An Alternate Approach on Statistical Single-label Document Classification for Greek Newspaper Articles

Text classification is one of the most important sectors of machine learning theory. It enables a series of tasks among which are email spam filtering, text summarization and context identification. It has been one of the most intriguing tasks in the computer theory, as it represents the projection of human thought in machine form. Classification theory proposes a number of different techniques based on statistics (Naive Bayes Classifier (NBC), language models (LM)), artificial intelligence (Neural Networks) and decision trees among others. Classification systems are typically distinguished into single-label categorization and multi-label categorization systems, according to the number of categories they assign to each of the classified documents. In this paper, we present work undertaken in the area of single-label classification which resulted in a statistical classifier, based on the Naive Bayes assumption of statistical independence of word occurrence across a document. Our algorithm, takes into account cross-category word occurrence in deciding the class of a random document. Moreover, instead of estimating word co-occurrence in deciding a class, we estimate word contribution for a document to belong in a class. This approach outperforms other statistical classifiers as Naive Bayes Classifier and Language Models, as it was proven in our results.