Impact of Context on Keyword Identification and Use in Biomedical Literature Mining

The use of two statistical metrics in automatically identifying important keywords associated with a concept such as a gene by mining scientific literature is reviewed. Starting with a subset of MEDLINE (R) abstracts that contain the name or synonyms of a gene in their titles, the aforementioned metrics contrast the prevalence of specific words in these documents against a broader "background set" of abstracts. If a word occurs substantially more often in the document subset associated with a gene than in the background set that acts as a reference, then the word is viewed as capturing some specific attribute of the gene. The keywords thus automatically identified may be used as gene features in clustering algorithms. Since the background set is the reference against which keyword prevalence is contrasted, the authors hypothesize that different background document sets can lead to somewhat different sets of keywords to be identified as specific to a gene. Two different background sets are discussed that are useful for two somewhat different purposes, namely, characterizing the function of a gene, and clustering a set of genes based on their shared functional similarities. Experimental results that reveal the significance of the choice of background set are presented.