Abstract
Choosing the optimal threshold for the collocations extraction remains a manual task performed by experts. Until today, there is no serious work, based on deep studies, which explores possible solutions to automate the learning of the threshold in the statistical terminology field. In this paper, the authors try to spotlight on this problem by exploring, firstly, the evaluation performance techniques used in several scientific areas (such as biomedical and biometric) and applying them, subsequently, on the statistical terminology field. The experimental study gives promoters results. First, it shows the effectiveness of usual techniques (such as ROC and Precision-Recall curves) used to evaluate the performance of binary classification systems. Second, it provides a practical solution for automatic estimation of optimal thresholds for collocation extraction systems.