Keyphrase Extraction with Dynamic Graph Convolutional Networks and Diversified Inference
Qiangjuan Huang, Yuanxin Liu, Haoyu Zhang · 2022
The generative framework based on Sequence-to-Sequence is adopted widely in the field of Keyphrase extraction, which is aim to extract a set of phrases that can accurately express the topic from a given document, and it shows a competitive performance on various benchmarks. However, acquiring informative latent document representation and update global information from generated keyphrases is still unsolved with these Seq2Seq methods. In this paper, we propose a method with designing the dynamic syntactic graph encoding (DGCN) to alleviate the above two problems. First, we introduce the dependency information of document along with the GCN encoder to learn a better context vector for the task. In addition, the graph edge weights is updated when decoding keyphrases. Therefore the introduce system can model the relations between the keyphrases set explicitly. Experiments on 5 extraction datasets shows that our work is effective.