Thinking like an Author: A Zero-Shot Learning Approach to Keyphrase Generation with Large Language Model

Abstract

Keyphrase generation aims to automatically derive a set of phrases from a given document. Recently, automated keyphrase generation has gained prominence as a focal point of research in both industry and academia, showcasing notable advancements in performance. This paper aims to employ a Large Language Model (LLM) for a more comprehensive extraction of keyphrases from documents. We observe that many article authors adhere to a four-step process when selecting keyphrases for a document. Initially, they extract keyphrases directly from the document. In addition, some authors will also choose hypernyms of the keyphrases selected in the previous stage as new extended keyphrases. Subsequently, these authors obtain keyphrases from documents with similar topics, opting for the most appropriate ones to include. Finally, the authors organize the keyphrases obtained from the extraction, extension, and retrieval steps, discerning the final keyphrases through a ranking process. Motivated by these observations, we introduce a zero-shot learning approach for keyphrase generation using the LLM, which includes four parts. The extractor, extender, and retriever components are responsible for recalling candidate keyphrases, while the ranker component is tasked with ranking them. Experimental results demonstrate the effectiveness of our approach on various datasets. Moreover, ablation experiments shed light on the impact of each module on the model’s final performance.