Learning to extract from multiple perspectives for neural keyphrase extraction

Abstract

Keyphrase extraction is a fundamental task in natural language processing that aims at extracting a set of important phrases from a source document. Generally, when people understand documents and extract relevant information, they consider multiple perspectives, which helps to reduce misunderstandings and errors in extractions. However, most existing keyphrase extraction approaches focus on only one or two perspectives, resulting in inaccurate extractions. To address this issue, we propose a new neural keyphrase extraction model, called Multiple Perspectives Neural Keyphrase Extraction (MPNKE), which learns representations and estimates the importance of candidate phrases from multiple perspectives, similar to how a human would approach the task. Extensive experimental results on several benchmark keyphrase extraction datasets demonstrate that MPNKE outperforms existing state-of-the-art models in most cases.