Phrases as Subtopical Concepts in Scholarly Text

摘要

Retrieval of subtopical concepts from scholarly communication systems is now possible through a combination of text and metadata analysis, augmented by user search queries and click logs. Here we investigate how a "phrase", defined as a variable length sequence of vocabulary words, can be used to represent a concept. We present a method to extract such phrases from a text corpus, and rank them using a citation network measure, the compensated normalized link count (CNLC), which measures the extent to which they are propagated along the citation structure of articles. We validate the ranking with actively and passively determined metrics: comparison with hum an-assigned keywords, and comparison with passively harvested terms from search query logs. This method is demonstrated on full texts and abstracts from 7 years of high energy physics articles from the arXiv preprint database.