ier should perform well on the example. Dredze et al. (2009) consider several variants of MDR. We select the two best performing methods: MDR-L2, which uses the underlying algorithm of Crammer et al. =-=(2008)-=-, and MDR-KL, which uses the underlying algorithm of Dredze et al. (2008). We follow their approach to classifier training and parameter optimization. MTRL The multi-task relationship learning (MTRL)