Keyphrase extraction from scientific literature using joint geometric graph embedding matching

Keyphrase extraction is the task of selecting representative words and phrases from a document. Recent research has focused on keyphrase extraction via graph-theoretic approaches, which leverage various graph-based centrality measures to locate important nodes in a constructed keyphrase candidate graph. In this thesis, we propose the use of an inexact graph matching algorithm for keyphrase extraction. We match graphs derived from test documents with graphs that have labeled keyphrases, and label as keyphrases the test nodes matching with known keyphrases. Our graph matching keyphrase extraction algorithm obtains an F-score of 14.6% on the standard SemEval 2010 Task 5 dataset and 37.1% on the well-known Inspec dataset. These scores are in line with time-tested algorithms on both datasets. We therefore conclude that inexact graph matching algorithms can be applied to keyphrase extraction successfully.