A Contrastive Learning Framework for Keyphrase Extraction

Keyphrase extraction aims to extract important phrases that reflect the main topics of a document. Recently, deep learning methods are used to model semantic information and rank candidates based on the similarities between the n-grams and the document. However, existing keyphrase extraction methods mainly caused the keyphrase extraction task to be independent of the embedding. Based on the fact that phrases that are semantically closer to the document are more likely to become keyphrases, we propose a novel contrastive learning strategy for supervised keyphrase extraction by integrating local and global Information of the document. A pre-trained RoBERTa model is used to model contextual information of sub-words in the document. Then, the embedding vectors of n-grams and the document are calculated by the convolution neural layers. Finally, we propose a novel loss function for efficiently ranking candidate phrases by combining n-gram features and document embeddings during the training of the model.