Natural Language Generation and Discourse Context: Computing Distractor Sets from the Focus Stack