CPSRank: Unsupervised Keyphrase Extraction via Contextual Perturbation

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

1

초록

The importance of a phrase within a document becomes most evident through its absence rather than its presence. Inspired by this observation, we redefine keyphrases as those whose removal most disrupts the document's meaning. Traditional unsupervised methods typically rely on document-level signals, such as term frequency or phrase-to document similarity, which overlook the contextual contribution of a phrase. This paper proposes CPSRank, an unsupervised keyphrase extraction method that evaluates the semantic importance of candidate phrases via a contextual perturbation score (CPS). The CPS quantifies the critical role of each phrase by combining contextual perturbation and content loss. CPSRank outperforms existing baselines in terms of F1 scores while providing deeper insights into the semantic value of keyphrases. We release our code at https://github.com/Splo2t/CPSRank.

키워드

contextual perturbation; information extraction; keyphrase extraction; semantic importance; unsupervised learning
제목
CPSRank: Unsupervised Keyphrase Extraction via Contextual Perturbation
저자
Yu, Hyunwook; Kim, Minju; Kim, Euijin; Kim, Mucheol
DOI
10.1145/3746252.3760945
발행일
2025
유형
Proceedings Paper
저널명
CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management
페이지
5454 ~ 5458