A Comparison of Different Loss Computations in Siamese Networks for Authorship Verification

Abstract

In this study, we consider the author verification problem as a binary classification problem where the aim is to identify whether the two inputs belong to the same author or not. For this task, we have used several Siamese convolutional neural networks-based models. The first model employs a Siamese network which is trained using binary cross-entropy loss after the absolute distance computation. In addition to this baseline model, we have implemented another model which employs a concatenation operation. Moreover, two contrastive loss-based models have also been implemented for the same task. Two publicly available benchmark datasets, IAM and CVL, have been used in the study. Training, validation and test datasets that have been used to train and test each of those models have been generated from those datasets. Experimental results show that the performances of the two cross-entropy loss-based models are similar and much better than that of the contrastive loss-based models.