Abstract: In large-scale and high-stakes test rating, we often have to mark unanswered test-items. If these items can be recognized automatically, the rating efficiency will be improved and the cost will be reduced. This study intends to develop a neural network based recognition technology of blank test items, and discuss its application in CET constructive item rating. While extracting the standard deviation of the standard deviations of row and column vectors of image pixel gray value matrix as characteristic parameters, choosing Elman as experimental model, traindx as training function, learngdm as learning function, tansig and logsig as the transfer function of hidden layer and output layer, and optimizing through adjusting the number of neurons in hidden layer, this technology succeeds in obtaining a good recognition effect with little computing effort, thus saving human resources while guaranteeing the correctness of recognition.