In this paper we explore the use of phrases occurring maximally in text as features for sentiment classification of product reviews. The goal is to find in a statistical way representative words and phrases used typically in positive and negative reviews. The approach does not rely on predefined sentiment lexicons, and the motivation for this is that potentially every word could be considered as expressing something positive and/or negative in different situations, and that the context and the personal attitude of the opinion holder should be taken into account when determining the polarity of the phrase, instead of doing this out of particular context.
展开▼