Multi-level Memory Network with CRFs for Keyphrase Extraction

Keyphrase, that concisely describe the high-level topics discussed in a document, are very useful for a wide range of natural language processing (NLP) tasks. Current popular supervised methods for keyphrase extraction commonly cannot effectively utilize the long-range contextual information in text. In this paper, we focus on how to effectively exploit the long-range contextual information to improve the keyphrase extraction performance. Specifically, we propose a multi-level memory network with the conditional random fields (CRFs), which allows to have unrestricted access to the long-range and local contextual information in text. We first design the multi-level memory network with sentence level and document level to enhance the text representation. Then, we integrate the multi-level memory network with the CRFs, which has an advantage in modeling the local contextual information. Compared with the recent state-of-the-art methods, our model can achieve better results through experiments on two datasets.