Unsupervised Keyphrase Extraction via Interpretable Neural Networks

Anonymous

Abstract: Keyphrase extraction aims at automatically extracting a list of "important'' phrases which represent the key concepts in a document. Traditionally, it has been approached from an information-theoretic angle using phrase co-occurrence statistics. This work proposes a novel unsupervised approach to keyphrase extraction that uses a more intuitive notion of phrase importance, inspired by interpretability research. In particular, we use a self-explaining neural model to measure the predictive impact of input phrases on downstream task performance, and consider the resulting interpretations as document keyphrases for the target task. We show the efficacy of our approach on four datasets in two domains---scientific publications and news articles---attaining state-of-the-art results in unsupervised keyphrase extraction.