Improved score aggregation for authorship verification

The Impostors method is one of the most successful solvers of author verification problems. Given a pair of texts, it aims to find whether the same author wrote them or not. This paper describes a proposed approach with the primary objective of achieving a higher classification accuracy. This higher accuracy is achieved by modifying the vector representations of input texts, such that the effect of them possibly being in different domains is reduced. Such vector modification factors are obtained by the addition of a computational step that empirically estimates the expected difference, or ratio, between the questioned texts’ similarity scores against their in-domain samples. Our evaluation confirms that our proposed approach is capable of achieving higher classification accuracy than the original method. Despite the size of the evaluation dataset, some of the increases in the classification accuracy are large enough to allow for observing statistically significant, very significant, and highly significant gains.