Hierank: Hierarchical Ranking Model for Keyphrase Extraction

The semantics of natural language texts contain hidden hierarchical structures, and utilizing hierarchical semantic structures in keyphrase extraction tasks can help models gain a deeper understanding of text semantics. In this paper, we propose HIERank, a keyphrase extraction method based on hierarchical semantic structure, which combines multilevel information to construct hierarchical semantic relevance. Specifically, we constructed three-level hierarchical relationships, phrase-sentence, sentence-document and document-document corresponding structural relationships, to build richer multi-level information through modeling hierarchical correlation. In addition, we propose a contrastive learning fine-tuning method based on keyphrase extraction tasks, which is used to train BERT and further improve the extraction performance of HIERank. The experimental results demonstrate that HIERank significantly outperform the compared baselines.