines how people adapt to both human and computer conversational partners, through variation in pronunciation (e.g, dialect), rhythm, word choice, sentence structure, and perspective. Adaptations include those that make processing easier for both partners, such as converging on the same wording; they also include adjustments made by one partner "for" the other. Findings from human dialog will be applied to systems that use speech recognition and generation, with the goals of (1) getting users to adapt utterances to forms the system can process more robustly, and (2) whenever feasible, adapting the system's vocabulary, dialect, and perspective to the user's' needs.

The project brings together methods and theoretical perspectives from computer science, linguistics and psychology to advance theories and improve applications. Methods include laboratory experiments, corpus studies, and simulation studies, integrated with prototyping and evaluation of spoken dialog systems. Three applications are planned: a picture matching game, a PDA-based calendar system, and a telephone-based course evaluation system for an undergraduate community.

The project will enhance training of young scientists in computer science, linguistics and psychology, and will include underrepresented groups in basic research and user-interface engineering. The broader impact will be a scientific foundation for developing flexible and robust spoken dialog systems that serve the needs of diverse users. Specific application opportunities include mobile computing, educational assessment, machine speech recognition of non-standard pronunciation, and tutoring systems for the hearing impaired or second-language learners.

When people reformulate a problem space, previously unseen structure emerges. This process can be decomposed into two steps: People must first recognize and then exploit novel structure. We suggest that both of these steps can be improved by experienced application of creative nominalization. Here, nominalization refers to the process of recognizing a novel concept and naming it appropriately. This project demonstrates that experience in nominalization can improve problem solving and that successful training and experience on nominalization has the potential to enhance people?s intrinsic motivation, and thereby effectiveness, with respect to creative aspects of problem solving. In parallel, the project explores the potential for nominalization as a strategy to enhance machine-learning agents in reinforcement learning environments. Inspired by research on animal learning, reinforcement learning is a branch of artificial intelligence research concerned with creating motivated, learning agents. In the reinforcement-learning setting, nominalization has the potential to create a first-class object, something that can be directly manipulated, recorded, analyzed, and composed with other objects to form higher-order structures. In addition, reinforcement-learning researchers have recently begun to consider how learning might be enhanced with intrinsic motivation to explore problem spaces. Thus nominalization can function in reinforcement-learning settings both as a direct strategy and indirectly via intrinsic motivation. The most significant broader impact of this project will be to provide a new intervention that will enhance the creativity and efficacy of problem solvers working alone or in collaborative groups. If successful, the relative simplicity of the intervention and its general applicability would make it a prime candidate for wide dispersal to people in disparate walks of like.