Gazetteer-Guided Keyphrase Generation from Research Papers

The task of keyphrase generation aims to generate the key phrases that capture the primary content of a document. An external domain-specific gazetteer can assist in generating keyphrases that are literally absent in the document (i.e., do not match any contiguous sub-sequence of source text) but relevant to the content of the document. In this paper, we present a technique to integrate knowledge from a gazetteer in order to improve keyphrase generation from research papers. We also present a copy mechanism that helps our model to utilize the gazetteer vocabulary to deal with the out-of-vocabulary words in keyphrases. Since constructing and maintaining relevant high-quality gazetteer by hand is very expensive, we also propose a method for automatic construction of a gazetteer given the input document, by leveraging similar documents in the training corpus. The thus constructed gazetteer helps focus on corpus-level information carried by other similar documents. Although this external information is crucial, it is never considered in previous studies. Experiments on real world datasets of research papers demonstrate that our proposed approach improves the performance of the state-of-the-art keyphrase generation models.