A Survey on Recent Advances in Keyphrase Extraction from Pre-trained Language Models

Findings of the Association for Computational Linguistics: EACL 2023, pages 2153–2164 May 2-6, 2023 ©2023 Association for Computational Linguistics A Survey on Recent Advances in Keyphrase Extraction from Pre-trained Language Models Mingyang Song, Yi Feng and Liping Jing∗ Beijing Key Lab of Traffic Data Analysis and Mining Beijing Jiaotong University Beijing, China {mingyang.song,21112027,lpjing}@bjtu.edu.cn Abstract Keyphrase Extraction (KE) is a critical com- ponent in Natural Language Processing (NLP) systems for selecting a set of phrases from the document that could summarize the important information discussed in the document. Typ- ically, a keyphrase extraction system can sig- nificantly accelerate the speed of information retrieval and help people get first-hand infor- mation from a long document quickly and ac- curately. Specifically, keyphrases are capable of providing semantic metadata characterizing documents and producing an overview of the content of a document. In this paper, we in- troduce keyphrase extraction, present a review of the recent studies based on pre-trained lan- guage models, offer interesting insights on the different approaches, highlight open issues, and give a comparative experimental study of pop- ular supervised as well as unsupervised tech- niques on several datasets. To encourage more instantiations, we release the related files men- tioned in this paper1. 1 Introduction Keyphrase extraction is a fundamental task in NLP for identifying and extracting a set of keyphrases from the document that could summarize the im- portant information discussed in the source docu- ment (Hasan and Ng, 2014; Papagiannopoulou and Tsoumakas, 2019). Keyphrases have enabled ac- curate and fast searching for the document from a large text corpus and have exhibited their potential in improving many NLP tasks, such as text summa- rization (Zhang et al., 2004). Various information filtering and extracting techniques are becoming critical with the ever-increasing amount of text data. Owing to its potential importance, keyphrase ex- traction has received more and more attention from ∗Corresponding author. 1https://github.com/MySong7NLPer/ KeyphraseExtractionSurvey NLP researchers. However, the keyphrase extrac- tion task is far from being solved: state-of-the-art performance on keyphrase extraction is still lower than other core NLP tasks. Our goal in this pa- per is to investigate the state-of-the-art models in keyphrase extraction, examine the primary sources of errors made by existing systems, and discuss the challenges ahead. The first keyphrase extraction task was organized by Turney (1999), which defines the keyphrase ex- traction task as “the automatic selection of impor- tant and topical phrases from the body of a doc- ument”. Since then, there have been numerous keyphrase extraction models (Witten et al., 1999; Turney, 2000; Tomokiyo and Hurst, 2003; Hulth, 2004; Wan and Xiao, 2008a; Jiang et al., 2009; Liu et al., 2009; Grineva et al., 2009; Nguyen and Phan, 2009; Bougouin et al., 2013; Caragea et al., 2014; Danesh et al., 2015; Bougouin et al., 2016; Florescu and Caragea, 2017a; Campos et al., 2018a; Alzaidy et al., 2019). In the past two decades, keyphrase extraction methods have experienced the develop- ment from traditional approaches to deep learning methods (Hasan and Ng, 2014; Papagiannopoulou and Tsoumakas, 2019). …

Content truncated to 500 words