Hyperbolic Deep Keyphrase Generation

Abstract

Keyphrases can concisely describe the high-level topics discussed in a document, and thus keyphrase prediction compresses document’s hierarchical semantic information into a few important representative phrases. Numerous methods have been proposed to use the encoder-decoder framework in Euclidean space to generate keyphrases. However, their ability to capture the hierarchical structures is limited by the nature of Euclidean space. To this end, we propose a new research direction that aims to encode the hierarchical semantic information of a document into the low-dimensional representation and then decompress it to generate keyphrases in a hyperbolic space, which can effectively capture the underlying semantic hierarchical structures. In addition, we propose a novel hyperbolic attention mechanism to selectively focus on the high-level phrases in hierarchical semantics. To the best of our knowledge, this is the first study to explore a hyperbolic network for keyphrase generation. The experimental results illustrate that our method outperforms fifteen state-of-the-art methods across five datasets.