A Structured Model for Joint Learning of Argument Roles and Predicate Senses