Handling Long-Tail Queries with Slice-Aware Conversational Systems