Stopwords Identification by Means of Characteristic and Discriminant Analysis

Stopwords are meaningless, non-significant terms that frequently occur in a document. They should be removed, like a noise. Traditionally, two different approaches of building a stoplist have been used: the former considers the most frequent terms looking at a language (e.g., english stoplist), the other includes the most occurring terms in a document collection. In several tasks, e.g., text classification and clustering, documents are typically grouped into categories. We propose a novel approach aimed at automatically identifying specific stopwords for each category. The proposal relies on two unbiased metrics that allow to analyze the informative content of each term; one measures the discriminant capability and the latter measures the characteristic capability. For each term, the former is expected to be high in accordance with the ability to distinguish a category against others, whereas the latter is expected to be high according to how the term is frequent and common over all categories. A preliminary study and experiments have been performed, pointing out our insight. Results confirm that, for each domain, the metrics easily identify specific stoplist wich include classical and category-dependent stopwords.