Exploiting Phonological Constraints for Handshape Inference in ASL Video