Abstract: A method of, and system for, extracting topic words from a collection of documents across multiple and potentially very large number of domains. Documents are selected and ranked based on similarity with at least one seed word, which defines a topic. Seed words may be entered directly by a user or provided by another application. Keywords are extracted from documents determined to be a sufficiently good match to the topic and may be displayed to the user or used as input into word prediction or word analysis and display software. Documents are determined to be a sufficiently good match to the topic using an iterative algorithm starting with the best match and selecting documents containing keywords sufficiently similar to the previously selected documents.