Adding numbers to text classification

Abstract

Many real-world problems involve a combination of both text- and numerical-valued features. For example, in email classification, it is possible to use instance representations that consider not only the text of each message, but also numerical-valued features such as the length of the message or the time of day at which it was sent. Text-classification methods have thus far not easily incorporated numerical features. In earlier work we described an approach for converting numerical features into bags of tokens so that text classification methods can be applied to numerical classification problems, and showed that the resulting learning methods are competitive with traditional numerical classification methods. In this paper we use this as a way to learn on problems that involve a combination of text and numbers. We show that the results outperform competing methods. Further, we show that selecting a best classification method using text-only features and then adding numerical features to the problem (as might happen if numerical features are only later added to a pre existing text-classification problem) gives performance that rivals a more time-consuming approach of evaluating all classification methods using the full set of both text and numerical features.