Confidence Learning from Noisy Labels for Arabic Dialect Identification

In this paper, we propose a new deep learning network for Arabic Dialect Identification (ADI) that addresses the incorrect label problem using confidence information. The dataset recently released for the MGB-5 ADI challenge includes a small amount of verified labels but a large amount of noisy labels,
which makes the ADI task very challenging. We propose a confidence learning network (CLN) that utilizes a multi-task learning strategy to handle confidence information by leveraging the verified and noisy label sets. The proposed CLN employs a confidence refinement module using a Gumbel softmax sampler that generates representations with more discriminative capabilities. Experimental results demonstrate that the proposed CLN model shows higher classification accuracy than conventional state of-the-art ADI systems.