A Phrase-Level Attention Enhanced CRF for Keyphrase Extraction

Since sequence labeling-based methods take into account the dependencies between neighbouring labels, they have been widely used for keyphrase prediction. Existing methods mainly focus on the word-level sequence labeling over the word-level features, and fail to capture the phrase-level information (i.e., inner properties of multi-word keyphrases). In this paper, we concentrate on how to effectively capture the phrase-level features and then integrate them with the word-level features to improve the performance of keyphrase extraction in the sequence labeling-based method. Specifically, we propose a phrase-level attention enhanced conditional random field (PAE-CRF) model for keyphrase extraction, which consists of two major modules: a phrase-level attention module that captures phrase-level features, and a phrase-level attention enhanced CRF module that integrates the phrase-level attention information with the word-level features into CRF to extract keyphrases. Finally, these two modules are jointly trained to help them learn complementary information from each other. Compared with the recent state-of-the-art methods, our model can achieve better results through experiments on four benchmark datasets. The code and keyphrase prediction results of our model are available in public at  https://github.com/pae-crf/PAE-CRF .