This paper proposes a new classification scheme for reading test items and reports on its usefulness and limitations when actually applied to three reading tests. In this scheme, an item is placed in one of nine categories based on its position on two dimensions. The first dimension concerns the size of the relevant text portion. If an item can be correctly answered by processing one sentence or a smaller portion, it will be placed in the category SENTENCE; if an item requires two successive sentences to be processed, it will be categorized as LOCAL, and if two non-successive sentences or three or more successive sentences need to be comprehended, the item will be termed GLOBAL. The second dimension concerns the type of cognitive activity required to respond correctly to an item. If an item can be answered by merely comprehending the explicit meaning of the portion and if key words used in the question and the relevant text are identical, the item will be placed in the category VERBATIM; if the answer is explicitly stated but in paraphrase, it will be termed PARAPHRASE; and if the implicit meaning of the relevant part needs to be comprehended to answer the item, it will be classified as IMPLICIT. Thus, each dimension having three categories, the result will be a 9-category (3 categories times 3 categories) taxonomy. Four testing experts classified three reading tests, consisting of 35 items each, by this classification scheme. The results indicated somewhat less than optimal inter-classifier reliability, when the VERBATIM/PARAPHRASE/IMPLICIT dimension was involved, in particular. To compensate for the low agreement rate, Rasch Rating Scale Model (Wright & Masters 1982) was applied to spot 'misfit' ratings to be cyclically deleted and to compute the final position (in Rasch logits) of each item on each dimension. The resultant scatter plots graphically illustrated the characteristics of each reading test, implying a high practical utility of the proposed taxonomy.
