The focus is on the utility of bilinear models for score building in contingency tables and contrasts it with the correspondence analysis approach. The groundwork for using bilinear models for score building was laid time ago, and a set of rules for the instrumental variable against which the target variable is classified has been previously specified. Typical bilinear models used for this purpose include the log-multiplicative model and the correspondence analysis model. While this approach seems at first sight to be promising, there are issues relating to empty cells and sample size which often mean that the model fails to form exactly as intended. We discuss whether the mentioned rules need extending and determine whether similar rules are needed for correspondence analysis. An example is used from the problem of scaling crime harm and impact from survey data.