Fast and Constrained Absent Keyphrase Generation by Prompt-Based Learning

• DOI: 10.1609/aaai.v36i10.21402
• Publication Date: 2022-07-04T11:46:30Z
• Authors: Huanqin Wu; Baijiaxin Ma; Wei Liu; Tao Chen; Dan Nie
• Keywords: Sequence (biology); Text generation

Generating absent keyphrases, which do not appear in the input document, is challenging keyphrase prediction task. Most previous works treat problem as an autoregressive sequence-to-sequence generation task, demonstrates promising results for generating grammatically correct and fluent keyphrases. However, such end-to-end process with a complete data-driven manner unconstrained, prone to generate keyphrases inconsistent document. In addition, existing decoding method makes of must be done from left right, leading slow speed during inference. this paper, we propose constrained prompt-based learning fashion. Specifically, prompt will created firstly based on keywords, are defined overlapping words between Then, mask-predict decoder used constraint prompt. Experiments benchmarks have demonstrated effectiveness our approach. evaluate performance information retrieval perspective. The result shows that approach can more consistent improve document performance. What’s more, non-autoregressive manner, model up by 8.67× compared method.