Cross-comparison for Two-dimensional Text Categorization

The organization of large text collections is the main goal of automated text categorization. In particular, the final aim is to classify documents into a certain number of pre-defined categories in an efficient way and with as much accuracy as possible. On-line and run-time services, such as personalization services and information filtering services, have increased the importance of effective and efficient document categorization techniques. In the last years, a wide range of supervised learning algorithms have been applied to this problem [1]. Recently, a new approach that exploits a two-dimensional summarization of the data for text classification was presented [2]. This method does not go through a selection of words phase; instead, it uses the whole dictionary to present data in intuitive way on two-dimensional graphs. Although successful in terms of classification effectiveness and efficiency (as recently showed in [3]), this method presents some unsolved key issues: the design of the training algorithm seems to be ad hoc for the Reuters-21578 collection; the evaluation has only been done only on the 10 most frequent classes of the Reuters-21578 dataset; the evaluation lacks measure of significance in most parts; the method adopted lacks a mathematical justification. We focus on the first three aspects, leaving the fourth as the future work.