I examine the Brown and Levinson (1978) model of politeness. On the assumption that a model of face attention must involve impolite as well as polite interactive behaviour, I construct an analogous model to deal with what I term Face Attack Acts. I show that an extension of a politeness model in this way reveals serious flaws in some hypotheses central to Brown and Levinson's work. I apply the principles of Relevance Theory (Sperber and Wilson 1986)to the extended model, and show how the theory can offer an explanatory account of face attention, on a continuum from polite to impolite. Relevance Theory makes predictions about the interpretation of utterances in context, which explain how the face-oriented aspect of interaction is recovered by hearers. I show that Relevance Theory provides a motivated way of linking utterances with facts about power asymmetries and group-membership which are seen to be recoverable by the interpretive process. I apply the resulting face-attention model of utterance interpretation to examples of the use of language to encode power and communicate assumptions about social behaviour and status. In terms of previous accounts of utterance interpretation, particularly Grice's Cooperative Principle and maxims, the account of face attention which incorporates Relevance Theory has greater explanatory power. In practical terms, this application of Relevance Theory is shown to be illuminating in raising the assumptions underlying non-cooperative communication to a conscious level, at which their validity can be sustained.