Rating news documents for similarity

Reviewer: Nagiza F. Samatova

Readers of electronic news on the Web invariably seek answers for the six classic questions: who, what, where, when, why, and how. The authors address the challenge of how to extract these attributes automatically. This paper describes a system that recognizes dates and names of people, locations, and organizations, and rates news items based on their similarity to user-specified attribute phrases. The recognizer works by spotting capitalized name-phrases. The classifier places them into date, person, location, or organization categories by matching names from category-specific dictionaries. For analysis, authors use clean data with an XML-like mark up for the title, author, content, etc., which is not the case for most Web news sources full of extraneous information (e.g., advertisements). The accuracy of the classifier largely depends on the completeness of domain-specific dictionaries but the efficiency deteriorates with their size. It is unclear what trials were selected to achieve 90% classification accuracy and how efficient the recognizer and classifier are. To retrieve and rate news items, authors propose a similarity measure dependent on the similarity of attribute lists. Performance is measured by precision only. The precision of over 90% is demonstrated on a small set of 78 news items. Since the attribute lists are of indeterminate length, the efficiency of calculating the similarity is unclear. Unfortunately, this research is neither placed in context with the current literature on rating structured documents nor compared with traditional key-word retrieval systems. Thus, this work is somewhat preliminary and its applicability in the real world is questionable.