Improved identification of tweets that mention books: Selection of effective features

Abstract

In this paper, we assessed the effectiveness of different types of features for the identification of tweets on Twitter that mention books among tweets that contain the same strings as full book titles. In the previous work, the bag-of-words based features were taken from the context of individual tweets. While performance was reasonable, we identified room for improvement in terms of the extraction of features. We proposed additional types of features such as words appearing in the profiles of tweet authors, POS tags of mentioned book titles, and bibliographic elements within tweets, e.g. authors and publishers. We conducted a grid search for all combinations of the above feature sets, and observed performance improvements suitable for practical applications.