ABSTRACT
The spoken performances of the test takers must be rated in some way. It is almost axiomatic that, because language use is a multicomponential phenomenon, requiring interlocutors to negotiate meanings, no two listeners hear the same message. This aspect of language use is a source of bias in test scores. It leads language test developers to severely limit which features of a performance they require raters to attend to in making their ratings. They hope that, if raters focus attention only on pronunciation, grammar, fluency and comprehensibility, for example, the many other features of the discourse will not influence them. There is mounting evidence that this is a vain hope.
