DAKE: Document-Level Attention for Keyphrase Extraction

Keyphrases provide a concise representation of the topical content of a document and they are helpful in various downstream tasks. Previous approaches for keyphrase extraction model it as a sequence labelling task and use local contextual information to understand the semantics of the input text but they fail when the local context is ambiguous or unclear. We present a new framework to improve keyphrase extraction by utilizing additional supporting contextual information. We retrieve this additional information from other sentences within the same document. To this end, we propose Document-level Attention for Keyphrase Extraction (DAKE), which comprises Bidirectional Long Short-Term Memory networks that capture hidden semantics in text, a document-level attention mechanism to incorporate document level contextual information, gating mechanisms which help to determine the influence of additional contextual information on the fusion with local contextual information, and Conditional Random Fields which capture output label dependencies. Our experimental results on a dataset of research papers show that the proposed model outperforms previous state-of-the-art approaches for keyphrase extraction.