Analyzing the Use of Character-Level Translation with Sparse and Noisy Datasets