Precision and recall are the most used performance measures in pattern recognition and information retrieval to evaluate algorithms for predicting the classes. Precision is the number of correct results (true positives) divided by the number of all results. Respectively, recall is the number of correct results divided by the number of expected results. Their main limitation is that this ground truth (G) and the predicted results (P) must come from the same finite set, which is often small and mutually exclusive. However, these assumptions do not always hold in every application.

Here, we generalize the classical precision and recall measures to be used with similarity measures. We introduce new measures called soft precision and soft recall using soft cardinality concept developed by Jimenez et al. We also derive soft F1 measure from the soft recall and precision measures. We demonstrate how the new measure can be applied in natural language processing using both syntactic and semantic text similarity measures.

You may edit the fields in the example below: