Nearest Neighbor Prompt Learning Model for Few-shot Process Text Entity Recognition

To address the entity extraction from the process text under few-shot settings, this paper proposes a Nearest Neighbor Prompt Learning Model (NN-PLM). This model utilizes word-building role of characters in the process text to model an explicit memory, and retrieves character-label pairs which have similar word-building role with the predicted characters through a nearest-neighbor search. By correcting the prediction probability of the prompt learning model, the NN-PLM model improves entity recognition performance under few-shot settings. Experimental results demonstrate the effectiveness of NN-PLM compared with nine methods, with 11.88% and 7.62% improvements in F₁ in 5-shot and 10-shot settings, respectively.