Probable Human Knowledge Elicitation for Efficient Human-in-the-Loop Text Classification without Any Label-Related Data