Re-examining automatic keyphrase extraction approaches in scientific articles

We tackle two major issues in automatic keyphrase extraction using scientific articles: candidate selection and feature engineering. To develop an efficient candidate selection method, we analyze the nature and variation of keyphrases and then select candidates using regular expressions. Secondly, we re-examine the existing features broadly used for the supervised approach, exploring different ways to enhance their performance. While most other approaches are supervised, we also study the optimal features for unsupervised keyphrase extraction. Our research has shown that effective candidate selection leads to better performance as evaluation accounts for candidate coverage. Our work also attests that many of existing features are also usable in unsupervised extraction.