next previous contents
Next: Typing of attributes Up: Measures Previous: Reliability of measures

Compositional measures

Measures may be compound. That is, the value of an attribute may itself be structured. This is familiar from work in computational linguistics on feature structures. An example drawn from the TEMAA work on evaluating spelling checkers again concerns the lexical coverage of a spelling checker. It may well be that it is not simply the number of entries in the dictionary that matters, but the coverage of high frequency words or of some specialised technical domain, or a composite of both of these with perhaps some other factors. The value taken by an attribute reporting on lexical coverage may therefore be a composite, calculated on the basis of putting together an average or even a weighted average of the values assigned to two or more other attributes. Any one of these attributes may in its turn be a composite: there is no theoretical limit on the depth of the hierarchy of attribute-value pairs. It is clear too that, where a measure uses structured values, the validity of the measure as a whole depends on the validity of its sub-measures.