Keyphrase Generation is the task of predicting keyphrases: short text sequences that convey the main semantic meaning of a document. In this paper, we introduce a keyphrase generation approach that makes use of a Generative Adversarial Networks (GANs) architecture. In our system, the Generator produces a sequence of keyphrases for an input document. The Discriminator, in turn, tries to distinguish between machine generated and human curated keyphrases. We propose a novel Discriminator architecture based on a BERT pretrained model fine-tuned for Sequence Classification. We train our proposed architecture using only a small subset of the standard available training dataset, amounting to less than 1% of the total, achieving a great level of data efficiency. The resulting model is evaluated on five public datasets, obtaining competitive and promising results with respect to four state-of-the-art generative models.

Efficient Keyphrase Generation with GANs

Giuseppe Lancioni
;
Saida Saad Mohamed Mahmoud;Beatrice Portelli
;
Giuseppe Serra
;
Carlo Tasso
2021-01-01

Abstract

Keyphrase Generation is the task of predicting keyphrases: short text sequences that convey the main semantic meaning of a document. In this paper, we introduce a keyphrase generation approach that makes use of a Generative Adversarial Networks (GANs) architecture. In our system, the Generator produces a sequence of keyphrases for an input document. The Discriminator, in turn, tries to distinguish between machine generated and human curated keyphrases. We propose a novel Discriminator architecture based on a BERT pretrained model fine-tuned for Sequence Classification. We train our proposed architecture using only a small subset of the standard available training dataset, amounting to less than 1% of the total, achieving a great level of data efficiency. The resulting model is evaluated on five public datasets, obtaining competitive and promising results with respect to four state-of-the-art generative models.
File in questo prodotto:
File Dimensione Formato  
IRCDL_2021_paper_1.pdf

accesso aperto

Descrizione: Articolo principale
Tipologia: Versione Editoriale (PDF)
Licenza: Creative commons
Dimensione 291.31 kB
Formato Adobe PDF
291.31 kB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11390/1207164
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? ND
social impact