Optimizating one-class techniques applied to verify information extractors

One-class techniques are classification algorithms, largely unsupervised, that learn using a single class. The problem of verify the information obtained by an informaction extractor, could be considered as a one-class problem because the verifier is build only using instances of classes we want to extract. We propose the use of a multi-level classifier based on One-class techniques to solve the problem of verify information. The need for this new proposal arises from the bad behavior of One-class techniques when they use categorical characteristics. As we shall see, its use significantly improves the performance of all of the algorithms studied. To evaluate the performance obtained by these techniques and modifications, we use different databases proposed in the literature as well as nonparametric statistical test that will help us strengthen the statistical significance of the results achieved.