Multi-level Contrastive Learning for Keyphrase Generation

Since contrastive learning-based methods take into account the differences between positive and negative samples, they have been applied to the keyphrase prediction. Existing keyphrase generation methods based on contrastive learning mainly focus on context-aware phrase-level representations, but fail to capture sequence-level information (i.e., order dependencies between key-phrases). This paper aims to effectively capture sequence-level features and integrate them with phrase-level features to improve the performance of keyphrase generation. Specifically, we propose a multi-level contrastive learning method (Multi-CL) for keyphrase generation. At the sequence level, it constructs fully permuted keyphrase sequence representations and captures effective keyphrase order dependencies in the sequences. At the phrase level, it learns context-aware keyphrase representations. Both contrastive learning modules are jointly trained with the document as the anchor so that they can learn complementary information from each other. Compared with eight baseline methods, Multi-CL demonstrates superior performance in experiments conducted on five benchmark datasets.