Keyword identification is an important task that provides useful information for NLP applications including: document retrieval, clustering, and categorization, among others. State-of-the-art methods rely on local features of words (e.g. lexical, syntactic, and presentation features) to assess their candidacy as keywords. In this paper, we propose a novel keyword identification method that relies on representation of text abstracts as word graphs. The significance of the proposed method stems from a flexible data representation that expands the context of words to span multiple sentences and thus can enable capturing
of important non-local graph topological features. Specifically, graphlets (small subgraph patterns) were efficiently extracted and scored to reflect the statistical dependency between these graphlet patterns and words labeled as keywords. Experimental results demonstrate the capability of the graphlet patterns in a keyword identification task when applied to MEDLINE, a standard research abstract dataset.
