Multitask Kernel-based Learning with First-Order Logic Constraints