Why is a new model needed? Almost all work on conversation in philosophy and linguistics still adheres to a view put forward by Paul Grice over forty years ago, according to which conversation is driven by cooperative and even shared intentions amongst the participants, who are viewed as ideally rational agents. At the core of this view are cooperative principles such as: normally one believes what one says and one normally tries to help one's interlocutors achieve their goals. The latter in turn requires speakers to adopt shared intentions. While much of this view is intuitive, Grice's ideas have proved difficult to make into a formal and testable linguistic theory. And when researchers in linguistics and philosophy have succeeded in formalizing Grice, their sophisticated symbolic or logic based models of semantics and pragmatics contain relatively primitive and impractical models of agent interaction and decision making in conversation. The strong cooperative principles underlying Grice's view are not borne out in many real conversations.
Consider, for instance, the following exchange from a book on courtroom conversations:
In contrast to linguistic models of conversation, game theory proposes a sophisticated view of strategic decision making and agent action. Economists have studied issues like credible information that are pertinent to strategic conversation and have provided important results on when information can be relied upon in strategic settings [#!farrell93!#,#!rabin90!#,#!lipman:sippi:1995!#,#!lipman:2003!#]. This has had important consequences for economics at large: imperfections in information affect the efficiency of traditional free markets, as the economist Joseph Stiglitz has argued in his Nobel Prize address. But game theory tends to assume a very minimal view of the complex relationship between the linguistic signal and its relations to the discourse context that determine what is communicated in a conversation. Furthermore, game theorists usually assume that agents have a complete knowledge of their preferences and of the possible strategies in a conversation and that they are rational utility maximizers.
These assumptions of game theory are problematic in conversational settings. Communication with natural language is a highly noisy channel; people have to resolve ambiguities and the information conveyed is not fully determined by linguistic form but also by a context that is only partially observable and often not known. For instance, Bronston's prosecutor does not have accurate information about Bronson's conversational strategy in (c); he infers the implicature based on assumptions about the context. Such implicatures are crucial to determining what is the ``mistake'' that the prosecutor makes. Work in game theory on signalling or bargaining does not study such implicatures. In the example () below, it appears that B has not considered all the strategic options but rather discovers them as conversation proceeds.
This project combines the strengths of linguistic and game theoretic models of conversation to produce a sophisticated theory of discourse interpretation, informed by observations about the moves that agents actually make in discourse. On this model preferences, not cooperative principles, drive conversational moves, and speakers use conversation to elicit preferences from their partners. What agents say is based on what they perceive as maximizing the satisfaction of their preferences, given their estimation of the preferences of other conversational participants and of their importance in achieving the agents' goals. Preferences of conversational agents are sometimes aligned but often divergent. Grice's cooperative understanding of conversation will be a special case of my more general view of conversation as a strategic activity. Not only will this view mark a major change in linguistic and philosophical thinking about discourse, but it will also benefit game theoretic approaches to communication with applications to the theory of credible information.