It has been shown that prior knowledge and information are organized according to categories, and that also background knowledge plays an important role in classification. The purpose of this study is first, to investigate the relationship between background knowledge and text classification, and second, to incorporate this relationship in a computational model. Our behavioral results demonstrate that participants with access to background knowledge (experts), overall performed significantly better than those without access to this knowledge (novices). More importantly, we show that experts rely more on relational features than surface features, an aspect that bag-of-words methods fail to capture. We then propose a computational model for text classification which incorporates background knowledge. This model is built upon vector-based representation methods and achieves significantly more accurate results over other models that were tested.