Abstract
In text categorization, a well-known problem related to document length is that larger term counts in longer documents cause classification algorithms to become biased. The effect of document length can be eliminated by normalizing term counts, thus reducing the bias towards longer documents. This gives us term frequency (TF), which in conjunction with inverse document frequency (IDF) became the most commonly used term weighting scheme to capture the importance of a term in a document and corpus. However, normalization may cause term frequency of a term in a related document to become equal or smaller than its term frequency in an unrelated document, thus perturbing a term's strength from its true worth. In this paper, we solve this problem by introducing a non-linear mapping of term frequency. This alternative to TF is called binned term count (BTC). The newly proposed term frequency factor trims large term counts before normalization, thus moderating the normalization effect on large documents. To investigate the effectiveness of BTC, we compare it against the original TF and its more recently proposed alternative named modified term frequency (MTF). In our experiments, each of these term frequency factors (BTC, TF, and MTF) is combined with four well-known collection frequency factors (IDF), RF, IGM, and MONO and the performance of each of the resulting term weighting schemes is evaluated on three standard datasets (Reuters (R8-21578), 20-Newsgroups, and WebKB) using support vector machines and K-nearest neighbor classifiers. To determine whether BTC is statistically better than TF and MTF, we have applied the paired two-sided t-test on the macro F-1 results. Overall, BTC is found to be 52% statistically significant than TF and MTF. Furthermore, the highest macro F-1 value on the three datasets was achieved by BTC-based term weighting schemes.