Дата: 17-07-2022, 23:12
So as to acknowledge the case that there’s no corresponding entity of the queried slot kind, we introduce token to pad the output, and in follow, we use "none" as token to make the model output extra natural. So as to tag multiple entity of the identical slot type, we introduce ";" as to divide multiple entity of the identical slot type. Such inverse prompting only requires a one-flip prediction for every slot sort and drastically accelerates the prediction. 1) We introduce the concept of inverse prediction to prompting strategies for slot tagging tasks, which vastly hurries up the prediction process. V is the variety of label types (4-instances in Fig. 1), which therefore drastically hurries up the prediction. For the instance in Fig. 1, we use an inverse prompt to change the input as "book a flight from Beijing to New York tomorrow morning. On the coaching time, we pre-construct the prompt with answers comparable to "book a flight from beijing to new york tomorrow morning" departure refers to new york . This part will present how to perform coaching and inference with the prompts.