Towards unsupervised keyphrase extraction via an autoregressive approach

Abstract

Keyphrase extraction is a technique used to capture the core information of documents and is an upstream task for advanced information retrieval systems, particularly in the academic realm. Current unsupervised methods are primarily built on a score-and-rank framework with a consistent inability to acquire mutual information between extracted keyphrases, especially with graph-based models. Utilizing the autoregressive structure that is typically used in sequence-to-sequence text generation models, we propose a plug-and-play optimizer named C-Decay that can be integrated into any graph-based unsupervised keyphrase extraction model for a stable performance boost, and that mitigates the bias of certain semantically or lexically dominant tokens by optimizing the origin score distribution output by graph-based models directly. The architecture of C-Decay includes the keyphrase pool, the gain vector and the decay factor, where the keyphrase pool is designed to realize an autoregressive structure and the gain vector and the decay factor are the optimization operator. Herein, we examine three graph-based models integrated with C-Decay, and the experiment is conducted on four datasets KDD, Semeval, Nguyen, and Krapivin. Moreover, we prove that C-Decay can improve accuracy and F-Measure by an average of approximately 50% and 20%, respectively.