PromptRank: Unsupervised Keyphrase Extraction Using Prompt

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9788–9801 July 9-14, 2023 ©2023 Association for Computational Linguistics PromptRank: Unsupervised Keyphrase Extraction Using Prompt Aobo Kong1 Shiwan Zhao2 Hao Chen3 Qicheng Li1∗ Yong Qin1 Ruiqi Sun3 Xiaoyan Bai3 1TMCC, CS, Nankai University 2Independent Researcher 3Enterprise & Cloud Research Lab, Lenovo Research 1kongaobo9@163.com 2zhaosw@gmail.com 1{liqicheng, qinyong}@nankai.edu.cn 3{chenhao31, sunrq2, baixy8}@lenovo.com Abstract The keyphrase extraction task refers to the au- tomatic selection of phrases from a given doc- ument to summarize its core content. State- of-the-art (SOTA) performance has recently been achieved by embedding-based algorithms, which rank candidates according to how simi- lar their embeddings are to document embed- dings. However, such solutions either strug- gle with the document and candidate length discrepancies or fail to fully utilize the pre- trained language model (PLM) without further fine-tuning. To this end, in this paper, we pro- pose a simple yet effective unsupervised ap- proach, PromptRank, based on the PLM with an encoder-decoder architecture. Specifically, PromptRank feeds the document into the en- coder and calculates the probability of generat- ing the candidate with a designed prompt by the decoder. We extensively evaluate the proposed PromptRank on six widely used benchmarks. PromptRank outperforms the SOTA approach MDERank, improving the F1 score relatively by 34.18%, 24.87%, and 17.57% for 5, 10, and 15 returned results, respectively. This demon- strates the great potential of using prompt for unsupervised keyphrase extraction. We release our code at this url. 1 Introduction Keyphrase extraction aims to automatically select phrases from a given document that serve as a suc- cinct summary of the main topics, assisting read- ers in quickly comprehending the key information, and facilitating numerous downstream tasks like information retrieval, text mining, summarization, etc. Existing keyphrase extraction work can be di- vided into two categories: supervised and unsuper- vised approaches. With the development of deep learning, supervised keyphrase extraction methods have achieved great success by using advanced ar- chitectures, such as LSTM (Alzaidy et al., 2019; ∗Qicheng Li is the corresponding author. Sahrawat et al., 2020) and Transformer (Santosh et al., 2020; Nikzad-Khasmakhi et al., 2021; Mart- inc et al., 2022). However, supervised methods re- quire large-scale labeled training data and may gen- eralize poorly to new domains. Therefore, unsuper- vised keyphrase extraction methods, mainly includ- ing statistics-based (Florescu and Caragea, 2017a; Campos et al., 2020b), graph-based (Bougouin et al., 2013; Boudin, 2018), and embedding-based methods (Bennani-Smires et al., 2018; Zhang et al., 2022), are more popular in industry scenarios. Recent advancements in embedding-based ap- proaches have led to SOTA performances that can be further divided into two groups. The first group of methods, such as EmbedRank (Bennani-Smires et al., 2018) and SIFRank (Sun et al., 2020), embed the document and keyphrase candidates into a la- tent space, calculate the similarity between the em- beddings of the document and candidates, then se- lect the top-K most similar keyphrases. Due to the discrepancy in length between the document and its candidates, these approaches …

Content truncated to 500 words